research paper on artificial intelligence in marketing

Artificial Intelligence in Marketing: Volume 20

Table of contents, the state of ai research in marketing: active, fertile, and ready for explosive growth, the economics of artificial intelligence: a marketing perspective.

This article reviews the quantitative marketing literature on artificial intelligence (AI) through an economics lens. We apply the framework in Prediction Machines: The Simple Economics of Artificial Intelligence to systematically categorize 96 research papers on AI in marketing academia into five levels of impact, which are prediction, decision, tool, strategy, and society. For each paper, we further identify each individual component of a task, the research question, the AI model used, and the broad decision type. Overall, we find there are fewer marketing papers focusing on strategy and society, and accordingly, we discuss future research opportunities in those areas.

AI and Personalization

This chapter reviews the recent developments at the intersection of personalization and AI in marketing and related fields. We provide a formal definition of personalized policy and review the methodological approaches available for personalization. We discuss scalability, generalizability, and counterfactual validity issues and briefly touch upon advanced methods for online/interactive/dynamic settings. We then summarize the three evaluation approaches for static policies – the Direct method, the Inverse Propensity Score (IPS) estimator, and the Doubly Robust (DR) method. Next, we present a summary of the evaluation approaches for special cases such as continuous actions and dynamic settings. We then summarize the findings on the returns to personalization across various domains, including content recommendation, advertising, and promotions. Next, we discuss the work on the intersection between personalization and welfare. We focus on four of these welfare notions that have been studied in the literature: (1) search costs, (2) privacy, (3) fairness, and (4) polarization. We conclude with a discussion of the remaining challenges and some directions for future research.

Artificial Intelligence and Pricing

As businesses become more sophisticated and welcome new technologies, artificial intelligence (AI)-based methods are increasingly being used for firms' pricing decisions. In this review article, we provide a survey of research in the area of AI and pricing. On the upside, research has shown that algorithms allow companies to achieve unprecedented advantages, including real-time response to demand and supply shocks, personalized pricing, and demand learning. However, recent research has uncovered unforeseen downsides to algorithmic pricing that are important for managers and policy-makers to consider.

Leveraging AI for Content Generation: A Customer Equity Perspective

Advances in artificial intelligence have ushered in new opportunities for marketers in the domain of content generation. We discuss approaches that have emerged to generate text and image content. Drawing on the customer equity framework, we then discuss the potential applications of automated content generation for customer acquisition, relationship development, and customer retention. We conclude by discussing important considerations that businesses must make prior to adopting automated content generation.

Artificial Intelligence and User-Generated Data Are Transforming How Firms Come to Understand Customer Needs

We provide an overview of how artificial intelligence is transforming the identification, structuring, and prioritization of customer needs – known as the voice of the customer (VOC). First, we summarize how the VOC helps firms gain insights on using user-generated data. Second, we discuss the types of user-generated data and the challenges associated with analyzing each type of data. Third, we describe common methods, matched to the firms' goals and the structure of the data, that are used to analyze the VOC. Fourth, and most importantly, we map the methods to relevant applications, providing guidance to select the appropriate method to address the desired research questions.

Artificial Intelligence Applications to Customer Feedback Research: A Review

In this paper, we aim to provide a comprehensive overview of customer feedback literature, highlighting the burgeoning role of artificial intelligence (AI). Customer feedback has long been a valuable source of customer insights for businesses and market researchers. While previously survey focused, customer feedback in the digital age has evolved to be rich, interactive, multimodal, and virtually real time. Such explosion in feedback content has also been accompanied by a rapid development of AI and machine learning technologies that enable firms to understand and take advantage of these high-velocity data sources. Yet, some of the challenges with traditional surveys remain, such as self-selection concerns of who chooses to participate and what attributes they give feedback on. In addition, these new feedback channels face other unique challenges like review manipulation and herding effects due to their public and democratic nature. Thus, while the AI toolkit has revolutionized the area of customer feedback, extracting meaningful insights requires complementing it with the appropriate social science toolkit. We begin by touching upon conventional customer feedback research and chart its evolution through the years as the nature of available data and analysis tools develop. We conclude by providing recommendations for future questions that remain to be explored in this field.

Natural Language Processing in Marketing

The increasing importance and proliferation of text data provide a unique opportunity and novel lens to study human communication across a myriad of business and marketing applications. For example, consumers compare and review products online, individuals interact with their voice assistants to search, shop, and express their needs, investors seek to extract signals from firms' press releases to improve their investment decisions, and firms analyze sales call transcripts to increase customer satisfaction and conversions. However, extracting meaningful information from unstructured text data is a nontrivial task. In this chapter, we review established natural language processing (NLP) methods for traditional tasks (e.g., LDA for topic modeling and lexicons for sentiment analysis and writing style extraction) and provide an outlook into the future of NLP in marketing, covering recent embedding-based approaches, pretrained language models, and transfer learning for novel tasks such as automated text generation and multi-modal representation learning. These emerging approaches allow the field to improve its ability to perform certain tasks that we have been using for more than a decade (e.g., text classification). But more importantly, they unlock entirely new types of tasks that bring about novel research opportunities (e.g., text summarization, and generative question answering). We conclude with a roadmap and research agenda for promising NLP applications in marketing and provide supplementary code examples to help interested scholars to explore opportunities related to NLP in marketing.

Marketing Through the Machine's Eyes: Image Analytics and Interpretability

The growth of social media and the sharing economy is generating abundant unstructured image and video data. Computer vision techniques can derive rich insights from unstructured data and can inform recommendations for increasing profits and consumer utility – if only the model outputs are interpretable enough to earn the trust of consumers and buy-in from companies. To build a foundation for understanding the importance of model interpretation in image analytics, the first section of this article reviews the existing work along three dimensions: the data type (image data vs. video data), model structure (feature-level vs. pixel-level), and primary application (to increase company profits vs. to maximize consumer utility). The second section discusses how the “black box” of pixel-level models leads to legal and ethical problems, but interpretability can be improved with eXplainable Artificial Intelligence (XAI) methods. We classify and review XAI methods based on transparency, the scope of interpretability (global vs. local), and model specificity (model-specific vs. model-agnostic); in marketing research, transparent, local, and model-agnostic methods are most common. The third section proposes three promising future research directions related to model interpretability: the economic value of augmented reality in 3D product tracking and visualization, field experiments to compare human judgments with the outputs of machine vision systems, and XAI methods to test strategies for mitigating algorithmic bias.

Deep Learning in Marketing: A Review and Research Agenda

The expansion of marketing data is encouraging the growing use of deep learning (DL) in marketing. I summarize the intuition behind deep learning and explain the mechanisms of six popular algorithms: three discriminative (convolutional neural network (CNN), recurrent neural network (RNN), and Transformer), two generative (variational autoencoder (VAE) and generative adversarial networks (GAN)), and one RL (DQN). I discuss what marketing problems DL is useful for and what fueled its growth in recent years. I emphasize the power and flexibility of DL for modeling unstructured data when formal theories and knowledge are absent. I also describe future research directions.

Anthropomorphism in Artificial Intelligence: A Review of Empirical Work Across Domains and Insights for Future Research

Anthropomorphism in Artificial Intelligence (AI)-powered devices is being used increasingly frequently in consumer-facing situations (e.g., AI Assistants such as Alexa, virtual agents in websites, call/chat bots, etc.), and therefore, it is essential to understand anthropomorphism in AI both to understand consequences for consumers and to optimize firms' product development and marketing. Extant literature is fragmented across several domains and is limited in the marketing domain. In this review, we aim to bring together the insights from different fields and develop a parsimonious conceptual framework to guide future research in fields of marketing and consumer behavior.

Methodology

We conduct a review of empirical articles published until November 2021 in Financial Times Top 50 (FT50) journals as well as in 41 additional journals selected across several disciplinary domains: computer science, robotics, psychology, marketing, and consumer behavior.

Based on literature review and synthesis, we propose a three-step guiding framework for future research and practice on AI anthropomorphism.

Research Implications

Our proposed conceptual framework informs marketing and consumer behavior domains with findings accumulated in other research domains, offers important directions for future research, and provides a parsimonious guide for marketing managers to optimally utilize anthropomorphism in AI to the benefit of both firms and consumers.

Originality/Value

We contribute to the emerging literature on anthropomorphism in AI in three ways. First, we expedite the information flow between disciplines by integrating insights from different fields of inquiry. Second, based on our synthesis of literature, we offer a conceptual framework to organize the outcomes of AI anthropomorphism in a tidy and concise manner. Third, based on our review and conceptual framework, we offer key directions to guide future research endeavors.

• Olivier Toubia

We’re listening — tell us what you think

Something didn’t work….

Report bugs here

All feedback is valuable

Please share your general feedback

Join us on our journey

Platform update page.

Visit emeraldpublishing.com/platformupdate to discover the latest news and updates

Questions & More Information

Answers to the most commonly asked questions here

Artificial Intelligence in Marketing Thesis Paper

Key takeaways.

Artificial Intelligence (AI) in Marketing - read the research paper (PDF):

How does Artificial Intelligence (AI) affect marketing?

Artificial intelligence is a burgeoning technology, industry, and field of study. While interest levels regarding its applications in marketing have not yet translated into widespread adoption, AI and its subcategories like Machine Learning (ML) and Deep Learning (DL) hold tremendous potential for vastly altering how marketing is done. AI, ML and DL offer an improvement to current marketing tactics, as well as entirely new ways of creating and distributing value to customers. For example, programmatic advertising and social media marketing can allow for a more comprehensive view of customer behavior, predictive analytics, and deeper insights through integration with AI. New marketing tools like biometrics, voice, and conversational user interfaces offer novel ways to add value for brands and consumers alike. These innovations all carry similar characteristics of hyper-personalization, efficient spending, scalable experiences, and deep insights.

Adapting to an AI Marketing landscape

There are many important issues that need to be addressed before AI is extensively implemented, including the potential for it to be used maliciously, its effects on job displacement, and the technology itself. The recent progress of AI in marketing is indicative that it will be adopted by a majority of companies soon. The long-term implications of vast implementation are crucial to consider, as an AI-powered industry entails fundamental changes to the skill-sets required to thrive, the way marketers and brands work, and consumer expectations.

Artificial Intelligence in Marketing (Research Paper - Honors Thesis)

Table of Contents

Introduction, scope and methodology.

A brief history and origin story of AI.

Consumer Perceptions

How do people perceive AI currently? What are the brand safety implications of using AI for marketing?

Defining Artificial Intelligence

Key AI terminology and vocabulary to know.

Artificial Intelligence (AI)

Artificial general intelligence (agi), machine learning (ml), deep learning (dl), natural language processing (nlp), signal processing, ai in marketing today.

How are brands currently using AI in marketing today? What are some examples of how AI can be applied to marketing?

1:1 Marketing

How AI enables hyper-personalization, or 1-to-1 marketing at-scale.

Levels of AI Implementation

A framework for analyzing the varying degrees in which companies are implementing AI currently, from off-the-shelf AI products to full-scale deployment.

Programmatic Advertising

How does AI affect programmatic digital advertising? The impact of AI on Facebook Ads, Google Ads, and more.

Transparency , Distrust, and Fraud

Omnichannel marketing, retargeting, organizational structure.

How companies are structuring their teams to enable successful AI projects.

Image Recognition and Computer Vision

A look at various applications of AI image recognition and computer vision models in marketing.

Social Media

Segmentation and targeting, facial recognition, interactive marketing through biometrics.

Applications of AI for creative use cases in marketing.

Personalized Narratives

Localization, audio generation, image curation, augmentation, data synergy.

How AI-powered chatbots can be used in marketing, eCommerce, and customer service.

Customer Service

Personal assistants, chatbot management, personalized ui and ux.

Using AI to deliver personalized digital experiences.

Applications of AI for voice.

Advertising on Voice Assistants

Impact of voice on seo, shopping using voice, automated phone calls, pitfalls of ai and areas of improvement.

What are some of the main weaknesses of AI in its current state? How can AI be improved?

Malicious AI

Job displacement, underlying technology, where marketers should go.

How can marketers adapt to an AI Marketing landscape? How can brands and non-programmers get started with AI?

The Individual's Perspective

The brand's perspective, ai readiness framework, implementation.

What are the steps for implementing AI? What are the prerequisites for a successful AI launch?

Requirements

Long-term implications.

How will AI affect marketing long-term? What changes might we see as AI becomes more widely adopted? How does the relationship between brands and consumers change in an AI Marketing landscape?

Impact of AI on Consumer Behavior

Increased importance of brand purpose, human-centered technology.

Citations for Artificial Intelligence in Marketing (Research Paper)

Get personalized, actionable resources sent straight to your inbox.

• Share full article

Advertisement

Supported by

In One Key A.I. Metric, China Pulls Ahead of the U.S.: Talent

China has produced a huge number of top A.I. engineers in recent years. New research shows that, by some measures, it has already eclipsed the United States.

By Paul Mozur and Cade Metz

Paul Mozur reported from Taipei, Taiwan, and Cade Metz from San Francisco.

When it comes to the artificial intelligence that powers chatbots like ChatGPT, China lags behind the United States . But when it comes to producing the scientists behind a new generation of humanoid technologies, China is pulling ahead.

New research shows that China has by some metrics eclipsed the United States as the biggest producer of A.I. talent, with the country generating almost half the world’s top A.I. researchers. By contrast, about 18 percent come from U.S. undergraduate institutions, according to the study , from MacroPolo, a think tank run by the Paulson Institute, which promotes constructive ties between the United States and China.

The findings show a jump for China, which produced about one-third of the world’s top talent three years earlier. The United States, by contrast, remained mostly the same. The research is based on the backgrounds of researchers whose papers were published at 2022’s Conference on Neural Information Processing Systems. NeurIPS, as it is known, is focused on advances in neural networks , which have anchored recent developments in generative A.I.

The talent imbalance has been building for the better part of a decade. During much of the 2010s, the United States benefited as large numbers of China’s top minds moved to American universities to complete doctoral degrees. A majority of them stayed in the United States. But the research shows that trend has also begun to turn, with growing numbers of Chinese researchers staying in China.

What happens in the next few years could be critical as China and the United States jockey for primacy in A.I. — a technology that can potentially increase productivity, strengthen industries and drive innovation — turning the researchers into one of the most geopolitically important groups in the world.

Generative A.I. has captured the tech industry in Silicon Valley and in China, causing a frenzy in funding and investment. The boom has been led by U.S. tech giants such as Google and start-ups like OpenAI. That could attract China’s researchers, though rising tensions between Beijing and Washington could also deter some, experts said.

(The New York Times has sued OpenAI and Microsoft for copyright infringement of news content related to A.I. systems.)

China has nurtured so much A.I. talent partly because it invested heavily in A.I. education. Since 2018, the country has added more than 2,000 undergraduate A.I. programs, with more than 300 at its most elite universities, said Damien Ma, the managing director of MacroPolo, though he noted the programs were not heavily focused on the technology that had driven breakthroughs by chatbots like ChatGPT.

“A lot of the programs are about A.I. applications in industry and manufacturing, not so much the generative A.I. stuff that’s come to dominate the American A.I. industry at the moment,” he said.

While the United States has pioneered breakthroughs in A.I., most recently with the uncanny humanlike abilities of chatbots , a significant portion of that work was done by researchers educated in China.

Researchers originally from China now make up 38 percent of the top A.I. researchers working in the United States, with Americans making up 37 percent, according to the research. Three years earlier, those from China made up 27 percent of top talent working in the United States, compared with 31 percent from the United States.

“The data shows just how critical Chinese-born researchers are to the United States for A.I. competitiveness,” said Matt Sheehan, a fellow at the Carnegie Endowment for International Peace who studies Chinese A.I.

He added that the data seemed to show the United States was still attractive. “We’re the world leader in A.I. because we continue to attract and retain talent from all over the world, but especially China,” he said.

Pieter Abbeel, a professor at the University of California, Berkeley, and a founder of Covariant , an A.I. and robotics start-up, said working alongside large numbers of Chinese researchers was taken for granted inside the leading American companies and universities.

“It’s just a natural state of affairs,” he said.

In the past, U.S. defense officials were not too concerned about A.I. talent flows from China, partly because many of the biggest A.I. projects did not deal with classified data and partly because they reasoned that it was better to have the best minds available. That so much of the leading research in A.I. is published openly also held back worries.

Despite bans introduced by the Trump administration that prohibit entry to the United States for students from some military-linked universities in China and a relative slowdown in the flow of Chinese students into the country during Covid, the research showed large numbers of the most promising A.I. minds continued coming to the United States to study.

But this month, a Chinese citizen who was an engineer at Google was charged with trying to transfer A.I. technology — including critical microchip architecture — to a Beijing-based company that paid him in secret , according to a federal indictment.

The substantial numbers of Chinese A.I. researchers working in the United States now present a conundrum for policymakers, who want to counter Chinese espionage while not discouraging the continued flow of top Chinese computer engineers into the United States, according to experts focused on American competitiveness.

“Chinese scholars are almost leading the way in the A.I. field,” said Subbarao Kambhampati, a professor and researcher of A.I. at Arizona State University. If policymakers try to bar Chinese nationals from research in the United States, he said, they are “shooting themselves in the foot.”

The track record of U.S. policymakers is mixed. A policy by the Trump administration aimed at curbing Chinese industrial espionage and intellectual property theft has since been criticized for errantly prosecuting a number of professors. Such programs, Chinese immigrants said, have encouraged some to stay in China.

For now, the research showed, most Chinese who complete doctorates in the United States stay in the country, helping to make it the global center of the A.I. world. Even so, the U.S. lead has begun to slip, to hosting about 42 percent of the world’s top talent, down from about 59 percent three years ago, according to the research.

Paul Mozur is the global technology correspondent for The Times, based in Taipei. Previously he wrote about technology and politics in Asia from Hong Kong, Shanghai and Seoul. More about Paul Mozur

Cade Metz writes about artificial intelligence, driverless cars, robotics, virtual reality and other emerging areas of technology. More about Cade Metz

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• My Account Login
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Open access
• Published: 19 March 2024

TacticAI: an AI assistant for football tactics

• Zhe Wang   ORCID: orcid.org/0000-0002-0748-5376 1   na1 ,
• Petar Veličković   ORCID: orcid.org/0000-0002-2820-4692 1   na1 ,
• Daniel Hennes   ORCID: orcid.org/0000-0002-3646-5286 1   na1 ,
• Nenad Tomašev   ORCID: orcid.org/0000-0003-1624-0220 1 ,
• Laurel Prince 1 ,
• Michael Kaisers 1 ,
• Yoram Bachrach 1 ,
• Romuald Elie 1 ,
• Li Kevin Wenliang 1 ,
• Federico Piccinini 1 ,
• William Spearman 2 ,
• Ian Graham 3 ,
• Jerome Connor 1 ,
• Yi Yang 1 ,
• Adrià Recasens 1 ,
• Mina Khan 1 ,
• Nathalie Beauguerlange 1 ,
• Pablo Sprechmann 1 ,
• Pol Moreno 1 ,
• Nicolas Heess   ORCID: orcid.org/0000-0001-7876-9256 1 ,
• Michael Bowling   ORCID: orcid.org/0000-0003-2960-8418 4 ,
• Demis Hassabis 1 &
• Karl Tuyls   ORCID: orcid.org/0000-0001-7929-1944 5  

Nature Communications volume  15 , Article number:  1906 ( 2024 ) Cite this article

15k Accesses

237 Altmetric

Metrics details

• Computational science
• Information technology

Identifying key patterns of tactics implemented by rival teams, and developing effective responses, lies at the heart of modern football. However, doing so algorithmically remains an open research challenge. To address this unmet need, we propose TacticAI, an AI football tactics assistant developed and evaluated in close collaboration with domain experts from Liverpool FC. We focus on analysing corner kicks, as they offer coaches the most direct opportunities for interventions and improvements. TacticAI incorporates both a predictive and a generative component, allowing the coaches to effectively sample and explore alternative player setups for each corner kick routine and to select those with the highest predicted likelihood of success. We validate TacticAI on a number of relevant benchmark tasks: predicting receivers and shot attempts and recommending player position adjustments. The utility of TacticAI is validated by a qualitative study conducted with football domain experts at Liverpool FC. We show that TacticAI’s model suggestions are not only indistinguishable from real tactics, but also favoured over existing tactics 90% of the time, and that TacticAI offers an effective corner kick retrieval system. TacticAI achieves these results despite the limited availability of gold-standard data, achieving data efficiency through geometric deep learning.

Similar content being viewed by others

Automated recognition of the cricket batting backlift technique in video footage using deep learning architectures

Tevin Moodley, Dustin van der Haar & Habib Noorbhai

Synchronization of passes in event and spatiotemporal soccer data

Henrik Biermann, Rumena Komitova, … Daniel Memmert

A Random Forest approach to identify metrics that best predict match outcome and player ranking in the esport Rocket League

Tim D. Smithies, Mark J. Campbell, … Adam J. Toth

Introduction

Association football, or simply football or soccer, is a widely popular and highly professionalised sport, in which two teams compete to score goals against each other. As each football team comprises up to 11 active players at all times and takes place on a very large pitch (also known as a soccer field), scoring goals tends to require a significant degree of strategic team-play. Under the rules codified in the Laws of the Game 1 , this competition has nurtured an evolution of nuanced strategies and tactics, culminating in modern professional football leagues. In today’s play, data-driven insights are a key driver in determining the optimal player setups for each game and developing counter-tactics to maximise the chances of success 2 .

When competing at the highest level the margins are incredibly tight, and it is increasingly important to be able to capitalise on any opportunity for creating an advantage on the pitch. To that end, top-tier clubs employ diverse teams of coaches, analysts and experts, tasked with studying and devising (counter-)tactics before each game. Several recent methods attempt to improve tactical coaching and player decision-making through artificial intelligence (AI) tools, using a wide variety of data types from videos to tracking sensors and applying diverse algorithms ranging from simple logistic regression to elaborate neural network architectures. Such methods have been employed to help predict shot events from videos 3 , forecast off-screen movement from spatio-temporal data 4 , determine whether a match is in-play or interrupted 5 , or identify player actions 6 .

The execution of agreed-upon plans by players on the pitch is highly dynamic and imperfect, depending on numerous factors including player fitness and fatigue, variations in player movement and positioning, weather, the state of the pitch, and the reaction of the opposing team. In contrast, set pieces provide an opportunity to exert more control on the outcome, as the brief interruption in play allows the players to reposition according to one of the practiced and pre-agreed patterns, and make a deliberate attempt towards the goal. Examples of such set pieces include free kicks, corner kicks, goal kicks, throw-ins, and penalties 2 .

Among set pieces, corner kicks are of particular importance, as an improvement in corner kick execution may substantially modify game outcomes, and they lend themselves to principled, tactical and detailed analysis. This is because corner kicks tend to occur frequently in football matches (with ~10 corners on average taking place in each match 7 ), they are taken from a fixed, rigid position, and they offer an immediate opportunity for scoring a goal—no other set piece simultaneously satisfies all of the above. In practice, corner kick routines are determined well ahead of each match, taking into account the strengths and weaknesses of the opposing team and their typical tactical deployment. It is for this reason that we focus on corner kick analysis in particular, and propose TacticAI, an AI football assistant for supporting the human expert with set piece analysis, and the development and improvement of corner kick routines.

TacticAI is rooted in learning efficient representations of corner kick tactics from raw, spatio-temporal player tracking data. It makes efficient use of this data by representing each corner kick situation as a graph—a natural representation for modelling relationships between players (Fig.  1 A, Table  2 ), and these player relationships may be of higher importance than the absolute distances between them on the pitch 8 . Such a graph input is a natural candidate for graph machine learning models 9 , which we employ within TacticAI to obtain high-dimensional latent player representations. In the Supplementary Discussion section, we carefully contrast TacticAI against prior art in the area.

A How corner kick situations are converted to a graph representation. Each player is treated as a node in a graph, with node, edge and graph features extracted as detailed in the main text. Then, a graph neural network operates over this graph by performing message passing; each node’s representation is updated using the messages sent to it from its neighbouring nodes. B How TacticAI processes a given corner kick. To ensure that TacticAI’s answers are robust in the face of horizontal or vertical reflections, all possible combinations of reflections are applied to the input corner, and these four views are then fed to the core TacticAI model, where they are able to interact with each other to compute the final player representations—each internal blue arrow corresponds to a single message passing layer from ( A ). Once player representations are computed, they can be used to predict the corner’s receiver, whether a shot has been taken, as well as assistive adjustments to player positions and velocities, which increase or decrease the probability of a shot being taken.

Uniquely, TacticAI takes advantage of geometric deep learning 10 to explicitly produce player representations that respect several symmetries of the football pitch (Fig.  1 B). As an illustrative example, we can usually safely assume that under a horizontal or vertical reflection of the pitch state, the game situation is equivalent. Geometric deep learning ensures that TacticAI’s player representations will be identically computed under such reflections, such that this symmetry does not have to be learnt from data. This proves to be a valuable addition, as high-quality tracking data is often limited—with only a few hundred matches played each year in every league. We provide an in-depth overview of how we employ geometric deep learning in TacticAI in the “Methods” section.

From these representations, TacticAI is then able to answer various predictive questions about the outcomes of a corner—for example, which player is most likely to make first contact with the ball, or whether a shot will take place. TacticAI can also be used as a retrieval system—for mining similar corner kick situations based on the similarity of player representations—and a generative recommendation system, suggesting adjustments to player positions and velocities to maximise or minimise the estimated shot probability. Through several experiments within a case study with domain expert coaches and analysts from Liverpool FC, the results of which we present in the next section, we obtain clear statistical evidence that TacticAI readily provides useful, realistic and accurate tactical suggestions.

To demonstrate the diverse qualities of our approach, we design TacticAI with three distinct predictive and generative components: receiver prediction, shot prediction, and tactic recommendation through guided generation, which also correspond to the benchmark tasks for quantitatively evaluating TacticAI. In addition to providing accurate quantitative insights for corner kick analysis with its predictive components, the interplay between TacticAI’s predictive and generative components allows coaches to sample alternative player setups for each routine of interest, and directly evaluate the possible outcomes of such alternatives.

We will first describe our quantitative analysis, which demonstrates that TacticAI’s predictive components are accurate at predicting corner kick receivers and shot situations on held-out test corners and that the proposed player adjustments do not strongly deviate from ground-truth situations. However, such an analysis only gives an indirect insight into how useful TacticAI would be once deployed. We tackle this question of utility head-on and conduct a comprehensive case study in collaboration with our partners at Liverpool FC—where we directly ask human expert raters to judge the utility of TacticAI’s predictions and player adjustments. The following sections expand on the specific results and analysis we have performed.

In what follows, we will describe TacticAI’s components at a minimal level necessary to understand our evaluation. We defer detailed descriptions of TacticAI’s components to the “Methods” section. Note that, all our error bars reported in this research are standard deviations.

Benchmarking TacticAI

We evaluate the three components of TacticAI on a relevant benchmark dataset of corner kicks. Our dataset consists of 7176 corner kicks from the 2020 to 2021 Premier League seasons, which we randomly shuffle and split into a training (80%) and a test set (20%). As previously mentioned, TacticAI operates on graphs. Accordingly, we represent each corner kick situation as a graph, where each node corresponds to a player. The features associated with each node encode the movements (velocities and positions) and simple profiles (heights and weights) of on-pitch players at the timestamp when the corresponding corner kick was being taken by the attacking kicker (see the “Methods” section), and no information of ball movement was encoded. The graphs are fully connected; that is, for every pair of players, we will include the edge connecting them in the graph. Each of these edges encodes a binary feature, indicating whether the two players are on opposing teams or not. For each task, we generated the relevant dataset of node/edge/graph features and corresponding labels (Tables  1 and 2 , see the “Methods” section). The components were then trained separately with their corresponding corner kick graphs. In particular, we only employ a minimal set of features to construct the corner kick graphs, without encoding the movements of the ball nor explicitly encoding the distances between players into the graphs. We used a consistent training-test split for all benchmark tasks, as this made it possible to benchmark not only the individual components but also their interactions.

Accurate receiver and shot prediction through geometric deep learning

One of TacticAI’s key predictive models forecasts the receiver out of the 22 on-pitch players. The receiver is defined as the first player touching the ball after the corner is taken. In our evaluation, all methods used the same set of features (see the “Receiver prediction” entry in Table  1 and the “Methods” section). We leveraged the receiver prediction task to benchmark several different TacticAI base models. Our best-performing model—achieving 0.782 ± 0.039 in top-3 test accuracy after 50,000 training steps—was a deep graph attention network 11 , 12 , leveraging geometric deep learning 10 through the use of D 2 group convolutions 13 . We supplement this result with a detailed ablation study, verifying that both our choice of base architecture and group convolution yielded significant improvements in the receiver prediction task (Supplementary Table  2 , see the subsection “Ablation study” in the “Methods” section). Considering that corner kick receiver prediction is a highly challenging task with many factors that are unseen by our model—including fatigue and fitness levels, and actual ball trajectory—we consider TacticAI’s top-3 accuracy to reflect a high level of predictive power, and keep the base TacticAI architecture fixed for subsequent studies. In addition to this quantitative evaluation with the evaluation dataset, we also evaluate the performance of TacticAI’s receiver prediction component in a case study with human raters. Please see the “Case study with expert raters” section for more details.

For shot prediction, we observe that reusing the base TacticAI architecture to directly predict shot events—i.e., directly modelling the probability \({\mathbb{P}}(\,{{\mbox{shot}}}| {{\mbox{corner}}}\,)\) —proved challenging, only yielding a test F 1 score of 0.52 ± 0.03, for a GATv2 base model. Note that here we use the F 1 score—the harmonic mean of precision and recall—as it is commonly used in binary classification problems over imbalanced datasets, such as shot prediction. However, given that we already have a potent receiver predictor, we decided to use its output to give us additional insight into whether or not a shot had been taken. Hence, we opted to decompose the probability of taking a shot as

where \({\mathbb{P}}(\,{{\mbox{receiver}}}| {{\mbox{corner}}}\,)\) are the probabilities computed by TacticAI’s receiver prediction system, and \({\mathbb{P}}(\,{{\mbox{shot}}}| {{\mbox{receiver}}},{{\mbox{corner}}}\,)\) models the conditional shot probability after a specific player makes first contact with the ball. This was implemented through providing an additional global feature to indicate the receiver in the corresponding corner kick (Table  1 ) while the architecture otherwise remained the same as that of receiver prediction (Supplementary Fig.  2 , see the “Methods” section). At training time, we feed the ground-truth receiver as input to the model—at inference time, we attempt every possible receiver, weighing their contributions using the probabilities given by TacticAI’s receiver predictor, as per Eq. ( 1 ). This two-phased approach yielded a final test F 1 score of 0.68 ± 0.04 for shot prediction, which encodes significantly more signal than the unconditional shot predictor, especially considering the many unobservables associated with predicting shot events. Just as for receiver prediction, this performance can be further improved using geometric deep learning; a conditional GATv2 shot predictor with D 2 group convolutions achieves an F 1 score of 0.71 ± 0.01.

Moreover, we also observe that, even just through predicting the receivers, without explicitly classifying any other salient features of corners, TacticAI learned generalisable representations of the data. Specifically, team setups with similar tactical patterns tend to cluster together in TacticAI’s latent space (Fig.  2 ). However, no clear clusters are observed in the raw input space (Supplementary Fig.  1 ). This indicates that TacticAI can be leveraged as a useful corner kick retrieval system, and we will present our evaluation of this hypothesis in the “Case study with expert raters” section.

We visualise the latent representations of attacking and defending teams in 1024 corner kicks using t -SNE. A latent team embedding in one corner kick sample is the mean of the latent player representations on the same attacking ( A – C ) or defending ( D ) team. Given the reference corner kick sample ( A ), we retrieve another corner kick sample ( B ) with respect to the closest distance of their representations in the latent space. We observe that ( A ) and ( B ) are both out-swing corner kicks and share similar patterns of their attacking tactics, which are highlighted with rectangles having the same colours, although they bear differences with respect to the absolute positions and velocities of the players. All the while, the latent representation of an in-swing attack ( C ) is distant from both ( A ) and ( B ) in the latent space. The red arrows are only used to demonstrate the difference between in- and out-swing corner kicks, not the actual ball trajectories.

Lastly, it is worth emphasising that the utility of the shot predictor likely does not come from forecasting whether a shot event will occur—a challenging problem with many imponderables—but from analysing the difference in predicted shot probability across multiple corners. Indeed, in the following section, we will show how TacticAI’s generative tactic refinements can directly influence the predicted shot probabilities, which will then corresponds to highly favourable evaluation by our expert raters in the “Case study with expert raters” section.

Controlled tactic refinement using class-conditional generative models

Equipped with components that are able to potently relate corner kicks with their various outcomes (e.g. receivers and shot events), we can explore the use of TacticAI to suggest adjustments of tactics, in order to amplify or reduce the likelihood of certain outcomes.

Specifically, we aim to produce adjustments to the movements of players on one of the two teams, including their positions and velocities, which would maximise or minimise the probability of a shot event, conditioned on the initial corner setup, consisting of the movements of players on both teams and their heights and weights. In particular, although in real-world scenarios both teams may react simultaneously to the movements of each other, in our study, we focus on moderate adjustments to player movements, which help to detect players that are not responding to a tactic properly. Due to this reason, we simplify the process of tactic refinement through generating the adjustments for only one team while keeping the other fixed. The way we train a model for this task is through an auto-encoding objective: we feed the ground-truth shot outcome (a binary indicator) as an additional graph-level feature to TacticAI’s model (Table  1 ), and then have it learn to reconstruct a probability distribution of the input player coordinates (Fig.  1 B, also see the “Methods” section). As a consequence, our tactic adjustment system does not depend on the previously discussed shot predictor—although we can use the shot predictor to evaluate whether the adjustments make a measurable difference in shot probability.

This autoencoder-based generative model is an individual component that separates from TacticAI’s predictive systems. All three systems share the encoder architecture (without sharing parameters), but use different decoders (see the “Methods” section). At inference time, we can instead feed in a desired shot outcome for the given corner setup, and then sample new positions and velocities for players on one team using this probability distribution. This setup, in principle, allows for flexible downstream use, as human coaches can optimise corner kick setups through generating adjustments conditioned on the specific outcomes of their interest—e.g., increasing shot probability for the attacking team, decreasing it for the defending team (Fig.  3 ) or amplifying the chance that a particular striker receives the ball.

TacticAI makes it possible for human coaches to redesign corner kick tactics in ways that help maximise the probability of a positive outcome for either the attacking or the defending team by identifying key players, as well as by providing temporally coordinated tactic recommendations that take all players into consideration. As demonstrated in the present example ( A ), for a corner kick in which there was a shot attempt in reality ( B ), TacticAI can generate a tactically-adjusted setting in which the shot probability has been reduced, by adjusting the positioning of the defenders ( D ). The suggested defender positions result in reduced receiver probability for attacking players 2–5 (see bottom row), while the receiver probability of Attacker 1, who is distant from the goalpost, has been increased ( C ). The model is capable of generating multiple such scenarios. Coaches can inspect the different options visually and additionally consult TacticAI’s quantitative analysis of the presented tactics.

We first evaluate the generated adjustments quantitatively, by verifying that they are indistinguishable from the original corner kick distribution using a classifier. To do this, we synthesised a dataset consisting of 200 corner kick samples and their corresponding conditionally generated adjustments. Specifically, for corners without a shot event, we generated adjustments for the attacking team by setting the shot event feature to 1, and vice-versa for the defending team when a shot event did happen. We found that the real and generated samples were not distinguishable by an MLP classifier, with an F 1 score of 0.53 ± 0.05, indicating random chance level accuracy. This result indicates that the adjustments produced by TacticAI are likely similar enough to real corner kicks that the MLP is unable to tell them apart. Note that, in spite of this similarity, TacticAI recommends player-level adjustments that are not negligible—in the following section we will illustrate several salient examples of this. To more realistically validate the practical indistinguishability of TacticAI’s adjustments from realistic corners, we also evaluated the realism of the adjustments in a case study with human experts, which we will present in the following section.

In addition, we leveraged our TacticAI shot predictor to estimate whether the proposed adjustments were effective. We did this by analysing 100 corner kick samples in which threatening shots occurred, and then, for each sample, generated one defensive refinement through setting the shot event feature to 0. We observed that the average shot probability significantly decreased, from 0.75 ± 0.14 for ground-truth corners to 0.69 ± 0.16 for adjustments ( z  = 2.62,  p  < 0.001). This observation was consistent when testing for attacking team refinements (shot probability increased from 0.18 ± 0.16 to 0.31 ± 0.26 ( z  = −4.46,  p  < 0.001)). Moving beyond this result, we also asked human raters to assess the utility of TacticAI’s proposed adjustments within our case study, which we detail next.

Case study with expert raters

Although quantitative evaluation with well-defined benchmark datasets was critical for the technical development of TacticAI, the ultimate test of TacticAI as a football tactic assistant is its practical downstream utility being recognised by professionals in the industry. To this end, we evaluated TacticAI through a case study with our partners at Liverpool FC (LFC). Specifically, we invited a group of five football experts: three data scientists, one video analyst, and one coaching assistant. Each of them completed four tasks in the case study, which evaluated the utility of TacticAI’s components from several perspectives; these include (1) the realism of TacticAI’s generated adjustments, (2) the plausibility of TacticAI’s receiver predictions, (3) effectiveness of TacticAI’s embeddings for retrieving similar corners, and (4) usefulness of TacticAI’s recommended adjustments. We provide an overview of our study’s results here and refer the interested reader to Supplementary Figs.  3 – 5 and the  Supplementary Methods for additional details.

We first simultaneously evaluated the realism of the adjusted corner kicks generated by TacticAI, and the plausibility of its receiver predictions. Going through a collection of 50 corner kick samples, we first asked the raters to classify whether a given sample was real or generated by TacticAI, and then they were asked to identify the most likely receivers in the corner kick sample (Supplementary Fig.  3 ).

On the task of classifying real and generated samples, first, we found that the raters’ average F 1 score of classifying the real vs. generated samples was only 0.60 ± 0.04, with individual F 1 scores ( \({F}_{1}^{A}=0.54,{F}_{1}^{B}=0.64,{F}_{1}^{C}=0.65,{F}_{1}^{D}=0.62,{F}_{1}^{E}=0.56\) ), indicating that the raters were, in many situations, unable to distinguish TacticAI’s adjustments from real corners.

The previous evaluation focused on analysing realism detection performance across raters. We also conduct a study that analyses realism detection across samples. Specifically, we assigned ratings for each sample—assigning +1 to a sample if it was identified as real by a human rater, and 0 otherwise—and computed the average rating for each sample across the five raters. Importantly, by studying the distribution of ratings, we found that there was no significant difference between the average ratings assigned to real and generated corners ( z  = −0.34,  p  > 0.05) (Fig.  4 A). Hence, the real and generated samples were assigned statistically indistinguishable average ratings by human raters.

In task 1, we tested the statistical difference between the real corner kick samples and the synthetic ones generated by TacticAI from two aspects: ( A.1 ) the distributions of their assigned ratings, and ( A.2 ) the corresponding histograms of the rating values. Analogously, in task 2 (receiver prediction), ( B.1 ) we track the distributions of the top-3 accuracy of receiver prediction using those samples, and ( B.2 ) the corresponding histogram of the mean rating per sample. No statistical difference in the mean was observed in either cases (( A.1 ) ( z  = −0.34,  p  > 0.05), and ( B.1 ) ( z  = 0.97,  p  > 0.05)). Additionally, we observed a statistically significant difference between the ratings of different raters on receiver prediction, with three clear clusters emerging ( C ). Specifically, Raters A and E had similar ratings ( z  = 0.66,  p  > 0.05), and Raters B and D also rated in similar ways ( z  = −1.84,  p  > 0.05), while Rater C responded differently from all other raters. This suggests a good level of variety of the human raters with respect to their perceptions of corner kicks. In task 3—identifying similar corners retrieved in terms of salient strategic setups—there were no significant differences among the distributions of the ratings by different raters ( D ), suggesting a high level of agreement on the usefulness of TacticAI’s capability of retrieving similar corners ( F 1,4  = 1.01,  p  > 0.1). Finally, in task 4, we compared the ratings of TacticAI’s strategic refinements across the human raters ( E ) and found that the raters also agreed on the general effectiveness of the refinements recommended by TacticAI ( F 1,4  = 0.45,  p  > 0.05). Note that the violin plots used in B.1 and C – E model a continuous probability distribution and hence assign nonzero probabilities to values outside of the allowed ranges. We only label y -axis ticks for the possible set of ratings.

For the task of identifying receivers, we rated TacticAI’s predictions with respect to a rater as +1 if at least one of the receivers identified by the rater appeared in TacticAI’s top-3 predictions, and 0 otherwise. The average top-3 accuracy among the human raters was 0.79 ± 0.18; specifically, 0.81 ± 0.17 for the real samples, and 0.77 ± 0.21 for the generated ones. These scores closely line up with the accuracy of TacticAI in predicting receivers for held-out test corners, validating our quantitative study. Further, after averaging the ratings for receiver prediction sample-wise, we found no statistically significant difference between the average ratings of predicting receivers over the real and generated samples ( z  = 0.97,  p  > 0.05) (Fig.  4 B). This indicates that TacticAI was equally performant in predicting the receivers of real corners and TacticAI-generated adjustments, and hence may be leveraged for this purpose even in simulated scenarios.

There is a notably high variance in the average receiver prediction rating of TacticAI. We hypothesise that this is due to the fact that different raters may choose to focus on different salient features when evaluating the likely receivers (or even the amount of likely receivers). We set out to validate this hypothesis by testing the pair-wise similarity of the predictions by the human raters through running a one-away analysis of variance (ANOVA), followed by a Tukey test. We found that the distributions of the five raters’ predictions were significantly different ( F 1,4  = 14.46,  p  < 0.001) forming three clusters (Fig.  4 C). This result indicates that different human raters—as suggested by their various titles at LFC—may often use very different leads when suggesting plausible receivers. The fact that TacticAI manages to retain a high top-3 accuracy in such a setting suggests that it was able to capture the salient patterns of corner kick strategies, which broadly align with human raters’ preferences. We will further test this hypothesis in the third task—identifying similar corners.

For the third task, we asked the human raters to judge 50 pairs of corners for their similarity. Each pair consisted of a reference corner and a retrieved corner, where the retrieved corner was chosen either as the nearest-neighbour of the reference in terms of their TacticAI latent space representations, or—as a feature-level heuristic—the cosine similarities of their raw features (Supplementary Fig.  4 ) in our corner kick dataset. We score the raters’ judgement of a pair as +1 if they considered the corners presented in the case to be usefully similar, otherwise, the pair is scored with 0. We first computed, for each rater, the recall with which they have judged a baseline- or TacticAI-retrieved pair as usefully similar—see description of Task 3 in the  Supplementary Methods . For TacticAI retrievals, the average recall across all raters was 0.59 ± 0.09, and for the baseline system, the recall was 0.36 ± 0.10. Secondly, we assess the statistical difference between the results of the two methods by averaging the ratings for each reference–retrieval pair, finding that the average rating of TacticAI retrievals is significantly higher than the average rating of baseline method retrievals ( z  = 2.34,  p  < 0.05). These two results suggest that TacticAI significantly outperforms the feature-space baseline as a method for mining similar corners. This indicates that TacticAI is able to extract salient features from corners that are not trivial to extract from the input data alone, reinforcing it as a potent tool for discovering opposing team tactics from available data. Finally, we observed that this task exhibited a high level of inter-rater agreement for TacticAI-retrieved pairs ( F 1,4  = 1.01,  p  > 0.1) (Fig.  4 D), suggesting that human raters were largely in agreement with respect to their assessment of TacticAI’s performance.

Finally, we evaluated TacticAI’s player adjustment recommendations for their practical utility. Specifically, each rater was given 50 tactical refinements together with the corresponding real corner kick setups—see Supplementary Fig.  5 , and the “Case study design” section in the  Supplementary Methods . The raters were then asked to rate each refinement as saliently improving the tactics (+1), saliently making them worse (−1), or offering no salient differences (0). We calculated the average rating assigned by each of the raters (giving us a value in the range [− 1, 1] for each rater). The average of these values across all five raters was 0.7 ± 0.1. Further, for 45 of the 50 situations (90%), the human raters found TacticAI’s suggestion to be favourable on average (by majority voting). Both of these results indicate that TacticAI’s recommendations are salient and useful to a downstream football club practitioner, and we set out to validate this with statistical tests.

We performed statistical significance testing of the observed positive ratings. First, for each of the 50 situations, we averaged its ratings across all five raters and then ran a t -test to assess whether the mean rating was significantly larger than zero. Indeed, the statistical test indicated that the tactical adjustments recommended by TacticAI were constructive overall ( \({t}_{49}^{{{{{{{{\rm{avg}}}}}}}}}=9.20,\, p \, < \, 0.001\) ). Secondly, we verified that each of the five raters individually found TacticAI’s recommendations to be constructive, running a t -test on each of their ratings individually. For all of the five raters, their average ratings were found to be above zero with statistical significance ( \({t}_{49}^{A}=5.84,\, {p}^{A} \, < \, 0.001;{t}_{49}^{B}=7.88,\; {p}^{B} \, < \, 0.001;{t}_{49}^{C}=7.00,\; {p}^{C} \, < \, 0.001;{t}_{49}^{D}=6.04,\; {p}^{D} \, < \, 0.001;{t}_{49}^{E}=7.30,\, {p}^{E} \, < \, 0.001\) ). In addition, their ratings also shared a high level of inter-agreement ( F 1,4  = 0.45,  p  > 0.05) (Fig.  4 E), suggesting a level of practical usefulness that is generally recognised by human experts, even though they represent different backgrounds.

Taking all of these results together, we find TacticAI to possess strong components for prediction, retrieval, and tactical adjustments on corner kicks. To illustrate the kinds of salient recommendations by TacticAI, in Fig.  5 we present four examples with a high degree of inter-rater agreement.

These examples are selected from our case study with human experts, to illustrate the breadth of tactical adjustments that TacticAI suggests to teams defending a corner. The density of the yellow circles coincides with the number of times that the corresponding change is recognised as constructive by human experts. Instead of optimising the movement of one specific player, TacticAI can recommend improvements for multiple players in one generation step through suggesting better positions to block the opposing players, or better orientations to track them more efficiently. Some specific comments from expert raters follow. In A , according to raters, TacticAI suggests more favourable positions for several defenders, and improved tracking runs for several others—further, the goalkeeper is positioned more deeply, which is also beneficial. In B , TacticAI suggests that the defenders furthest away from the corner make improved covering runs, which was unanimously deemed useful, with several other defenders also positioned more favourably. In C , TacticAI recommends improved covering runs for a central group of defenders in the penalty box, which was unanimously considered salient by our raters. And in D , TacticAI suggests substantially better tracking runs for two central defenders, along with a better positioning for two other defenders in the goal area.

We have demonstrated an AI assistant for football tactics and provided statistical evidence of its efficacy through a comprehensive case study with expert human raters from Liverpool FC. First, TacticAI is able to accurately predict the first receiver after a corner kick is taken as well as the probability of a shot as the direct result of the corner. Second, TacticAI has been shown to produce plausible tactical variations that improve outcomes in a salient way, while being indistinguishable from real scenarios by domain experts. And finally, the system’s latent player representations are a powerful means to retrieve similar set-piece tactics, allowing coaches to analyse relevant tactics and counter-tactics that have been successful in the past.

The broader scope of strategy modelling in football has previously been addressed from various individual angles, such as pass prediction 14 , 15 , 16 , shot prediction 3 or corner kick tactical classification 7 . However, to the best of our knowledge, our work stands out by combining and evaluating predictive and generative modelling of corner kicks for tactic development. It also stands out in its method of applying geometric deep learning, allowing for efficiently incorporating various symmetries of the football pitch for improved data efficiency. Our method incorporates minimal domain knowledge and does not rely on intricate feature engineering—though its factorised design naturally allows for more intricate feature engineering approaches when such features are available.

Our methodology requires the position and velocity estimates of all players at the time of execution of the corner and subsequent events. Here, we derive these from high-quality tracking and event data, with data availability from tracking providers limited to top leagues. Player tracking based on broadcast video would increase the reach and training data substantially, but would also likely result in noisier model inputs. While the attention mechanism of GATs would allow us to perform introspection of the most salient factors contributing to the model outcome, our method does not explicitly model exogenous (aleatoric) uncertainty, which would be valuable context for the football analyst.

While the empirical study of our method’s efficacy has been focused on corner kicks in association football, it readily generalises to other set pieces (such as throw-ins, which similarly benefit from similarity retrieval, pass and/or shot prediction) and other team sports with suspended play situations. The learned representations and overall framing of TacticAI also lay the ground for future research to integrate a natural language interface that enables domain-grounded conversations with the assistant, with the aim to retrieve particular situations of interest, make predictions for a given tactical variant, compare and contrast, and guide through an interactive process to derive tactical suggestions. It is thus our belief that TacticAI lays the groundwork for the next-generation AI assistant for football.

We devised TacticAI as a geometric deep learning pipeline, further expanded in this section. We process labelled spatio-temporal football data into graph representations, and train and evaluate on benchmarking tasks cast as classification or regression. These steps are presented in sequence, followed by details on the employed computational architecture.

Raw corner kick data

The raw dataset consisted of 9693 corner kicks collected from the 2020–21, 2021–22, and 2022–23 (up to January 2023) Premier League seasons. The dataset was provided by Liverpool FC and comprises four separate data sources, described below.

Our primary data source is spatio-temporal trajectory frames (tracking data), which tracked all on-pitch players and the ball, for each match, at 25 frames per second. In addition to player positions, their velocities are derived from position data through filtering. For each corner kick, we only used the frame in which the kick is being taken as input information.

Secondly, we also leverage event stream data, which annotated the events or actions (e.g., passes, shots and goals) that have occurred in the corresponding tracking frames.

Thirdly, the line-up data for the corresponding games, which recorded the players’ profiles, including their heights, weights and roles, is also used.

Lastly, we have access to miscellaneous game data, which contains the game days, stadium information, and pitch length and width in meters.

Graph representation and construction

We assumed that we were provided with an input graph \({{{{{{{\mathcal{G}}}}}}}}=({{{{{{{\mathcal{V}}}}}}}},\,{{{{{{{\mathcal{E}}}}}}}})\) with a set of nodes \({{{{{{{\mathcal{V}}}}}}}}\) and edges \({{{{{{{\mathcal{E}}}}}}}}\subseteq {{{{{{{\mathcal{V}}}}}}}}\times {{{{{{{\mathcal{V}}}}}}}}\) . Within the context of football games, we took \({{{{{{{\mathcal{V}}}}}}}}\) to be the set of 22 players currently on the pitch for both teams, and we set \({{{{{{{\mathcal{E}}}}}}}}={{{{{{{\mathcal{V}}}}}}}}\times {{{{{{{\mathcal{V}}}}}}}}\) ; that is, we assumed all pairs of players have the potential to interact. Further analyses, leveraging more specific choices of \({{{{{{{\mathcal{E}}}}}}}}\) , would be an interesting avenue for future work.

Additionally, we assume that the graph is appropriately featurised. Specifically, we provide a node feature matrix, \({{{{{{{\bf{X}}}}}}}}\in {{\mathbb{R}}}^{| {{{{{{{\mathcal{V}}}}}}}}| \times k}\) , an edge feature tensor, \({{{{{{{\bf{E}}}}}}}}\in {{\mathbb{R}}}^{| {{{{{{{\mathcal{V}}}}}}}}| \times | {{{{{{{\mathcal{V}}}}}}}}| \times l}\) , and a graph feature vector, \({{{{{{{\bf{g}}}}}}}}\in {{\mathbb{R}}}^{m}\) . The appropriate entries of these objects provide us with the input features for each node, edge, and graph. For example, \({{{{{{{{\bf{x}}}}}}}}}_{u}\in {{\mathbb{R}}}^{k}\) would provide attributes of an individual player \(u\in {{{{{{{\mathcal{V}}}}}}}}\) , such as position, height and weight, and \({{{{{{{{\bf{e}}}}}}}}}_{uv}\in {{\mathbb{R}}}^{l}\) would provide the attributes of a particular pair of players \((u,\, v)\in {{{{{{{\mathcal{E}}}}}}}}\) , such as their distance, and whether they belong to the same team. The graph feature vector, g , can be used to store global attributes of interest to the corner kick, such as the game time, current score, or ball position. For a simplified visualisation of how a graph neural network would process such an input, refer to Fig.  1 A.

To construct the input graphs, we first aligned the four data sources with respect to their game IDs and timestamps and filtered out 2517 invalid corner kicks, for which the alignment failed due to missing data, e.g., missing tracking frames or event labels. This filtering yielded 7176 valid corner kicks for training and evaluation. We summarised the exact information that was used to construct the input graphs in Table  2 . In particular, other than player heights (measured in centimeters (cm)) and weights (measured in kilograms (kg)), the players were anonymous in the model. For the cases in which the player profiles were missing, we set their heights and weights to 180 cm and 75 kg, respectively, as defaults. In total, we had 385 such occurrences out of a total of 213,246( = 22 × 9693) during data preprocessing. We downscaled the heights and weights by a factor of 100. Moreover, for each corner kick, we zero-centred the positions of on-pitch players and normalised them onto a 10 m × 10 m pitch, and their velocities were re-scaled accordingly. For the cases in which the pitch dimensions were missing, we used a standard pitch dimension of 110 m × 63 m as default.

We summarised the grouping of the features in Table  1 . The actual features used in different benchmark tasks may differ, and we will describe this in more detail in the next section. To focus on modelling the high-level tactics played by the attacking and defending teams, other than a binary indicator for ball possession—which is 1 for the corner kick taker and 0 for all other players—no information of ball movement, neither positions nor velocities, was used to construct the input graphs. Additionally, we do not have access to the player’s vertical movement, therefore only information on the two-dimensional movements of each player is provided in the data. We do however acknowledge that such information, when available, would be interesting to consider in a corner kick outcome predictor, considering the prevalence of aerial battles in corners.

Benchmark tasks construction

TacticAI consists of three predictive and generative models, which also correspond to three benchmark tasks implemented in this study. Specifically, (1) Receiver prediction, (2) Threatening shot prediction, and (3) Guided generation of team positions and velocities (Table  1 ). The graphs of all the benchmark tasks used the same feature space of nodes and edges, differing only in the global features.

For all three tasks, our models first transform the node features to a latent node feature matrix, \({{{{{{{\bf{H}}}}}}}}={f}_{{{{{{{{\mathcal{G}}}}}}}}}({{{{{{{\bf{X}}}}}}}},\, {{{{{{{\bf{E}}}}}}}},\, {{{{{{{\bf{g}}}}}}}})\) , from which we could answer queries: either about individual players—in which case we learned a relevant classifier or regressor over the h u vectors (the rows of H )—or about the occurrence of a global event (e.g. shot taken)—in which case we classified or regressed over the aggregated player vectors, ∑ u h u . In both cases, the classifiers were trained using stochastic gradient descent over an appropriately chosen loss function, such as categorical cross-entropy for classifiers, and mean squared error for regressors.

For different tasks, we extracted the corresponding ground-truth labels from either the event stream data or the tracking data. Specifically, (1) We modelled receiver prediction as a node classification task and labelled the first player to touch the ball after the corner was taken as the target node. This player could be either an attacking or defensive player. (2) Shot prediction was modelled as graph classification. In particular, we considered a next-ball-touch action by the attacking team as a shot if it was a direct corner, a goal, an aerial, hit on the goalposts, a shot attempt saved by the goalkeeper, or missing target. This yielded 1736 corners labelled as a shot being taken, and 5440 corners labelled as a shot not being taken. (3) For guided generation of player position and velocities, no additional label was needed, as this model relied on a self-supervised reconstruction objective.

The entire dataset was split into training and evaluation sets with an 80:20 ratio through random sampling, and the same splits were used for all tasks.

Graph neural networks

The central model of TacticAI is the graph neural network (GNN) 9 , which computes latent representations on a graph by repeatedly combining them within each node’s neighbourhood. Here we define a node’s neighbourhood, \({{{{{{{{\mathcal{N}}}}}}}}}_{u}\) , as the set of all first-order neighbours of node u , that is, \({{{{{{{{\mathcal{N}}}}}}}}}_{u}=\{v\,| \,(v,\, u)\in {{{{{{{\mathcal{E}}}}}}}}\}\) . A single GNN layer then transforms the node features by passing messages between neighbouring nodes 17 , following the notation of related work 10 , and the implementation of the CLRS-30 benchmark baselines 18 :

where \(\psi :{{\mathbb{R}}}^{k}\times {{\mathbb{R}}}^{k}\times {{\mathbb{R}}}^{l}\times {{\mathbb{R}}}^{m}\to {{\mathbb{R}}}^{{k}^{{\prime} }}\) and \(\phi :{{\mathbb{R}}}^{k}\times {{\mathbb{R}}}^{{k}^{{\prime} }}\to {{\mathbb{R}}}^{{k}^{{\prime} }}\) are two learnable functions (e.g. multilayer perceptrons), \({{{{{{{{\bf{h}}}}}}}}}_{u}^{(t)}\) are the features of node u after t GNN layers, and ⨁ is any permutation-invariant aggregator, such as sum, max, or average. By definition, we set \({{{{{{{{\bf{h}}}}}}}}}_{u}^{(0)}={{{{{{{{\bf{x}}}}}}}}}_{u}\) , and iterate Eq. ( 2 ) for T steps, where T is a hyperparameter. Then, we let \({{{{{{{\bf{H}}}}}}}}={f}_{{{{{{{{\mathcal{G}}}}}}}}}({{{{{{{\bf{X}}}}}}}},\, {{{{{{{\bf{E}}}}}}}},\, {{{{{{{\bf{g}}}}}}}})={{{{{{{{\bf{H}}}}}}}}}^{(T)}\) be the final node embeddings coming out of the GNN.

It is well known that Eq. ( 2 ) is remarkably general; it can be used to express popular models such as Transformers 19 as a special case, and it has been argued that all discrete deep learning models can be expressed in this form 20 , 21 . This makes GNNs a perfect framework for benchmarking various approaches to modelling player–player interactions in the context of football.

Different choices of ψ , ϕ and ⨁ yield different architectures. In our case, we utilise a message function that factorises into an attentional mechanism, \(a:{{\mathbb{R}}}^{k}\times {{\mathbb{R}}}^{k}\times {{\mathbb{R}}}^{l}\times {{\mathbb{R}}}^{m}\to {\mathbb{R}}\) :

yielding the graph attention network (GAT) architecture 12 . In our work, specifically, we use a two-layer multilayer perceptron for the attentional mechanism, as proposed by GATv2 11 :

where \({{{{{{{{\bf{W}}}}}}}}}_{1},\, {{{{{{{{\bf{W}}}}}}}}}_{2}\in {{\mathbb{R}}}^{k\times h}\) , \({{{{{{{{\bf{W}}}}}}}}}_{e}\in {{\mathbb{R}}}^{l\times h}\) , \({{{{{{{{\bf{W}}}}}}}}}_{g}\in {{\mathbb{R}}}^{m\times h}\) and \({{{{{{{\bf{a}}}}}}}}\in {{\mathbb{R}}}^{h}\) are the learnable parameters of the attentional mechanism, and LeakyReLU is the leaky rectified linear activation function. This mechanism computes coefficients of interaction (a single scalar value) for each pair of connected nodes ( u ,  v ), which are then normalised across all neighbours of u using the \({{{{{{{\rm{softmax}}}}}}}}\) function.

Through early-stage experimentation, we have ascertained that GATs are capable of matching the performance of more generic choices of ψ (such as the MPNN 17 ) while being more scalable. Hence, we focus our study on the GAT model in this work. More details can be found in the subsection “Ablation study” section.

Geometric deep learning

In spite of the power of Eq. ( 2 ), using it in its full generality is often prone to overfitting, given the large number of parameters contained in ψ and ϕ . This problem is exacerbated in the football analytics domain, where gold-standard data is generally very scarce—for example, in the English Premier League, only a few hundred games are played every season.

In order to tackle this issue, we can exploit the immense regularity of data arising from football games. Strategically equivalent game states are also called transpositions, and symmetries such as arriving at the same chess position through different move sequences have been exploited computationally since the 1960s 22 . Similarly, game rotations and reflections may yield equivalent strategic situations 23 . Using the blueprint of geometric deep learning (GDL) 10 , we can design specialised GNN architectures that exploit this regularity.

That is, geometric deep learning is a generic methodology for deriving mathematical constraints on neural networks, such that they will behave predictably when inputs are transformed in certain ways. In several important cases, these constraints can be directly resolved, directly informing neural network architecture design. For a comprehensive example of point clouds under 3D rotational symmetry, see Fuchs et al. 24 .

To elucidate several aspects of the GDL framework on a high level, let us assume that there exists a group of input data transformations (symmetries), \({\mathfrak{G}}\) under which the ground-truth label remains unchanged. Specifically, if we let y ( X ,  E ,  g ) be the label given to the graph featurised with X ,  E ,  g , then for every transformation \({\mathfrak{g}}\in {\mathfrak{G}}\) , the following property holds:

This condition is also referred to as \({\mathfrak{G}}\) -invariance. Here, by \({\mathfrak{g}}({{{{{{{\bf{X}}}}}}}})\) we denote the result of transforming X by \({\mathfrak{g}}\) —a concept also known as a group action. More generally, it is a function of the form \({\mathfrak{G}}\times {{{{{{{\mathcal{S}}}}}}}}\to {{{{{{{\mathcal{S}}}}}}}}\) for some state set \({{{{{{{\mathcal{S}}}}}}}}\) . Note that a single group element, \({\mathfrak{g}}\in {\mathfrak{G}}\) can easily produce different actions on different \({{{{{{{\mathcal{S}}}}}}}}\) —in this case, \({{{{{{{\mathcal{S}}}}}}}}\) could be \({{\mathbb{R}}}^{| {{{{{{{\mathcal{V}}}}}}}}| \times k}\) ( X ), \({{\mathbb{R}}}^{| {{{{{{{\mathcal{V}}}}}}}}| \times | {{{{{{{\mathcal{V}}}}}}}}| \times l}\) ( E ) and \({{\mathbb{R}}}^{m}\) ( g ).

It is worth noting that GNNs may also be derived using a GDL perspective if we set the symmetry group \({\mathfrak{G}}\) to \({S}_{| {{{{{{{\mathcal{V}}}}}}}}}|\) , the permutation group of \(| {{{{{{{\mathcal{V}}}}}}}}|\) objects. Owing to the design of Eq. ( 2 ), its outputs will not be dependent on the exact permutation of nodes in the input graph.

Frame averaging

A simple mechanism to enforce \({\mathfrak{G}}\) -invariance, given any predictor \({f}_{{{{{{{{\mathcal{G}}}}}}}}}({{{{{{{\bf{X}}}}}}}},\, {{{{{{{\bf{E}}}}}}}},\, {{{{{{{\bf{g}}}}}}}})\) , performs frame averaging across all \({\mathfrak{G}}\) -transformed inputs:

This ensures that all \({\mathfrak{G}}\) -transformed versions of a particular input (also known as that input’s orbit) will have exactly the same output, satisfying Eq. ( 5 ). A variant of this approach has also been applied in the AlphaGo architecture 25 to encode symmetries of a Go board.

In our specific implementation, we set \({\mathfrak{G}}={D}_{2}=\{{{{{{{{\rm{id}}}}}}}},\leftrightarrow,\updownarrow,\leftrightarrow \updownarrow \}\) , the dihedral group. Exploiting D 2 -invariance allows us to encode quadrant symmetries. Each element of the D 2 group encodes the presence of vertical or horizontal reflections of the input football pitch. Under these transformations, the pitch is assumed completely symmetric, and hence many predictions, such as which player receives the corner kick, or takes a shot from it, can be safely assumed unchanged. As an example of how to compute transformed features in Eq. ( 6 ), ↔( X ) horizontally reflects all positional features of players in X (e.g. the coordinates of the player), and negates the x -axis component of their velocity.

Group convolutions

While the frame averaging approach of Eq. ( 6 ) is a powerful way to restrict GNNs to respect input symmetries, it arguably misses an opportunity for the different \({\mathfrak{G}}\) -transformed views to interact while their computations are being performed. For small groups such as D 2 , a more fine-grained approach can be assumed, operating over a single GNN layer in Eq. ( 2 ), which we will write shortly as \({{{{{{{{\bf{H}}}}}}}}}^{(t)}={g}_{{{{{{{{\mathcal{G}}}}}}}}}({{{{{{{{\bf{H}}}}}}}}}^{(t-1)},\, {{{{{{{\bf{E}}}}}}}},\, {{{{{{{\bf{g}}}}}}}})\) . The condition that we need a symmetry-respecting GNN layer to satisfy is as follows, for all transformations \({\mathfrak{g}}\in {\mathfrak{G}}\) :

that is, it does not matter if we apply \({\mathfrak{g}}\) it to the input or the output of the function \({g}_{{{{{{{{\mathcal{G}}}}}}}}}\) —the final answer is the same. This condition is also referred to as \({\mathfrak{G}}\) -equivariance, and it has recently proved to be a potent paradigm for developing powerful GNNs over biochemical data 24 , 26 .

To satisfy D 2 -equivariance, we apply the group convolution approach 13 . Therein, views of the input are allowed to directly interact with their \({\mathfrak{G}}\) -transformed variants, in a manner very similar to grid convolutions (which is, indeed, a special case of group convolutions, setting \({\mathfrak{G}}\) to be the translation group). We use \({{{{{{{{\bf{H}}}}}}}}}_{{\mathfrak{g}}}^{(t)}\) to denote the \({\mathfrak{g}}\) -transformed view of the latent node features at layer t . Omitting E and g inputs for brevity, and using our previously designed layer \({g}_{{{{{{{{\mathcal{G}}}}}}}}}\) as a building block, we can perform a group convolution as follows:

Here, ∥ is the concatenation operation, joining the two node feature matrices column-wise; \({{\mathfrak{g}}}^{-1}\) is the inverse transformation to \({\mathfrak{g}}\) (which must exist as \({\mathfrak{G}}\) is a group); and \({{\mathfrak{g}}}^{-1}{\mathfrak{h}}\) is the composition of the two transformations.

Effectively, Eq. ( 8 ) implies our D 2 -equivariant GNN needs to maintain a node feature matrix \({{{{{{{{\bf{H}}}}}}}}}_{{\mathfrak{g}}}^{(t)}\) for every \({\mathfrak{G}}\) -transformation of the current input, and these views are recombined by invoking \({g}_{{{{{{{{\mathcal{G}}}}}}}}}\) on all pairs related together by applying a transformation \({\mathfrak{h}}\) . Note that all reflections are self-inverses, hence, in D 2 , \({\mathfrak{g}}={{\mathfrak{g}}}^{-1}\) .

It is worth noting that both the frame averaging in Eq. ( 6 ) and group convolution in Eq. ( 8 ) are similar in spirit to data augmentation. However, whereas standard data augmentation would only show one view at a time to the model, a frame averaging/group convolution architecture exhaustively generates all views and feeds them to the model all at once. Further, group convolutions allow these views to explicitly interact in a way that does not break symmetries. Here lies the key difference between the two approaches: frame averaging and group convolutions rigorously enforce the symmetries in \({\mathfrak{G}}\) , whereas data augmentation only provides implicit hints to the model about satisfying them. As a consequence of the exhaustive generation, Eqs. ( 6 ) and ( 8 ) are only feasible for small groups like D 2 . For larger groups, approaches like Steerable CNNs 27 may be employed.

Network architectures

While the three benchmark tasks we are performing have minor differences in the global features available to the model, the neural network models designed for them all have the same encoder–decoder architecture. The encoder has the same structure in all tasks, while the decoder model is tailored to produce appropriately shaped outputs for each benchmark task.

Given an input graph, TacticAI’s model first generates all relevant D 2 -transformed versions of it, by appropriately reflecting the player coordinates and velocities. We refer to the original input graph as the identity view, and the remaining three D 2 -transformed graphs as reflected views.

Once the views are prepared, we apply four group convolutional layers (Eq. ( 8 )) with a GATv2 base model (Eqs. ( 3 ) and ( 4 )) as the \({g}_{{{{{{{{\mathcal{G}}}}}}}}}\) function. Specifically, this means that, in Eqs. ( 3 ) and ( 4 ), every instance of \({{{{{{{{\bf{h}}}}}}}}}_{u}^{(t-1)}\) is replaced by the concatenation of \({({{{{{{{{\bf{h}}}}}}}}}_{{\mathfrak{h}}}^{(t-1)})}_{u}\parallel {({{{{{{{{\bf{h}}}}}}}}}_{{{\mathfrak{g}}}^{-1}{\mathfrak{h}}}^{(t-1)})}_{u}\) . Each GATv2 layer has eight attention heads and computes four latent features overall per player. Accordingly, once the four group convolutions are performed, we have a representation of \({{{{{{{\bf{H}}}}}}}}\in {{\mathbb{R}}}^{4\times 22\times 4}\) , where the first dimension corresponds to the four views ( \({{{{{{{{\bf{H}}}}}}}}}_{{{{{{{{\rm{id}}}}}}}}},\, {{{{{{{{\bf{H}}}}}}}}}_{\leftrightarrow },\, {{{{{{{{\bf{H}}}}}}}}}_{\updownarrow },\, {{{{{{{{\bf{H}}}}}}}}}_{\leftrightarrow \updownarrow }\in {{\mathbb{R}}}^{22\times 4}\) ), the second dimension corresponds to the players (eleven on each team), and the third corresponds to the 4-dimensional latent vector for each player node in this particular view. How this representation is used by the decoder depends on the specific downstream task, as we detail below.

For receiver prediction, which is a fully invariant function (i.e. reflections do not change the receiver), we perform simple frame averaging across all views, arriving at

and then learn a node-wise classifier over the rows of \({{{{{{{{\bf{H}}}}}}}}}^{{{{{{{{\rm{node}}}}}}}}}\in {{\mathbb{R}}}^{22\times 4}\) . We further decode H node into a logit vector \({{{{{{{\bf{O}}}}}}}}\in {{\mathbb{R}}}^{22}\) with a linear layer before computing the corresponding softmax cross entropy loss.

For shot prediction, which is once again fully invariant (i.e. reflections do not change the probability of a shot), we can further average the frame-averaged features across all players to get a global graph representation:

and then learn a binary classifier over \({{{{{{{{\bf{h}}}}}}}}}^{{{{{{{{\rm{graph}}}}}}}}}\in {{\mathbb{R}}}^{4}\) . Specifically, we decode the hidden vector into a single logit with a linear layer and compute the sigmoid binary cross-entropy loss with the corresponding label.

For guided generation (position/velocity adjustments), we generate the player positions and velocities with respect to a particular outcome of interest for the human coaches, predicted over the rows of the hidden feature matrix. For example, the model may adjust the defensive setup to decrease the shot probability by the attacking team. The model output is now equivariant rather than invariant—reflecting the pitch appropriately reflects the predicted positions and velocity vectors. As such, we cannot perform frame averaging, and take only the identity view’s features, \({{{{{{{{\bf{H}}}}}}}}}_{{{{{{{{\rm{id}}}}}}}}}\in {{\mathbb{R}}}^{22\times 4}\) . From this latent feature matrix, we can then learn a conditional distribution from each row, which models the positions or velocities of the corresponding player. To do this, we extend the backbone encoder with conditional variational autoencoder (CVAE 28 , 29 ). Specifically, for the u -th row of H id , h u , we first map its latent embedding to the parameters of a two-dimensional Gaussian distribution \({{{{{{{\mathcal{N}}}}}}}}({\mu }_{u}| {\sigma }_{u})\) , and then sample the coordinates and velocities from this distribution. At training time, we can efficiently propagate gradients through this sampling operation using the reparameterisation trick 28 : sample a random value \({\epsilon }_{u} \sim {{{{{{{\mathcal{N}}}}}}}}(0,1)\) for each player from the unit Gaussian distribution, and then treat μ u  +  σ u ϵ u as the sample for this player. In what follows, we omit edge features for brevity. For each corner kick sample X with the corresponding outcome o (e.g. a binary value indicating a shot event), we extend the standard VAE loss 28 , 29 to our case of outcome-conditional guided generation as

where h u is the player embedding corresponding to the u th row of H id , and \({\mathbb{KL}}\) is Kullback–Leibler (KL) divergence. Specifically, the first term is the generation loss between the real player input x u and the reconstructed sample decoded from h u with the decoder p ϕ . Using the KL term, the distribution of the latent embedding h u is regularised towards p ( h u ∣ o ), which is a multivariate Gaussian in our case.

A complete high-level summary of the generic encoder–decoder equivariant architecture employed by TacticAI can be summarised in Supplementary Fig.  2 . In the following section, we will provide empirical evidence for justifying these architectural decisions. This will be done through targeted ablation studies on our predictive benchmarks (receiver prediction and shot prediction).

Ablation study

We leveraged the receiver prediction task as a way to evaluate various base model architectures, and directly quantitatively assess the contributions of geometric deep learning in this context. We already see that the raw corner kick data can be better represented through geometric deep learning, yielding separable clusters in the latent space that could correspond to different attacking or defending tactics (Fig.  2 ). In addition, we hypothesise that these representations can also yield better performance on the task of receiver prediction. Accordingly, we ablate several design choices using deep learning on this task, as illustrated by the following four questions:

Does a factorised graph representation help? To assess this, we compare it against a convolutional neural network (CNN 30 ) baseline, which does not leverage a graph representation.

Does a graph structure help? To assess this, we compare against a Deep Sets 31 baseline, which only models each node in isolation without considering adjacency information—equivalently, setting each neighbourhood \({{{{{{{{\mathcal{N}}}}}}}}}_{u}\) to a singleton set { u }.

Are attentional GNNs a good strategy? To assess this, we compare against a message passing neural network 32 , MPNN baseline, which uses the fully potent GNN layer from Eq. ( 2 ) instead of the GATv2.

Does accounting for symmetries help? To assess this, we compare our geometric GATv2 baseline against one which does not utilise D 2 group convolutions but utilises D 2 frame averaging, and one which does not explicitly utilise any aspect of D 2 symmetries at all.

Each of these models has been trained for a fixed budget of 50,000 training steps. The test top- k receiver prediction accuracies of the trained models are provided in Supplementary Table  2 . As already discussed in the section “Results”, there is a clear advantage to using a full graph structure, as well as directly accounting for reflection symmetry. Further, the usage of the MPNN layer leads to slight overfitting compared to the GATv2, illustrating how attentional GNNs strike a good balance of expressivity and data efficiency for this task. Our analysis highlights the quantitative benefits of both graph representation learning and geometric deep learning for football analytics from tracking data. We also provide a brief ablation study for the shot prediction task in Supplementary Table  3 .

Training details

We train each of TacticAI’s models in isolation, using NVIDIA Tesla P100 GPUs. To minimise overfitting, each model’s learning objective is regularised with an L 2 norm penalty with respect to the network parameters. During training, we use the Adam stochastic gradient descent optimiser 33 over the regularised loss.

All models, including baselines, have been given an equal hyperparameter tuning budget, spanning the number of message passing steps ({1, 2, 4}), initial learning rate ({0.0001, 0.00005}), batch size ({128, 256}) and L 2 regularisation coefficient ({0.01, 0.005, 0.001, 0.0001, 0}). We summarise the chosen hyperparameters of each TacticAI model in Supplementary Table  1 .

Data availability

The data collected in the human experiments in this study have been deposited in the Zenodo database under accession code https://zenodo.org/records/10557063 , and the processed data which is used in the statistical analysis and to generate the relevant figures in the main text are available under the same accession code. The input and output data generated and/or analysed during the current study are protected and are not available due to data privacy laws and licensing restrictions. However, contact details of the input data providers are available from the corresponding authors on reasonable request.

Code availability

All the core models described in this research were built with the Graph Neural Network processors provided by the CLRS Algorithmic Reasoning Benchmark 18 , and their source code is available at https://github.com/google-deepmind/clrs . We are unable to release our code for this work as it was developed in a proprietary context; however, the corresponding authors are open to answer specific questions concerning re-implementations on request. For general data analysis, we used the following freely available packages: numpy v1.25.2 , pandas v1.5.3 , matplotlib v3.6.1 , seaborn v0.12.2 and scipy v1.9.3 . Specifically, the code of the statistical analysis conducted in this study is available at https://zenodo.org/records/10557063 .

The International Football Association Board (IFAB). Laws of the Game (The International Football Association Board, 2023).

Tuyls, K. et al. Game plan: what AI can do for football, and what football can do for AI. J. Artif. Intell. Res. 71 , 41–88 (2021).

Article   Google Scholar  

Goka, R., Moroto, Y., Maeda, K., Ogawa, T. & Haseyama, M. Prediction of shooting events in soccer videos using complete bipartite graphs and players’ spatial–temporal relations. Sensors 23 , 4506 (2023).

Article   ADS   PubMed   PubMed Central   Google Scholar  

Omidshafiei, S. et al. Multiagent off-screen behavior prediction in football. Sci. Rep. 12 , 8638 (2022).

Article   ADS   CAS   PubMed   PubMed Central   Google Scholar  

Lang, S., Wild, R., Isenko, A. & Link, D. Predicting the in-game status in soccer with machine learning using spatiotemporal player tracking data. Sci. Rep. 12 , 16291 (2022).

Baccouche, M., Mamalet, F., Wolf, C., Garcia, C. & Baskurt, A. Action classification in soccer videos with long short-term memory recurrent neural networks. In International Conference on Artificial Neural Networks (eds Diamantaras, K., Duch, W. & Iliadis, L. S.) pages 154–159 (Springer, 2010).

Shaw, L. & Gopaladesikan, S. Routine inspection: a playbook for corner kicks. In Machine Learning and Data Mining for Sports Analytics: 7th International Workshop, MLSA 2020, Co-located with ECML/PKDD 2020 , Proceedings, Ghent, Belgium, September 14–18, 2020, Vol . 7 , 3–16 (Springer, 2020).

Araújo, D. & Davids, K. Team synergies in sport: theory and measures. Front. Psychol. 7 , 1449 (2016).

Article   PubMed   PubMed Central   Google Scholar  

Veličković, P. Everything is connected: graph neural networks. Curr. Opin. Struct. Biol. 79 , 102538 (2023).

Article   PubMed   Google Scholar  

Bronstein, M. M., Bruna, J., Cohen, T. & Veličković, P. Geometric deep learning: grids, groups, graphs, geodesics, and gauges. arXiv preprint arXiv:2104.13478 (2021).

Brody, S., Alon, U. & Yahav, E. How attentive are graph attention networks? In International Conference on Learning Representations (ICLR, 2022). https://openreview.net/forum?id=F72ximsx7C1 .

Veličković, P. et al. Graph attention networks. In International Conference on Learning Representations (ICLR, 2018). https://openreview.net/forum?id=rJXMpikCZ .

Cohen, T. & Welling, M. Group equivariant convolutional networks. In International Conference on Machine Learning (eds Balcan, M. F. & Weinberger, K. Q.) 2990–2999 (PMLR, 2016).

Honda, Y., Kawakami, R., Yoshihashi, R., Kato, K. & Naemura, T. Pass receiver prediction in soccer using video and players’ trajectories. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) 3502–3511 (2022). https://ieeexplore.ieee.org/document/9857310 .

Hubáček, O., Sourek, G. & Železný, F. Deep learning from spatial relations for soccer pass prediction. In MLSA@PKDD/ECML (eds Brefeld, U., Davis, J., Van Haaren, J. & Zimmermann, A.) Vol. 11330, (Lecture Notes in Computer Science, Springer, Cham, 2018).

Sanyal, S. Who will receive the ball? Predicting pass recipient in soccer videos. J Visual Commun. Image Represent. 78 , 103190 (2021).

Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O. & Dahl, G. E. Neural message passing for quantum chemistry. In International Conference on Machine Learning (Precup, D. & Teh, Y. W.) 1263–1272 (PMLR, 2017).

Veličković, P. et al. The CLRS algorithmic reasoning benchmark. In International Conference on Machine Learning (eds Chaudhuri, K. et al.) 22084–22102 (PMLR, 2022).

Vaswani, A. et al. Attention is all you need. In Advances in Neural Information Processing Systems (eds Guyon, I. et al.) Vol. 30 (Curran Associates, Inc., 2017).

Veličković, P. Message passing all the way up. In ICLR 2022 Workshop on Geometrical and Topological Representation Learning (GTRL, 2022). https://openreview.net/forum?id=Bc8GiEZkTe5 .

Baranwal, A., Kimon, F. & Aukosh, J. Optimality of message-passing architectures for sparse graphs. In Thirty-seventh Conference on Neural Information Processing Systems (2023). https://papers.nips.cc/paper_files/paper/2023/hash/7e991aa4cd2fdf0014fba2f000f542d0-Abstract-Conference.html .

Greenblatt, R. D., Eastlake III, D. E. & Crocker, S. D. The Greenblatt chess program. In Proc. Fall Joint Computer Conference , 14–16 , 801–810 (Association for Computing Machinery, 1967). https://dl.acm.org/doi/10.1145/1465611.1465715 .

Schijf, M., Allis, L. V. & Uiterwijk, J. W. Proof-number search and transpositions. ICGA J. 17 , 63–74 (1994).

Fuchs, F., Worrall, D., Fischer, V. & Welling, M. SE(3)-transformers: 3D roto-translation equivariant attention networks. Adv. Neural Inf. Process. Syst. 33 , 1970–1981 (2020).

Google Scholar  

Silver, D. et al. Mastering the game of Go with deep neural networks and tree search. Nature 529 , 484–489 (2016).

Article   ADS   CAS   PubMed   Google Scholar  

Satorras, V. G., Hoogeboom, E. & Welling, M. E ( n ) equivariant graph neural networks. In International Conference on Machine Learning (eds Meila, M. & Zhang, T.) 9323–9332 (PMLR, 2021).

Cohen, T. S. & Welling, M. Steerable CNNs. In International Conference on Learning Representations (ICLR, 2017). https://openreview.net/forum?id=rJQKYt5ll .

Kingma, D. P. & Welling, M. Auto-encoding variational Bayes. In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14–16, 2014, Conference Track Proceedings (ICLR, 2014). https://openreview.net/forum?id=33X9fd2-9FyZd .

Sohn, K., Lee, H. & Yan, X. Learning structured output representation using deep conditional generative models. In Advances in Neural Information Processing Systems (eds Cortes, C, Lawrence, N., Lee, D., Sugiyama, M. & Garnett, R.) Vol. 28 (Curran Associates, Inc., 2015).

Fernández, J. & Bornn, L. Soccermap: a deep learning architecture for visually-interpretable analysis in soccer. In Machine Learning and Knowledge Discovery in Databases. Applied Data Science and Demo Track: European Conference, ECML PKDD 2020, Ghent, Belgium, September 14–18, 2020, Proceedings, Part V (eds Dong, Y., Ifrim, G., Mladenić, D., Saunders, C. & Van Hoecke, S.) 491–506 (Springer, 2021).

Zaheer, M. et al. Deep sets. In Advances in Neural Information Processing Systems Vol. 30 (eds Guyon, I., et al.) (Curran Associates, Inc., 2017).

Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O. & Dahl, G. E. Neural message passing for quantum chemistry. In Proc. 34th International Conference on Machine Learning , Vol. 70 of Proceedings of Machine Learning Research, 6–11 Aug 2017 (eds Precup, D. & Whye Teh, Y) 1263–1272 (PMLR, 2017).

Kingma, G. E. & Ba, J. Adam: a method for stochastic optimization. In ICLR (Poster) , (eds Bengio, Y. & LeCun, Y.) (International Conference of Learning Representations (ICLR), 2015). https://openreview.net/forum?id=8gmWwjFyLj .

Download references

Acknowledgements

We gratefully acknowledge the support of James French, Timothy Waskett, Hans Leitert and Benjamin Hervey for their extensive efforts in analysing TacticAI’s outputs. Further, we are thankful to Kevin McKee, Sherjil Ozair and Beatrice Bevilacqua for useful technical discussions, and Marc Lanctôt and Satinder Singh for reviewing the paper prior to submission.

Author information

These authors contributed equally: Zhe Wang, Petar Veličković, Daniel Hennes.

Authors and Affiliations

Google DeepMind, 6-8 Handyside Street, London, N1C 4UZ, UK

Zhe Wang, Petar Veličković, Daniel Hennes, Nenad Tomašev, Laurel Prince, Michael Kaisers, Yoram Bachrach, Romuald Elie, Li Kevin Wenliang, Federico Piccinini, Jerome Connor, Yi Yang, Adrià Recasens, Mina Khan, Nathalie Beauguerlange, Pablo Sprechmann, Pol Moreno, Nicolas Heess & Demis Hassabis

Liverpool FC, AXA Training Centre, Simonswood Lane, Kirkby, Liverpool, L33 5XB, UK

William Spearman

Liverpool FC, Kirkby, UK

University of Alberta, Amii, Edmonton, AB, T6G 2E8, Canada

Michael Bowling

Google DeepMind, London, UK

You can also search for this author in PubMed   Google Scholar

Contributions

Z.W., D. Hennes, L.P. and K.T. coordinated and organised the research effort leading to this paper. P.V. and Z.W. developed the core TacticAI models. Z.W., W.S. and I.G. prepared the Premier League corner kick dataset used for training and evaluating these models. P.V., Z.W., D. Hennes and N.T. designed the case study with human experts and Z.W. and P.V. performed the qualitative evaluation and statistical analysis of its outcomes. Z.W., P.V., D. Hennes, N.T., L.P., M. Kaisers, Y.B., R.E., L.K.W., F.P., W.S., I.G., N.H., M.B., D. Hassabis and K.T. contributed to writing the paper and providing feedback on the final manuscript. J.C., Y.Y., A.R., M. Khan, N.B., P.S. and P.M. contributed valuable technical and implementation discussions throughout the work’s development.

Corresponding authors

Correspondence to Zhe Wang , Petar Veličković or Karl Tuyls .

Ethics declarations

Competing interests.

The authors declare no competing interests but the following competing interests: TacticAI was developed during the course of the Authors’ employment at Google DeepMind and Liverpool Football Club, as applicable to each Author.

Peer review

Peer review information.

Nature Communications thanks Rui Luo and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Supplementary information, peer review file, rights and permissions.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Wang, Z., Veličković, P., Hennes, D. et al. TacticAI: an AI assistant for football tactics. Nat Commun 15 , 1906 (2024). https://doi.org/10.1038/s41467-024-45965-x

Download citation

Received : 13 October 2023

Accepted : 08 February 2024

Published : 19 March 2024

DOI : https://doi.org/10.1038/s41467-024-45965-x

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

By submitting a comment you agree to abide by our Terms and Community Guidelines . If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

March 18, 2024

Too Much Trust in AI Poses Unexpected Threats to the Scientific Process

It’s vital to “keep humans in the loop” to avoid humanizing machine-learning models in research

By Lauren Leffer

Moor Studio/Getty Images

Machine-learning models are quickly becoming common tools in scientific research. These artificial intelligence systems are helping bioengineers discover new potential antibiotics , veterinarians interpret animals’ facial expressions , papyrologists read words on ancient scrolls , mathematicians solve baffling problems and climatologists predict sea-ice movements . Some scientists are even probing large language models’ potential as proxies or replacements for human participants in psychology and behavioral research. In one recent example, computer scientists ran ChatGPT through the conditions of the Milgram shock experiment —the famous study on obedience in which people gave what they believed were increasingly painful electric shocks to an unseen person when told to do so by an authority figure—and other well-known psychology studies. The artificial intelligence model responded in a similar way as humans did —75 percent of simulated participants administered shocks of 300 volts and above.

But relying on these machine-learning algorithms also carries risks. Some of those risks are commonly acknowledged, such as generative AI’s tendency to spit out occasional “hallucinations” (factual inaccuracies or nonsense). Artificial intelligence tools can also replicate and even amplify human biases about characteristics such as race and gender. And the AI boom, which has given rise to complex, trillion-variable models, requires water- and energy-hungry data centers that likely have high environmental costs.

One big risk is less obvious, though potentially very consequential: humans tend to automatically attribute a great deal of authority and trust to machines. This misplaced faith could cause serious problems when AI systems are used for research , according to a paper published in early March in Nature .

On supporting science journalism

If you're enjoying this article, consider supporting our award-winning journalism by subscribing . By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.

“These tools are being anthropomorphized and framed as humanlike and superhuman. We risk inappropriately extending trust to the information produced by AI,” says the new paper’s co-author Molly Crockett , a cognitive psychologist and neuroscientist at Princeton University. AI models are human-made products, and they “represent the views and positions of the people who developed them,” says Lisa Messeri , a Yale University sociocultural anthropologist who worked with Crockett on the paper. Scientific American spoke with both researchers to learn more about the ways scientists use AI—and the potential effects of trusting this technology too much.

[ An edited transcript of the interview follows. ]

Why did you write this paper?

LISA MESSERI: [Crockett] and I started seeing and sharing all sorts of large, lofty promises of what AI could offer the scientific pipeline and scientific community. When we really started to think we needed to write something was when we saw claims that large language models could become substitutions for human subjects in research. These claims, given our years of conversation, seemed wrong-footed.

MOLLY CROCKETT: I have been using machine learning in my own research for several years, [and] advances in AI are enabling scientists to ask questions we couldn’t ask before. But, as I’ve been doing this research and observing that excitement among colleagues, I have developed a sense of uneasiness that’s been difficult to shake.

Beyond using large language models to replace human participants, how are scientists thinking about deploying AI?

CROCKETT: Previously we helped write a response to a study in [ Proceedings of the National Academy of Sciences USA ] that claimed machine learning could be used to predict whether research would [be replicable] just from the words in a paper.... That struck us as technically implausible. But more broadly, we’ve discovered that scientists are talking about using AI tools to make their work more objective and to be more productive.

We found that both of those goals are quite risky and open up scientists to producing more while understanding less. The worry is that we’re going to think that these tools are helping us to understand the world better, when in reality they might actually be distorting our view.

MESSERI: We categorize the AI uses we observed in our review into four categories: the Surrogate, the Oracle, the Quant and the Arbiter. The Surrogate is what we’ve already discussed—it replaces human subjects. The Oracle is an AI tool that is asked to synthesize the existing corpus of research and produce something, such as a review or new hypotheses. The Quant is AI that is used by scientists to process the intense amount of data out there—maybe produced by those machine surrogates. AI Arbiters are like [the tools described] in the [ PNAS ] replication study [Crockett] mentioned, tools for evaluating and adducting research. We call these visions for AI because they’re not necessarily being executed today in a successful or clean way, but they’re all being explored and proposed.

For each of these uses, you’ve pointed out that even if AI’s hallucinations and other technical problems are solved, risks remain. What are those risks?

CROCKETT: The overarching metaphor we use is this idea of monoculture, which comes from agriculture. Monocultures are very efficient. They improve productivity. But they’re vulnerable to being invaded by pests or disease; you’re more likely to lose the whole crop when you have a monoculture versus a diversity of what you’re growing. Scientific monocultures, too, are vulnerable to risks such as errors propagating throughout the whole system. This is especially the case with the foundation models in AI research, where one infrastructure is being used and applied across many domains. If there’s some error in that system, it can have widespread effects.

We identify two kinds of scientific monocultures that can arise with widespread AI adoption. The first is the monoculture of knowing. AI tools are only suited to answer certain kinds of questions. Because these tools boost productivity, the overall set of research questions being explored could become tailored to what AI is good at.

Then there’s the monoculture of the knower , where AI tools come to replace human thinkers. And because AI tools have a specific standpoint, this eliminates the diversity of different human perspectives from research production. When you have many different kinds of minds working on a scientific problem, you’re more likely to spot false assumptions or missed opportunities.

Both monocultures could lead to cognitive illusions.

What do you mean by illusions?

MESSERI: One example that’s already out there in psychology is the illusion of explanatory depth. Basically, when someone in your community claims they know something, you tend to assume you know that thing as well.

In your paper you cite research demonstrating that using a search engine can trick someone into believing they know something—when really they only have online access to that knowledge. And students who use AI assistant tools to respond to test questions end up thinking they understand a topic better than they do.

MESSERI: Exactly. Building off that one illusion of explanatory depth, we also identify two others. First, the illusion of exploratory breadth, where someone thinks they’re examining more than they are: There are an infinite number of questions we could ask about science and about the world. We worry that with the expansion of AI, the questions that AI is well suited to answer will be mistaken for the entire field of questions one could ask. Then there’s the risk of an illusion of objectivity. Either there’s an assumption that AI represents all standpoints or there’s an assumption that AI has no standpoint at all. But at the end of the day, AI tools are created by humans coming from a particular perspective.

How can scientists avoid falling into these traps? How can we mitigate these risks?

MESSERI: There’s the institutional level where universities and publishers dictate research. These institutions are developing partnerships with AI companies. We have to be very circumspect about the motivations behind that.... One mitigation strategy is just to be incredibly forthright about where the funding for AI is coming from and who benefits from the work being done on it.

CROCKETT: At the institutional level, funders, journal editors and universities can be mindful of developing a diverse portfolio of research to ensure that they’re not putting all the resources into research that uses a single-AI approach. In the future, it might be necessary to consciously protect resources for the kinds of research that can’t be addressed with AI tools.

And what sort of research is that?

CROCKETT: Well, as of right now, AI cannot think like a human. Any research about human thought and behavior, and also qualitative research, is not addressable with AI tools.

Would you say that in the worst-case scenario, AI poses an existential threat to human scientific knowledge production? Or is that an overstatement?

CROCKETT: I don’t think that it’s an overstatement. I think we are at a crossroads around how we decide what knowledge is and how we proceed in the endeavor of knowledge production.

Is there anything else you think is important for the public to really understand about what’s happening with AI and scientific research?

MESSERI: From the perspective of reading media coverage of AI, it seems as though this is some preordained, inevitable “evolution” of scientific and technical development. But as an anthropologist of science and technology, I would really like to emphasize that science and tech don’t proceed in an inevitable direction. It is always human-driven. These narratives of inevitability are themselves a product of human imagination and come from mistaking the desire by some to be a prophecy for all. Everyone, even nonscientists, can be part of questioning this narrative of inevitability by imagining the different futures that might come true instead.

CROCKETT: Being skeptical about AI in science doesn’t require being a hater of AI in science and technology. We love science. I’m excited about AI and its potential for science. But just because an AI tool is being used in science does not mean that it is automatically better science.

As scientists, we are trained to deny our humanness. We’re trained that human experience, bias and opinion have no place in the scientific method. The future of autonomous, AI “self-driving” labs is the pinnacle of realizing that sort of training. But increasingly we are seeing evidence that diversity of thought, experience and training in humans that do the science is vital for producing robust, innovative and creative knowledge. We don’t want to lose that. To keep the vitality of scientific knowledge production, we need to keep humans in the loop.

Advertisement

A strategic framework for artificial intelligence in marketing

• Conceptual/Theoretical Paper
• Open access
• Published: 04 November 2020
• Volume 49 , pages 30–50, ( 2021 )

Cite this article

You have full access to this open access article

• Ming-Hui Huang 1 &
• Roland T. Rust 2  

161k Accesses

332 Citations

20 Altmetric

Explore all metrics

The authors develop a three-stage framework for strategic marketing planning, incorporating multiple artificial intelligence (AI) benefits: mechanical AI for automating repetitive marketing functions and activities, thinking AI for processing data to arrive at decisions, and feeling AI for analyzing interactions and human emotions. This framework lays out the ways that AI can be used for marketing research, strategy (segmentation, targeting, and positioning, STP), and actions. At the marketing research stage, mechanical AI can be used for data collection, thinking AI for market analysis, and feeling AI for customer understanding. At the marketing strategy (STP) stage, mechanical AI can be used for segmentation (segment recognition), thinking AI for targeting (segment recommendation), and feeling AI for positioning (segment resonance). At the marketing action stage, mechanical AI can be used for standardization, thinking AI for personalization, and feeling AI for relationalization. We apply this framework to various areas of marketing, organized by marketing 4Ps/4Cs, to illustrate the strategic use of AI.

Similar content being viewed by others

Artificial Intelligence: The Revolution in Marketing

Artificial Intelligence in Marketing and Organizational Decision-Making: Some Challenges and Concerns

Artificial Intelligence Techniques in Precision Marketing: A Multi-criteria Analysis and Comparative Study

Avoid common mistakes on your manuscript.

Artificial intelligence (AI) in marketing is currently gaining importance, due to increasing computing power, lower computing costs, the availability of big data, and the advance of machine learning algorithms and models. We see wide applications of AI in various areas of marketing. For example, Amazon.com ’s Prime Air uses drones to automate shipping and delivery. Domino’s pizza is experimenting with autonomous cars and delivery robots to deliver pizza to the customer’s door. RedBalloon uses Albert’s AI marketing platform to discover and reach new customers. Macy’s On Call uses natural language processing to provide an in-store personal assistant to customers. Lexus uses IBM Watson to write its TV commercial scripts, “Driven by Intuition.” Affectiva, based on affective analytics, recognizes consumers’ emotions while they are watching commercials. Replika, a machine learning-based chatbot, provides emotional comfort to consumers by mimicking their styles of communication. It has even been asserted that AI will change the future of marketing substantially (Davenport et al. 2020 ; Rust 2020 ). However, academic marketing research to date provides insufficient guidance about how best to leverage the benefits of AI for marketing impact.

The academic literature on AI in marketing may be sorted into four main types. These are (1) technical AI algorithms for solving specific marketing problems (e.g., Chung et al. 2009 ; Chung et al. 2016 ; Dzyabura and Hauser 2011 , 2019 ), (2) customers’ psychological reactions to AI (e.g., Luo et al. 2019 ; Mende et al. 2019 ), (3) effects of AI on jobs and society (e.g., Autor and Dorn 2013 ; Frey and Osborne 2017 ; Huang and Rust 2018 ), and (4) managerial and strategic issues related to AI (e.g., Fountaine et al. 2019 ; Huang and Rust 2020 ).

The fourth literature stream, managerial issues related to AI, is currently dominated by consultants gravitating to the latest hot topic, and largely lacks a solid academic basis, albeit there are some recent studies trying to tackle strategic marketing issues. Examples include unstructured data for various areas of marketing (Balducci and Marinova 2018 ), analytics for consumer value in healthcare (Agarwal et al. 2020 ), machine learning prediction for mobile marketing personalization (Tong et al. 2020 ), in-store technology (e.g., robots, smart displays, or augmented reality) for convenience or social presence (Grewal et al. 2020 ), and AI for personalized customer engagement (Kumar et al. 2019 ).

To facilitate the strategic use of AI in marketing, we develop a three-stage framework, from marketing research, to marketing strategy (segmentation, targeting, and positioning, STP), to marketing actions (4Ps/4Cs), for strategic marketing planning incorporating AI. This strategic AI framework is based on a more nuanced perspective of the technical development of AI, existing studies on AI and marketing, and current and future AI applications. It can be used for strategic marketing planning, for organizing the existing AI marketing studies, and for identifying research gaps in AI marketing.

This paper contributes to the strategic application of AI in marketing by developing a framework that guides the strategic planning of AI in marketing in a systematic and actionable manner. It is achieved by bringing together diverse AI literatures on algorithms (e.g., Bauer and Jannach 2018 ; Davis and Marcus 2015 ), psychology (e.g., Lee et al. 2018 ; Leung et al. 2018 ), societal effects (e.g., Autor and Dorn 2013 ; Frey and Osborne 2017 ), and managerial implications (e.g., Huang et al. 2019 ) to explore what those literatures can tell us about managing AI in marketing. Marketing is an applied field, and using the more foundational literatures to inform marketing practice is an important role for marketing academia. This paper also contributes to strategic marketing research by providing a systematic and rigorous approach to identifying research gaps that bridge strategic AI marketing practice and research.

Conceptual foundation

The marketing research–strategy–action cycle.

We propose a three-stage strategic planning framework based on the marketing research–marketing strategy–marketing action cycle. Similar cycles have been proposed, such as Deming’s ( 1986 ) plan-do-check-act cycle, but that cycle omits the role of strategy. Our cycle views strategic planning as a circular process, starting from conducting marketing research to understand the market, the firm, the competitors, and the customers; to developing strategies for segmentation, targeting, and positioning; and to designing specific marketing actions to execute the strategy. This cycle does not stop at marketing actions. The execution of marketing actions will feed back to marketing research as market data, which constitutes a continuous cycle for marketing research–strategy–action, as illustrated in Fig.  1 .

AI and strategic marketing decisions

Multiple AI intelligences

Figure 1 illustrates that AI can play critical roles in all three strategic marketing stages. It shows that there are multiple AI intelligences that a marketer can leverage: mechanical, thinking, and feeling.

We conceptualize AI as the use of computational machinery to emulate capabilities inherent in humans, such as doing physical or mechanical tasks, thinking, and feeling; the multiple AI intelligence view considers that, rather than treating AI as a thinking machine, AI can be designed to have multiple intelligences, as humans have, for different tasks. Ordered by the difficulty with which AI can address them, there are mechanical, thinking, and feeling AI intelligences (Huang and Rust 2018 ; Huang et al. 2019 ).

Mechanical AI is designed for automating repetitive and routine tasks. For example, remote sensing, machine translation, classification algorithms, clustering algorithms, and dimensionality reduction are some current technologies that can be considered mechanical AI.

Thinking AI is designed for processing data to arrive at new conclusions or decisions. The data are typically unstructured. Thinking AI is good at recognizing patterns and regularities in data, for example, text mining, speech recognition, and facial recognition. Machine learning, neural networks, and deep learning (neural networks with additional layers) are some of the current methods by which thinking AI processes data. IBM Watson, expert systems, and recommender systems are some current applications for decision making.

Feeling AI is designed for two-way interactions involving humans, and/or for analyzing human feelings and emotions. Some current technologies include sentiment analysis, natural language processing (NLP), text-to-speech technology, recurrent neural networks (RNN), chatbots for mimicking human speech, embodied and embedded virtual agents for human interactions, and robots with customized hardware for sensing affective signals (McDuff and Czerwinski 2018 ).

It is important to note two qualifications of this multiple AI intelligences view. First, although we set up three intelligences, the assignment of some applications to a particular intelligence is mainly based on the purpose they are used for. Sometimes those applications may have some elements of more than one intelligence; thus, suggesting that the three intelligences are fuzzy sets (Varki et al. 2000 ). For example, facial recognition that is trying to identify someone is thinking AI (e.g., customs uses it to identify potential terrorists), while facial recognition that is trying to figure out somebody’s emotional state from their facial expression is feeling AI (e.g., advertisers use it to identify audience responses to ads). Second, we do not have true feeling AI yet Footnote 1 ; thus, the current practice is to use thinking AI to analyze emotional data (e.g., affective analytics) and two-way interactions (e.g., chatbots and social bots). Emotional data are distinct from cognitive data, in that they are contextual, individual-specific, and typically multimodal (speech, gestures, and language). Such data are about the individual in context, meaning that feeling AI needs to incorporate contextual and individual-specific data into modeling the emotional state of an individual.

Multiple benefits of AI

Each of the AI intelligences can deliver its unique benefit: mechanical AI is best for standardization, thinking AI is good for personalization, and feeling AI is ideal for relationalization (Huang and Rust 2020 ).

Mechanical AI provides standardization benefits due to its ability to be consistent. In marketing, various forms of mechanical AI have been used to provide a standardization benefit; for example, collaborative robots (cobots) help with packaging (Colgate et al. 1996 ), drones distribute physical goods, self-service robots deliver service, and service robots automate social presence in frontline (Mende et al. 2019 ; van Doorn et al. 2017 ). All these applications aim to generate standardized, consistent, and reliable outcomes.

Thinking AI provides personalization benefits, due to its ability to recognize patterns from data (e.g., text mining, speech recognition, facial recognition). Any marketing functions and activities that can benefit from personalized outcomes should consider thinking AI. The most common applications in marketing are various personalized recommendation systems (Chung et al. 2009 ; Chung et al. 2016 ), such as Netflix movie recommendations and Amazon cross-selling recommendations.

Feeling AI provides relationalization benefits (i.e., personalizes relationships), due to its capability to recognize and respond to emotions. Any marketing functions or activities that require interaction and communication, with the goal of relational benefits (e.g., when customer lifetime value is high) should consider feeling AI—one example being customer service. A broad range of marketing functions involves feelings, for example, customer satisfaction, customer complaints, customer moods, and emotions in advertising, etc., and can make use of feeling AI.

The strategic AI framework

We propose a three-stage strategic framework for using AI in marketing that leverages the three AI intelligences and their benefits, as shown in Fig. 1 . At the marketing research stage, AI is used for market intelligence, including mechanical AI for data collection, thinking AI for market analysis, and feeling AI for customer understanding.

At the marketing strategy stage, AI is used for the strategic decisions of segmentation, targeting, and positioning. Specifically, mechanical AI is ideal for discovering novel customer preference patterns in unstructured data, thinking AI is ideal for recommending the best segment(s) to target, and feeling AI is ideal for communicating with the targeted customers about the product. Footnote 2

At the marketing action stage, AI is used for the benefits of standardization, personalization, and relationalization, individually or synergistically. Marketers need to decide which AI intelligence(s) to use for which marketing actions. For example, payment and delivery are functions that can benefit from standardization by using mechanical AI, such as automatic payment and delivery tracking. Digital marketing can benefit from personalization by using thinking AI, such as various recommendation systems. Customer service and frontline customer interaction can benefit from relationalization by using feeling AI, such as social robots greeting customers and conversational AI providing customer service. The discussion of the strategic use of AI in marketing action is organized in terms of the marketing 4Ps/4Cs, to balance both the marketer and customer sides. Table 1 defines various strategic elements of this strategic AI framework, Table 2 illustrates marketing actions by using mechanical AI for standardization, thinking AI for personalization, and feeling AI for relationalization, with various examples and current and future scenarios, Footnote 3 and Table 3 summarizes the existing literature for using AI in each of the strategic elements. We discuss this three-stage framework in the following sections.

Marketing research

At this strategic stage, mechanical AI can be used for data collection, thinking AI for market analysis, and feeling AI for customer understanding.

Mechanical AI for data collection

Mechanical AI can automate data collection about the market, the environment, the firm, the competitors, and the customers. In the digitally connected world, market data can be easily tracked and monitored. Data sensing, tracking, and collection are routine, repetitive tasks that can be easily automated by mechanical AI.

Existing studies have shown various ways of using mechanical AI for data collection. For example, customer intelligence, including data about consumers, their activities, and their environments, can be collected if they use connected devices (Cooke and Zubcsek 2017 ), product usage and consumption experience can be visualized with Internet of Things (IoT) (Ng and Wakenshaw 2017 ), various advanced technologies and analytics can capture unstructured marketing activity data (Balducci and Marinova 2018 ), in-car sensors can track driving behavior for determining insurance premiums (Soleymanian et al. 2019 ), and retail technologies, such as heat maps, video surveillance, and beacons, can be used for profiling and recognizing retail shoppers (Kirkpatrick 2020 ). These studies show that, given the repetitive, routine, but high-volume nature of market data, mechanical AI can collect data efficiently at scale.

The data collection capability of mechanical AI is not limited to observable behavioral data; it can also be used to facilitate survey or experimental data collection to capture consumer psychographics, opinions, and attitudes. For example, human administration and supervision of ongoing surveys are no longer required and can be automated. SurveyMonkey and SurveyCake are two commercial survey platforms that automate survey design and data collection.

Thinking AI for market analysis

Thinking AI can be used to identify competitors in a well-defined market or outside options in a new market, and to derive insights for a product’s competitive advantages (i.e., the way the product can do better than competitors to meet customer demands). For example, supervised machine learning can be used for a mature market where the market structure is stable and known to marketers, whereas unsupervised machine learning can be used for a new market or spotting outside options where the market structure and trends are unstable and unknown to marketers.

In marketing practice, predictive analytics are commonly used to predict volatile market trends and customers’ heterogeneous preferences. For example, Gap, the fashion clothing brand, uses it to predict fast fashion trends to better meet customer needs, and Amazon uses it to predict a customer’s future orders (i.e., anticipatory fulfilling).

Existing studies have demonstrated various potential uses of thinking AI for market analysis. Automated text analysis can be applied for consumer research (Humphreys and Wang 2018 ), for marketing insights (i.e., prediction and understanding) (Berger et al. 2019 ), and for analyzing consumer consideration heuristics (Dzyabura and Hauser 2011 ). Machine learning algorithms and lexicon-based text classification can be used to analyze various social media datasets (Hartmann et al. 2019 ). Also, big data marketing analytics is now a mainstream approach for generating marketing insights (Berger et al. 2019 ; Chintagunta et al. 2016 ; Liu et al. 2016 ; Wedel and Kannan 2016 ).

Specific applications include mapping market structures for large retail assortments using a neural network language model, by analyzing the co-occurrences of products in shopping baskets (Gabel et al. 2019 ), detecting copycat mobile apps using a machine learning copycat-detection method (Wang et al. 2018 ), and aiding social media content engineering by employing natural language processing algorithms that discover the associations between social media marketing content and user engagement (Lee et al. 2018 ).

Feeling AI for customer understanding

Feeling AI can be used to understand existing and potential customer needs and wants, for example, who they are, what they want, and what their current solutions are. The major distinction between market analysis and customer understanding is that the latter often involves emotional data about customer sentiments, feelings, preferences, and attitudes. Thus, feeling AI can do a better job of understanding customers than mechanical AI and thinking AI, due to its capability of analyzing emotional data.

For existing customers, marketers can use feeling AI to answer the questions of whether they are happy with the product and why. Existing customers’ preferences are more stable, and a company has past and current transaction data for a deeper understanding. For example, Affectiva partnered with Ford to create AutoEmotive sentiment analysis, to try to figure out drivers’ emotional states.

For potential customers, marketers can use feeling AI to understand what they want and why they are happy with competitors or outside options. Potential customers’ needs and wants are more difficult to predict, and their emotional data are less available. In marketing practice, Albert AI, Adgorithm’s AI-driven marketing platform, has been used by RedBalloon to discover and reach new customers (Sutton 2018 ) and by Harley-Davidson to identify high potential customers based on the company’s customer relationship management database, personalizing the marketing campaign accordingly (Power 2017 ).

In academic research, existing studies have shown various approaches of using feeling AI to understand customers. For example, the sentiment expressed by consumers in social media (e.g., online reviews, tweets), including explicit and implicit language and discourse patterns, can be analyzed to understand consumer responses using their own language (Hewett et al. 2016 ; Humphreys and Wang 2018 ; Ordenes et al. 2017 ), the interaction between conversational AI and customers can be enhanced by applying analytical mapping to script appropriate response sequences that make customers feel that they have a “conversation” with AI (Avery and Steenburgh 2018 ), consumer consideration heuristics can be understood by machine learning (Dzyabura and Hauser 2011 ), and customer needs can be identified from user-generated content using convolutional neural network machine learning (Timoshenko and Hauser 2019 ).

Marketing strategy (STP)

At this strategic stage, marketers can use AI for the three key strategic decisions: segmentation, targeting, and positioning. However, before proceeding to specific STP decisions, marketers need to decide the overall strategic positioning to guide their STP decisions. Huang and Rust ( 2017 ) propose a technology-driven approach to positioning a firm’s strategy along the dimensions of standardization–personalization and transaction–relationship. A firm can pursue a commodity strategy that uses automated/robotic technology for efficiency, a relational strategy that cultivates the existing customers’ lifetime value, a static personalization strategy that uses cross-sectional big data analytics (e.g., like-minded customers) for personalization, or an adaptive personalization strategy that uses longitudinal customer data for dynamic personalization over time. This strategic positioning will guide firms’ STP decisions. For example, if a firm pursues the static personalization strategy, the firm may want to have a big, relevant, existing and potential customer database and let unsupervised machine learning explore patterns of preference or purchase behavior as the basis of targeting and positioning. If a firm pursues the adaptive personalization strategy, the firm may want to use supervised machine learning to continue analyzing existing customers’ satisfactions/dissatisfactions over time (which may not be big). However, if firms embrace a data-driven approach to STP, it can rely more heavily on AI to explore the STP possibilities.

In general, this stage of strategic decisions relies more on thinking AI, for its capability of processing data to arrive at new conclusions or decisions. However, it is worth reiterating that the assignment of some applications to a particular intelligence is mainly based on the purpose an application is used for. For example, when thinking AI becomes completely routinized, as is often the case in segmentation applications, it shares many of the characteristics of mechanical AI, because it simply identifies patterns from data routinely and repetitively, without involving much about the purpose of making new decisions (e.g., segmentation but not retargeting).

• Segmentation

Segmentation is to slice a market into pieces, with customers in each piece having unique needs and wants, for example, using gender to slice the shoe market into male and female shoes segments; and using price and quality to slice the air travel market into budget and premium airlines segments. Mechanical AI, especially the various mining and grouping techniques, has the strength of identifying novel patterns from data.

AI segmentation is flexible, in that it can disaggregate the market into segments of one (i.e., each individual customer is a segment) and can aggregate scattered long tails into one segment. Wang et al. ( 2017 ) demonstrate that transfer learning can be used to model the tail of the distribution, by learning from the head of the distribution and transferring the learning to the data-poor tail. This flexibility in aggregation and disaggregation allows marketers to find the right size of segment.

Existing studies have shown how data mining can be used to uncover patterns that are difficult for human marketers to see. For example, text-mining and machine learning can be used to automatically process and analyze loan requests to slice borrowers into good customers (those will pay back the loan) and bad customers (those will not) (Netzer et al. 2019 ), automated text analysis and correspondence analysis can be used for psychographic consumer segmentation in the art market (Pitt et al. 2020 ), data mining can be used to obtain tourist segments based on the meaning of destinations to consumers, that is better than the classic clustering methods (Valls et al. 2018 ), and retail customers can be micro-segmented based on their preferences for personalized recommendation (Dekimpe 2020 ).

Targeting is choosing the right segment(s) on which to focus the firm’s marketing actions. Slicing the market is more mechanical and can be done automatically by mechanical AI, given the relevant data. However, choosing the right segment requires domain knowledge, judgement, and intuition. Various technologies and analytics have been used for targeting, such as search engines using keywords searched and browsing history to target search consumers, and social media platforms using interests, content, and connections to target social media consumers (Liu 2020 ). The representative AI for this decision is recommendation engines that can recommend various potential targets for marketing managers’ final verdict, and predictive modeling that can be used to choose which segment to target.

Existing studies show that various thinking AI can be used for this purpose. Examples include targeting customers using a combination of statistical and data-mining techniques (Drew et al. 2001 ), screening and targeting cancer outreach marketing using machine learning and causal forests (Chen et al. 2020 ), optimizing promotion targeting for new customers using various machine learning methods (Simester et al. 2020 ), identifying the best targets for proactive churn programs from field experimental data using machine learning techniques (Ascarza 2018 ), and profiling digital consumers for targeting using online browsing data (Neumann et al. 2019 ).

• Positioning

Positioning bridges product attributes and customer benefits by finding a competitively advantageous position for the product in customers’ minds. This term is often associated with brand positioning or advertising positioning for its association with customer perceptions and communications to maintain a desirable perception. Daabes and Kharbat ( 2017 ) demonstrate how data mining techniques can be used to distill a customer-based perceptual map, as an alternative to marketer knowledge, from mining customers’ perceptions.

Compared with the mechanical-based segmentation and the thinking-based targeting, positioning is more about speaking to customers’ hearts, typically as a positioning statement or slogan in promotional communication. Gali et al. ( 2017 ) find that tourism positioning slogans in top destinations tend to emphasize the affective component.

Some successful positioning statements help brands to occupy a unique position in customers’ minds and thus succeed in the market for a long time. For example, Nike’s “Just do it,” Apple computer’s “Be different,” and McDonalds’ “I’m loving it” all communicate with customers by speaking to their hearts. Feeling AI, such as feeling analytics, is ideal for this strategic decision to help develop compelling slogans by understanding what resonates with target customers.

Academic research on this decision is sparse, indicating a research gap for using feeling AI to create compelling positioning.

Marketing action

At this strategic stage, marketers can use mechanical AI for standardization, thinking AI for personalization, and feeling AI for relationalization (Huang and Rust 2020 ). Depending on which benefit is desirable, a marketer can use multiple AI intelligences individually or collectively. We illustrate the use of AI intelligences in various areas of marketing with examples and current and future scenarios, and support the illustration using the existing literature. The discussion is organized by marketing 4Ps (product, price, place, and promotion) (Kotler and Keller 2006 ) and the corresponding 4Cs (consumer, cost, convenience, and communication) (Lauterborn 1990 ) to emphasize that the 4P actions need to be able to deliver consumer benefits.

Product (consumer)

Product (consmer) actions include goods and services as offerings to meet the consumer’s needs and wants. Such actions typically include product design, packaging, branding, and returns, and the associated customer services in these activities. We illustrate this decision using product/branding and customer service, with product/branding representing the “product” side and customer service representing the “consumer” side. Product and branding are related in that branding is the identity (e.g., name, symbol, logo) of a product. We put them together to maintain the simplicity of the table.

Product/branding focuses on product creation (including tangible goods R&D and production, and service innovation and process) and branding (i.e., the identity of a product).

Mechanical AI can be used for product/branding actions that can benefit from standardization. For example, brand logo design can be automated by a decision-tree like machine learning using multiple-choice questions, allowing small budget marketers to have AI-assisted branding (Avery 2018 ). Product adoption and acceptance can be tracked and monitored automatically. While enjoying the standardization benefit of mechanical AI, one existing study is cautious about automating product decisions, when those products are related to consumers’ identity (Leung et al. 2018 ).

Thinking AI can be used for product/branding actions that can benefit from personalization. For example, marketing analytics can predict market trends for product design that cater more precisely to target customers’ preferences, big data analytics can be used to inform product development to quickly adapt to consumer trends and changing preferences (Dekimpe 2020 ), topic modeling can advance service innovation and design (Antons and Breidbach 2018), adaptive systems can be used to personalize service to each consumer’s preference (Chung et al. 2009 ; Chung et al. 2016 ; Dzyabura and Hauser 2019 ; Liebman et al. 2019 ), and deep learning can be used to personalize point-of-interest recommendations (Guo et al. 2018 ).

Feeling AI can be used for product/branding actions that can benefit from relationalization. For example, conversational AI can be trained to have brand personality to interact with customers (Wilson and Daugherty 2018 ), machine learning can recommend TV programs based on the viewer’s mood, brands can track their reputation through text and sentiment analyses tweets, reviews, and social media posts (Rust et al. 2020 ), and chatbots can mimic customers’ communication style to provide emotional support. Kumar et al. ( 2019 ) provide a systematic exploration about the role of AI in personalized engagement marketing, an approach to create, communicate, and deliver personalized offering to customers. Huang and Rust ( 2020 ) show that feeling AI can be used to engage customers in service interaction.

From the consumer side, existing studies further show that customers have varying responses and attitudes toward using AI products. For example, consumers may be resistant to personal medical AI (Longoni et al. 2019 ), identity-based consumption automation (Leung et al. 2018 ), and anthropomorphized consumer robots (Kim et al. 2019 ; Mende et al. 2019 ). These studies put a boundary condition for marketers when using AI in the product/branding actions to generate positive customer responses.

Customer service is emotionally charged, yet is also costly. A marketer can handle customer service using the three AI intelligences to balance the cost/satisfaction tradeoff in serving customers.

Mechanical AI, such as text-based chatbots, is widely used online to handle a massive amount of routine customer service. Most customer questions can be answered by such bots. As long as such automation is not related to customers’ identity (Leung et al. 2018 ), it is easy to implement, cost-efficient, and scales up easily.

Thinking AI, such as natural language processing chatbots, can handle more diversified customers and idiosyncratic issues (e.g., multicultural customers with accents and contextual-dependent complaints). This is an AI version of the old telephone menu, except that customers are talking to chatbots, rather than human customer service agents. Although a recent study shows that customers may not feel comfortable yet about talking to chatbots (Luo et al. 2019 ), with the wider acceptance of AI and the further advance of chatbots, we can expect the acceptance to increase over time.

Feeling AI, such as Cogito’s emotional AI systems, can analyze the pace of speaking, energy and empathy, and common errors of conversations, and gives in-call guidance to customer service agents in call centers that make the conversations more natural and engaging.

Price (cost)

Price (cost) action includes the tasks of payment, price setting, and price negotiation, which is the cost that the consumer pays for the product.

The payment task is routine and can be handled best by mechanical AI. Apple Pay, Google Pay, PayPal, Amazon Payments, and Square are some popular automatic payment methods for online marketers.

The price setting task is calculation-intensive and analytical, which is the strength of thinking AI. Misra et al. ( 2019 ) demonstrate that multiarmed bandit algorithms from statistical machine learning can dynamically adjust online prices in real-time, even if price information is incomplete. Bauer and Jannach ( 2018 ) show that a machine-learning framework based on Bayesian inference can optimize online pricing even when data are updated frequently, and are sparse and noisy. Prices can also be personalized by incorporating consumer online WOM (Feng et al. 2019 ) and consumers’ private personal information (Montes et al. 2019 ). Dekimpe ( 2020 ) proposes that retailers can use big data to optimize dynamic best-response pricing algorithms that consider consumer choices, competitor actions, and supply parameters.

The price negotiation task is interactive; thus, feeling AI is in a better position to undertake this task. Pulles and Hartman ( 2017 ) hypothesize that interpersonal likeability would impact the price negotiation outcome in a B2B relationship, suggesting that interaction, communication, and sentiment may be critical for price negotiation.

Place (convenience)

Place (convenience) is the way that the consumer can access the product. We discuss two broad categories of place action: Retailing and frontline, virtually or physically, in which interactions play a key role; and distribution, logistics, and delivery, in which convenience is the key.

Retailing/frontline is the area of marketing that most employs embodied AI (i.e., robots) to facilitate frontline interactions.

Mechanical AI can be used to automate backend marketing processes and frontend interactions. In the backend, service processes can be automated (Huang and Rust 2018 ) and retail processes can be optimized using IoT (Grewal et al. 2018 ). In the frontend, service robots can interact with scale and consistency (Wirtz et al. 2018 ), and can automate social presence in the frontline (Mende et al. 2019 ; van Doorn et al. 2017 ). Frontline service robots are common; for example, Giant grocery uses the robot Marty to identify hazards in store (e.g., detecting milk spilled on the floor) and HaiDiLao hotpot uses robots to deliver soup base from kitchen to table side. Grocery shopping is typically repeat purchase, which does not involve too much interaction, communication, and emotion, and thus using mechanical AI to automate the marketing function is desirable.

At the thinking level, due to the direct customer contact nature of retailing, AI is used to facilitate in-store shopping for individual customers. Amazon Go, an experimental grocery store, uses facial recognition technology to identify and remember each customer, Macy’s On Call, a mobile shopping personal aid, provides in-store information to help customers locate items they are looking for, and Alibaba’s FashionAI system uses smart mirrors in sales floor and changing rooms to display items that each customer selects and suggests complementary items.

Feeling AI can be used to enhance interaction and engagement. For example, service robots can easily do surface acting (Wirtz et al. 2018 ), and “one-voice” AI can enhance customer engagement by integrating various interfaces involved in a customer’s journey (Singh et al. 2020 ). At the feeling level, various embodied robots are used to engage customers and optimize their experience. For example, Pepper robots are used by Marriott to greet and interact with customers. Hotels and travel typically involve more interactions and more emotions, and thus feeling AI naturally suits. Nevertheless, marketers need to be cautious, in that anthropomorphized robots are found to increase perceived warmth but decrease liking (Kim et al. 2019 ); thus, in the case of embodied frontline robots, marketers need to take the appearance of robots into consideration.

Distribution/logistics/delivery is an area of marketing in which many functions and processes can be highly automated; including packaging, inventory, warehousing, supply chain, logistics, and delivery, to provide convenience benefits to customers.

Tasks in distribution are mostly mechanical, routine, and repetitive; thus, the standardization benefit of mechanical AI fits well. We have seen cobots for packaging, drones for delivery, IoT for consumption tracking and order refilling, and self-service technologies for delivering service to customers directly. Such standardization provides a convenience benefit to customers.

Moving up to the thinking AI level, we have seen that a customer’s future orders and refills can be anticipated by predictive analytics, and ordered products can be delivered to customers using autonomous cars equipped with facial recognition technology (e.g., JD.com and Domino’s use self-driving cars for delivery).

So far, feeling AI is not as widely used as the other two AI intelligences for this marketing action, due to distribution’s mechanical and thinking nature.

Promotion (communication)

Promotion (communication) is the marketing communications between the consumer and the marketer. It can include personal selling, traditional mass media advertising, and more commonly nowadays direct marketing, database marketing, and digital marketing (social media marketing, mobile marketing, search engine optimization, etc.). All these can benefit from AI intelligences.

Mechanical AI is ideal for automating various repetitive, routine, and data intensive functions of promotion (Huang and Rust 2018 ). Most of these are about promotional media planning and executions. Examples include automating advertising media planning, scheduling, and buying; automating search campaigns execution, keywords researching, and bidding; automating social media targeting, retargeting, and posting. Especially considering the real-time nature of digital marketing, such automation greatly aids marketers’ efforts in the labor-intensive, high-time-pressure process.

Thinking AI has great potential for promotion content creation and personalization. For example, AI content writers can facilitate the generation of ad or post content. A recent example was a Lexus car commercial that used IBM Watson to create the “Driven by Intuition” commercial script. Content can be personalized and optimized to different customer profiles at different locations and different times. Kantar Analytics uses content analytics to help advertisers create content that shortens the idea-to-value time and maximizes content effectiveness (Gopinath 2019 ).

Feeling AI can be used to track real-time customer response to promotional messages (like, dislike, disgusted, funny, etc.) and then adjust what to deliver and what to emphasize in both media and content. At the feeling level, more real-time and accurate emotion sensing from posted messages can better engage customers and provide a better interaction experience (Hartmann et al. 2019 ; Lee et al. 2018 ).

Managerial implications

Our framework provides a roadmap for marketers to implement various AI intelligences in marketing, strategically and systematically. Table 4 summarizes the managerial implications of the framework by contrasting current marketing practice with the emerging AI-enabled marketing practice. The two practices should be viewed as two ends of a continuum, with more AI intelligences expected to be used for more strategic elements at the three stages over time. We discuss these implications below.

Implications for marketing research

At this stage, marketers need to decide (1) how to use AI to identify competitors (including competitors in the same industry and outside options) and to understand competitive advantages (i.e., the way the product can do better to meet customers’ needs), and (2) how to use AI to discover and understand existing and potential customers (i.e., who they are, what they want, and what their current solutions are) and to understand their preferences and feelings. For firms that embrace a theory-driven approach to marketing strategies, data and intelligences resulting from this stage play a critical role.

Data collection

Currently, surveys, experiments, interviews, panels, and sales data are still the major approaches for marketers to obtain data, even though the administration of these methods can be partially automated or facilitated by technology. Surveys and experimental methods tend to be more theory driven, while other methods tend to be more data driven. Marketers often also rely on thirty-party syndicated data (e.g., YouGov), especially for external data that are difficult for the firm to collect. These data are typically delayed, out of context, and ad hoc, meaning that they are collected periodically, after the fact (after consumption has occurred), and not during data generation.

By contrast, emerging practices automate most of the data collection by connecting technology (e.g., IoT, social networking sites, mobile apps), sensor technology (e.g., remote sensing, detection), and wearable technology (e.g., smart watch, Fitbit). These mechanical AI approaches track and capture real-time data when they are occurring. Thus, the data are in context, about the customer, and during the consumption experience. Such spontaneous data collection tends to be more data driven, but if theories can be developed priori to guide and update the continuous data collection, they can be theory driven as well.

Market analysis

Current market analysis, although marching toward machine learning-based analysis rapidly, still relies heavily on statistical analysis to analyze structured data for marketing insights. It is also common for firms to purchase third-party data and analysis, especially for external market and competitor analysis. Such analysis tends to be standardized across firms (with limited degree of customization), and thus the insights derived from it are less useful for deriving a unique value proposition. Firms also monitor and analyze first-party data, especially for firm marketing resource analysis and for existing customer analysis for which internal data are available.

By contrast, big data and machine learning-based analytics are the emerging approach for marketing insights. Online reviews, opinions, and behaviors all can be mined, and data can be in text, image, audio, or video. When the question at hand is clear (e.g., a mature brand), supervised machine learning can be used to conduct theory-driven analysis, whereas when the question at hand is unclear (e.g., a new brand), unsupervised machine learning can be applied to obtain data-driven insights. Balducci and Marinova ( 2018 ) offer a detailed description of various methods of analyzing unstructured data in marketing. More advanced approaches to marketing analysis include using deep learning methods such as predictive analytics, computational creativity, personalization algorithms, and natural language processing systems, to come up with intuitive suggestions for marketing strategies.

Customer understanding

Current practice relies heavily on focus groups to gain qualitative insights about customers. Focus groups are time consuming and labor intensive, not to mention not representative. Marketers also observe customers’ behaviors and choices, and their reactions to promotions to understand their preferences and the underlying reasons.

By contrast, data about customers’ feelings, moods, and emotions can be obtained directly from customers’ interaction with AI (e.g., conversational bots), rather than inferred from psychometrics, using conversational bots and analyzed using feeling analytics (e.g., posts on social media, voice recordings of customer interactions, and chat transcripts). Feeling analytics can identify customer insights with scale and cost-efficiently. Given that emotional data are personal and in context, understanding customers in context provides richer insights about who they are and what they like.

Implications for marketing strategy (STP)

At this stage, marketers can leverage the three AI intelligences for segmentation, targeting, and positioning, respectively. For firms that embrace a data-driven approach to marketing strategy, this stage may play a bigger role than conclusions derived from marketing research.

The current approach relies on the marketer’s intuition and domain knowledge to choose a limited number of segmentation variables with which to slice the market, such as demographics, psychographics, and behavioral variables. Such an approach sees customers as aggregate, not individual. For example, some customer equity models focus on segmenting customers based on their acquisition and churn rates, and do not see them as individually unique (e.g., Blattberg and Deighton 1996 ; Gupta et al. 2004 ). Artificial personas thus are often applied to these segments to help marketers make the aggregate segments more personal and relatable.

By contrast, when data mining is used to segment the market, it no longer requires marketers to decide segmentation variables on an a priori basis, because unsupervised machine learning can discover the patterns itself. A virtually unlimited number of variables can be used to slice the market in a novel way that often goes beyond any pattern that human marketers can see. It is like the customer lifetime value model, in which each customer is valuable in some way. The Target store knowing a daughter is pregnant before her father knows, by mining the daughter’s purchase patterns, is a classic example.

Currently targeting mostly uses marketers’ subjective judgment, based on the resources, the competitive advantage of the firm, and the value of the segment to the firm. It is typically at the segment level (not individual level), and often trades off segment size for effectiveness.

By contrast, after very refined segmentation, it is thinking AI’s turn to recommend the best segment(s) to target. It is very well likely to be a segment of one, since personalization is the strength of thinking AI. With the capability to slice the market in unlimited ways and at the individual customer level, targeting in emerging practice is more commonly at the individual customer level. For example, online ads use cookies to target individual customers by following them around, wherever they go on the Internet. The new targeting also is flexible, because it can aggregate individual customers into a segment, if they have similar preferences (e.g., like-minded customers recommendation, aggregating long-tail customers even when each individual customer may not be valuable), or it can disaggregate a segment, if heterogeneity within the segment becomes manifest. Targeting involves not just identifying segments but also determining whether they should be pursued. Whether they should be pursued or not is a matter of predicting the outcome if they pursue, and prediction at the individual level is only scalable with the help of AI.

Positioning is currently a human task, for it involves judgment, intuition, and creativity that machines are not particularly good at it yet (Davis and Marcus 2015 ; Schoenick et al. 2017 ). Kelly ( 2019 ) argues that creativity is not just about novelty but is also about social acceptability. A novel idea has to be accepted by community to be deemed as creative. Because creativity is socially embedded, a good positioning is in the eye of the targeted customers. Although we have seen an increasing number of examples of AI participating in the creative process, for example using AI to compose its own music and to write short stories, there is still a long way to go for AI to be as creative as humans while still maintaining strategic relevance. For example, the script of the 2018 Lexus car “Driven by Intuition” TV commercial was created by AI by applying the machine learning approach. Lexus fed machines with award-winning luxury ads, Lexus brand data, and emotion data, that were shown to connect with viewers, to tell the story about how Lexus generated the new ES executive saloon car. The commercial appears to have face validity as a luxury car commercial. However, this commercial may not be very strategic because the ad has an unclear customer segmentation and ambiguous positioning (Rust and Huang 2020 ). This real-world example illustrates that positioning can be expected to be a human-AI collaboration for the immediate future.

Implications for marketing action

At this stage, marketers can leverage the three AI intelligences for the 4Ps to serve the 4Cs. The key questions to answers are which AI to use and how to use it for marketing actions.

The current practice for product decisions is to use conjoint analysis to decide what levels of product attributes to include in product development, use test markets to decide whether and to what degree the product will be accepted, and use aggregate sales results after the launch of the product as a proxy for customer feedback.

The emerging practice is to use mechanical AI to automate production and service process (e.g., Huang and Rust 2020 ), use thinking AI, such as cognitive technology, to facilitate product research and development (which is currently more common in drug development), and use feeling AI, such social robots and conversational bots, to interact with customers, from which they obtain real-time, first-hand customer feedback about the product. Such a process can become an adaptive loop that improves the product continuously based on customer feedback.

The current practice is to list prices on retail stores, websites, or mobile apps, set the prices discriminately, based on segments, and have salespeople handle price negotiations. Price menus offline are difficult and labor intensive to change, and price menus online, though easier to update, are also easier to be compared. Price setting typically requires careful and extensive calculation, taking various factors into consideration. Price negotiation is more of an art than a science, especially for big ticket items.

The emerging practice is to use mechanical AI to automate price setting and changes, thinking AI for price personalization, and feeling AI for price negotiation. Price updating is a simple routine task, price setting can be achieved by the powerful calculating machine, thinking AI, and can be personalized taking individual customers’ preferences and sensitivity into consideration. Price negotiation can be done when feeling AI detects customer reactions to the offered price in real-time.

The current practice relies on self-service to automate routine delivery and labor-intensive physical distribution for ordering processing, materials handling, and delivery; unskilled labor at the frontline to offer homogeneous assistance, and frontline employees for emotional labor.

Distribution, logistics, and delivery can mostly be automated with mechanical AI, and is a fast-growing emerging practice, for example, product tracking systems for firms to track where the product is in the supply chain, and for customers to track when they can expect to receive the product. Thinking AI, such as personal shopping assistants, has been used to assist customers about where to find the product. Feeling AI, such as conversational bots, can be used to display emotions in service interaction without the need to actually experience emotions (Wirtz et al. 2018 ).

The media planning part of promotion has a higher degree of automation by mechanical AI, due to the repetitive nature of the task. The content creation part of promotion, though having a lower degree of automation, is increasingly handled by thinking AI, such as AI writers, to generate content on its own, or to stimulate human creativity. Customers’ reactions to promotion are still mostly measured using traditional marketing research methods. The emerging practice is to use feeling AI, such as feeling analytics, to sense, react, and adjust promotions in real-time based on customers’ emotional reactions.

Directions for future research

The framework lays out a stage-by-stage circular process for using different AI intelligences in marketing. It provides rich implications for future research based on the use of AI intelligences. We discuss these implications, organized by the three AI intelligences. The last column of Table 4 illustrates one example research question for each element.

When mechanical AI is used for data collection, it makes both the competition and the customer more transparent, making the governance of privacy issues more central to marketers. When thinking AI is used for market analysis, it turns theory-driven marketing research into data-driven, resulting in a debate about whether a data or theory approach to marketing research should be embraced. When feeling AI is used for customer understanding, it is as if AI can really understand emotions, when we don’t have true emotional machines yet. These issues give rise to various future research topics.

How would privacy and data security affect marketing’s use of AI for data collection? When mechanical AI is used to collect and integrate multiple sources of real-time, in-context consumer data, the risk of privacy infringement and the consequences of a data breach are much higher. It is more difficult to keep such “all-in-one” data anonymous, and consumers are more sensitive to data sharing and data breaches. We have seen the public outcry when Facebook permitted the unauthorized licensing of 30 million user accounts to Cambridge Analytica, with its brand reputation plummeting dramatically in two weeks (Rust et al. 2020 ). In the AI age, both competition and customers become more transparent, due to AI’s data collection capability. How should privacy and data security be handled? The existing studies suggest that both data and consumer characteristics need to be considered. It has been shown that some types of data are more sensitive, such as personal data, financial data, health data, or medical records. Agarwal et al. ( 2020 ) urge healthcare organizations to find ways of protecting patient privacy as a competitive advantage, because healthcare consumers are especially weak in protecting their own data. It has also been shown that some consumers are more sensitive to data sharing. Thomaz et al. ( 2020 ) identify two types of consumers who vary in willingness to share information. They advise that marketers need to understand which type of consumers they are dealing with when providing personalization. Some methods of protecting marketing data have been proposed (Schneider et al. 2017 , 2018 ). Future research will need to explore more delicate approaches to handle the ubiquitous data collected by AI.

How to balance data- and theory-driven market analysis? Traditional market research has the strength of hypothesis testing, whereas the emerging machine learning approach to analyzing unstructured big data has the merit of data exploration. How to balance the two approaches of inquiry and benefit from them? The Silicon Valley viewpoint is that prediction, driven by data, is everything. Balducci and Marinova ( 2018 ) show a data-driven approach to marketing management and research and Ma and Sun ( 2020 ) argue for machine learning methods for marketing research, while Lehmann ( 2020 ) proposes a method to blend theory and data in the evolving world of marketing research. These practices and studies suggest that AI will expand the role of exploration in the development of theory; thus, investigating how the two approaches can be balanced, integrated, or blended would allow marketing to leverage the benefit of data-driven market analysis.

What algorithms and models are needed for feeling AI? Much of the progress in technical algorithms has been made in the area of deep learning neural networks, applied to thinking AI problems, such as personalized advertising. The current neural network-based machine learning is mainly for prediction, rather than for understanding. We need algorithms to understand emotions and react to emotions appropriately (Rust and Huang 2020 ). There are different ways to understand emotions, such as understanding emotions in written language (text-based sentiment analysis), in oral conversation (natural language processing), or in facial expression (facial recognition). Although there is some work already in using thinking AI to analyze emotional data (e.g., affective analytics), with respect to facial recognition and natural language processing, models in this area are still fairly rudimentary.

• Marketing strategy

For the three strategic elements of segmentation, targeting, and positioning, mechanical AI can slice the market based on virtually an unlimited number of variables, thinking AI can recommend the target, even if it is a segment of AI customers, or a combination of AI and human customers, and feeling AI can team up with marketers to come up with creative positioning that touches consumers’ hearts. We discuss these implications.

How best to visualize segmentation? The emerging AI approach to segmentation can be based on an unlimited number of variables. This creates a challenge as to how best to visualize for marketers to make sense of these multi-dimensional segmentation results for targeting. Thinking AI can easily decipher such complexity, but human marketers may not. Earlier studies have attempted to visually display three-dimensional segmentation and positioning from a larger set of variables (e.g., Donthu and Rust 1989 ; Rust et al. 1992 ). There are some existing tools, but they are not designed for STP decisions. We need better and friendlier human-computer interfaces to summarize and display those extremely high dimensional results to inform human marketers.

What happens when the customer is AI? Alexa and Siri are not the only AI that can potentially buy things. Increasingly, customers use AI as their agents for information collection, price negotiation, or purchase. When it becomes more real that marketers will be catering to non-human customers (Rust 1997 , 2020 ) or a combination of humans and AI as customers, how AI customers would behave, and how marketers should serve them are both pertinent issues. Are customer emotions still pertinent? Will rational decision making become the only route to decisions? How will AI customers change the scope of consumer research?

How should marketers and AI collaborate, to resonate with customers? Positioning requires creativity and empathy about the preferred way that a customer would like to see a product. AI can optimize a positioning recommendation, but may not be mature enough to resonate with customers, given that we do not have true feeling AI yet. Thus, for the time being, it is important to explore the best approaches to marketer-AI collaboration for a positioning that resonates with customers. This is a question that involves exploring the role of AI intelligences in creativity; for example, to what extent should marketers let machines be creative on their own (e.g., the Lexus commercial example), or use them as creative support? What will be consumers’ attitudes toward machine-generated creativity? What will decide the acceptance or rejection of machine creativity?

For the four strategic elements of 4Ps/4Cs, as illustrated in Table 4 , all three AI intelligences can be applied to each element. We illustrate one research question for each of the elements, highlighting the level of AI intelligence that is likely to help the most.

How is AI best used in developing new products to meet customer needs and wants? Thinking AI can be innovative and creative, but it may not be straightforward as to whether the new products will meet customer needs and wants, because many of those needs and wants are implicit, rather than explicit. This is the challenge of converting product attributes into consumer benefits. Rogers’ ( 1962 ) diffusion of innovation theory suggests that both consumer and product innovativeness matter for new product acceptance. AI, with its capability to find patterns and regularities in data that otherwise would be hard to see, can be used to uncover consumers’ implicit needs and wants, and match those needs and wants to products at different stages of lifecycle. For example, for the earlier stages of product lifecycle, AI may play a bigger role, because consumers’ needs and wants tend to be implicit, whereas for the latter stages, human marketers may play a bigger role, because those needs and wants become more explicit.

How to manage AI-based price negotiation? Negotiating price to achieve price personalization can be expected to be more common and increasingly at the individual level (e.g., price bidding). The price discrimination literature (e.g., Montes et al. 2019 ) tends to view pricing as a process of marketers setting the price and consumers reacting to the price. In AI-based price negotiation, the process is dynamic and real-time. We have seen the popularity for technology giants to use AI for price negotiation, such as Google’s search keywords bidding and Facebook’s ad bidding for reaching the desirable customers. What will be the new mechanisms and methods for pricing when AI is used to negotiate prices more widely?

How to manage customer disengagement due to place automation? The goal of the place decision is to provide the customer a convenience benefit. When the entire place process is automated, which is increasingly common these days, there is no human-to-human contact, and thus it is likely for customers to become less engaged with the brand. How to avoid customer disengagement while striving to provide convenience? Are there ways to mitigate the part that is lost?

How to use AI to build strong relational bonds? When two-way communications and interactions are made easier and richer by AI, how are trust and commitment affected by interacting with AI? What will be gained, and what will be lost? Customers may have changed by interacting with AI. For example, when AI does more thinking, will customers think less when they may feel more ? To what extent does the advance of AI, from mechanical, to thinking, to feeling, lead to customers feeling more, at the expense of thinking, such as accepting fake news on social media at face value?

Current limitations of AI

Our strategic framework illustrates applications of multiple AI intelligences to various areas of marketing at different strategic decision stages. These applications are not without limitations. We discuss the major limitations of applying the three AI intelligences to marketing for marketers to use AI more wisely.

Limitations of mechanical AI

Although current mechanical AI has the strong capability of collecting and integrating multiple sources of data autonomously, very often contexts of the data are lost, creating problems in modeling, especially for emotional data. The automated process of data collection also makes customer intimacy less achievable because it is machines talking to machines.

Non-contextual data

Many data collected by mechanical AI are non-contextual. This is especially the case for emotional data because such data are about the individual in context, meaning modeling the emotional state of a consumer requires contextual and individual-specific data. Contextual data are often lost during interaction. One Dell AI expert said at a frontline service conference in 2019 that it is not that difficult to model emotions (meaning using the existing machine learning approach), but the difficulty lies in that emotional data are difficult to capture, and thus are not analyzed. For example, in a customer service interaction, the content and sentiment of the conversations are recorded, but not the context of the conversations. When an angry and frustrated customer calls, his way of talking may be different, depending on whether he is alone or with a group of friends, whether the weather is gloomy or sunny, or whether the traffic is jammed or smooth. Even if voice analytics can detect the sentiment of his voice, it cannot provide guidance to the customer agent as to why the customer is angry, and what the best way to respond is (Rust and Huang 2020 ).

Machine to machine interactions

Or communications are predicted to be the key emerging technology-enabled interactions in digital environments (Yadav and Pavlou 2020 ). Examples include an ATM getting authorization from the bank for a cash withdrawal, and a refrigerator sensor sending inventory information to a vendor via IoT. Mechanical AI plays an important role in those routine interactions; however, it may come at the cost of customer intimacy (e.g., Treacy and Wiersma 1997 ). To be able to figure out the customer side of strategy requires customer data, collected and compiled by mechanical AI, which is the only methodology that is powerful enough and scalable enough to capture individual-level data. When customers are removed from the interactions, it is more challenging for marketers to remain intimate with customers. Machine-to-machine interactions and communications will be a new form of customer interactions that calls for more future studies to tell marketers how to approach them, and how to keep a balance between operational efficiency and customer intimacy.

Limitations of thinking AI

Current thinking AI, though powerful, may not be neutral and transparent, which can result in biased recommendations or entail unintended consequences.

How thinking AI comes up with a certain recommendation often is not transparent to human marketers. The current dominant machine learning approach to AI designs machines to learn via a mapping mechanism (i.e., map input pairs (X,Y) to output Y=F(X)), not via cognitive reasoning (Lewis and Denning 2018 ). This results in the output being unexplainable because it does not answer the “why” question. Thus, there is a need for scholars to develop explainable AI so that thinking AI can be used for trustworthy and fair marketing exchanges (Rai 2020 ). Opaque AI also results in liability issues. If AI output is not transparent, when AI goes wrong, marketers who use the AI are likely to be held accountable. The accountability issue has emerged since the first fatal accident of an autonomous car. Thus, marketers, as AI users, when using thinking AI for strategic decisions, need to strive to use the most explainable AI, rather than just the most powerful AI, to keep the exchanges transparent to both the marketer and the customer.

Thinking AI is not neutral. If data input is erroneous or biased, output is likely to be biased too. However, biased input is not the only way AI bias can occur. For example, it has been shown that for loan decisions, discriminatory results can occur even if there is no bigotry programmed into the system, and the system only seeks to maximize profit (Ukanwa and Rust 2020 ). Researchers have also shown that gender bias can occur without any conscious (or unconscious) attempt to produce a biased outcome—using only an unbiased algorithm (Lambrecht and Tucker 2019 ).

Marketing researchers have addressed some issues, such as managers and management education underpreparing the next generation for feeling and emotional intelligence (Huang and Rust 2018 ; Huang et al. 2019 ), and IoT may impose constraints and restrictions on consumer journeys (Hoffman and Novak 2018 ; Novak and Hoffman 2019 ). Thus, in using thinking AI for market analysis, for targeting, and for personalized marketing actions, marketers need to be aware of the potential AI biases and have better knowledge about how AI learns to avoid AI biases.

Limitations of feeling AI

Although using feeling AI for two-way interactions involving humans and for analyzing human feelings and emotions is common in marketing due to the high-touch nature of many marketing functions (e.g., frontline interactions, customer service, and emotional ad appeals), we don’t yet have true emotional machines that can recognize, act, and react to human emotions appropriately. The substitutive use of mechanical and thinking AI for feeling AI may generate some unintended consequences.

Technology unreadiness

The fact that marketers are using “lower” intelligence AI for feeling functions (i.e., using mechanical AI to capture emotional data and using thinking AI to analyze emotional data) may inflate the perceived capability of AI to assist marketers in understanding customer emotions. For example, marketers may overly rely on such feeing AI to interact with customers, resulting in customer disengagement. Srinivasan et al. ( 2016 ) find that higher levels of customer activity on social media lead to disengagement (i.e., Facebook unlikes). Unlikes, as affective responses, may imply that customer responses may be polarized more easily if technologies (social media in their study) are not able to interact with customers appropriately.

Customer unreadiness

Another consideration is that customers may not be ready for interacting with feeling AI. Luo et al. ( 2019 ) find that many customers hang up on call-out marketing chatbots once they realize they are talking to bots. The Technology Readiness Index surveyed what people think about AI in the workplace and only 10% consider feeling AI to have the biggest impact in the past 5 years on their jobs, indicating that customers are not aware that AI can “have” feelings and may constitute a threat to their jobs (Espino 2019 ).

Contributions and conclusions

The most disruptive aspect of AI is that it replaces and improves upon human thinking capability. One of the most revolutionary characteristics of modern thinking AI is its ability to personalize by analyzing big data in an automatic way. This creates a quantum leap in marketing’s ability to target individual customers. Marketing primarily requires thinking intelligence and feeling intelligence. Until now there has been only limited ability of technology to help with those things. Now as thinking AI is advancing rapidly, it is gaining the ability to assume many of the thinking tasks in marketing. Eventually will even assume many of the feeling tasks in marketing, as AI develops further. Such efforts are already underway by researchers.

We see that marketers who cannot wait for technology to sufficiently advance use mechanical AI and thinking AI for feeling tasks, due to true feeling AI not being ready yet. We also see that AI intelligences may not be used in the most effective way (e.g., collecting customer data indiscriminately or accepting AI recommendation blindly). Thus, we develop this strategic framework to help marketers leverage the benefits of multiple AI intelligences for marketing impact. In this framework, we lay out the ways in which various AI intelligences can be used in marketing research, marketing strategy (STP), and marketing action (4Ps/4Cs). It shows the strategic roles that AI can play in marketing, as well as points out the limitations of current AI, to help marketers use AI wisely.

According to Huang and Rust ( 2018 ), true feeling AI needs to be able to recognize, emulate, and respond appropriately to human emotions.

The term “product” is used to refer to both tangible goods and intangible services.

The two strategic stages, marketing research and marketing strategy, are less relevant for this marketing practice illustration.

Agarwal, R., Dugas, M., Gao, G., & Kannan, P. K. (2020). Emerging technologies and analytics for a new era of value-centered marketing in healthcare. Journal of the Academy of Marketing Science, 48 (2), 9–23.

Article   Google Scholar  

Ascarza, E. (2018). Retention futility: Targeting high-risk customers might be ineffective. Journal of Marketing Research, 55 (1), 80–98.

Autor, D. H., & Dorn, D. (2013). The growth of low-skill service jobs and the polarization of the US labor market. American Economic Review, 103 (5), 1553–1597.

Avery, J. (2018). Tailor brands: Artificial intelligence-driven branding. Harvard Business School Case 519–017 , (August).

Avery, J., & Steenburgh, T. (2018). HubSpot and motion AI: Chatbot-enabled CRM. Harvard Business School case 518-067, February.

Balducci, B., & Marinova, D. (2018). Unstructured data in marketing. Journal of the Academy of Marketing Science, 46 (4), 557–590.

Bauer, J., & Jannach, D. (2018). Optimal pricing in e-commerce based on sparse and noisy data. Decision Support Systems, 106 (February), 53–63.

Berger, J., Humphreys, A., Ludwig, S., Moe, W. W., Netzer, O., & Schweidel, D. A. (2019). Uniting the tribes: Using text for marketing insight. Journal of Marketing, 84 (1), 1–25.

Blattberg, R. C., & Deighton, J. (1996). Manage marketing by the customer equity test. Harvard Business Review, 74 (4), 136–144.

Google Scholar  

Chen, Y., Lee, J. Y., Sridhar, S., Mittal, V., McCallister, K., & Singal, A. G. (2020). Improving cancer outreach effectiveness through targeting and economic assessments: Insights from a randomized field experiment. Journal of Marketing, 84 (3), 1–27.

Chintagunta, P., Hanssens, D. M., & Hauser, J. R. (2016). Editorial—Marketing science and big data. Marketing Science, 35 (3), 341–342.

Chung, T. S., Wedel, M., & Rust, R. T. (2016). Adaptive personalization using social networks. Journal of the Academy of Marketing Science, 44 (1), 66–87.

Chung, T. S., Rust, R. T., & Wedel, M. (2009). My mobile music: An adaptive personalization system for digital audio players. Marketing Science, 28 (1), 52–68.

Colgate, E., Wannasuphoprasit, W., & Peshkin, M. (1996). Cobots: Robots for collaboration with human operators. In Proceedings of the ASME Dynamic Systems and Control Division , New York, 58, 433-439.

Cooke, A. D. J., & Zubcsek, P. P. (2017). The connected consumer: Connected devices and the evolution of customer intelligence. Journal of the Association for Consumer Research, 2 (2), 164–178.

Daabes, A. S. A., & Kharbat, F. F. (2017). Customer-based perceptual map as a marketing intelligence source. International Journal of Economics and Business Research, 13 (4), 360–379.

Davenport, T., Guha, A., Grewal, D., & Bressgott, T. (2020). How artificial intelligence will change the future of marketing. Journal of the Academy of Marketing Science, 48 (2), 24–42.

Davis, E., & Marcus, G. (2015). Commonsense reasoning and commonsense knowledge in artificial intelligence. Communications of the ACM, 58 (9), 93–103.

Dekimpe, M. (2020). Retailing and retailing research in the age of big data analytics. International Journal of Research in Marketing, 37 , 3–14.

Deming, W. E. (1986). Out of the Crisis . Cambridge: Massachusetts Institute of Technology, Center for Advanced Engineering Study.

Donthu, N., & Rust, R.T. (1989). Estimating geographic customer densities using kernel density estimation. Marketing Science , 8 (2), 191–203.

Drew, J. H., Mani, D. R., Betz, A. L., & Datta, P. (2001). Targeting customers with statistical and data-mining techniques. Journal of Service Research, 3 (3), 205–219.

Dzyabura, D., & Hauser, J. R. (2011). Active machine learning for consideration heuristics. Marketing Science, 30 (5), 757–944.

Dzyabura, D., & Hauser, J. R. (2019). Recommending products when consumers learn their preferences weights. Marketing Science, 38 (3), 365–541.

Espino, A. (2019). Artificial intelligence: A snapshot into the future. The National Technology Readiness Survey , https://rockresearch.com/artificial-intelligence-snapshot-future/ .

Feng, J., Li, X., & Zhang, X. (2019). Online product reviews-triggered dynamic pricing: Theory and evidence. Information Systems Research, 30 (4), 1107–1123.

Fountaine, T., McCarthy, B., & Saleh, T. (2019). Building the AI-powered organization. Harvard Business Review , July–August, 63–73.

Frey, C. B., & Osborne, M. A. (2017). The future of employment: How susceptible are jobs to computerisation? Technological Forecasting and Social Change, 114 (January), 254–280.

Gabel, S., Guhl, D., & Klapper, D. (2019). P2V-MAP: Mapping market structures for large retail assortments. Journal of Marketing Research, 56 (4), 557–580.

Gali, N., Camprubi, R., & Donaire, J. A. (2017). Analyzing tourism slogans in top tourism destinations. Journal of Destination Marketing & Management, 6 (3), 243–251.

Gopinath, D. (2019). Human + machine: How content analytics delivers unsurpassed value to advertisers. MSI Lunch Lecture, (Sept 25).

Grewal, D., Motyka, S., & Levy, M. (2018). The evolution and future of retailing and retailing education. Journal of Marketing Education, 40 (1), 85–93.

Grewal, D., Noble, S. M., Roggeveen, A. L., & Nordfalt, J. (2020). The future of in-store technology. Journal of the Academy of Marketing Science, 48 (2), 96–113.

Guo, J., Zhang, W., Fan, W., & Li, W. (2018). Combining geographical and social influences with deep learning for personalized point-of-interest recommendation. Journal of Management Information Systems, 35 (4), 1121–1153.

Gupta, S., Lehmann, D. R., & Stuart, J. A. (2004). Valuing customers. Journal of Marketing Research, 41 (1), 7–18.

Hartmann, J., Huppertz, J., Schamp, C. P., & Heitmann, M. (2019). Comparing automated text classification methods. International Journal of Research in Marketing, 36 (1), 20–38.

Hewett, K., Rand, W., Rust, R. T., & van Heerde, H. (2016). Brand buzz in the echoverse. Journal of Marketing, 80 (3), 1–24.

Hoffman, D. L., & Novak, T. P. (2018). Consumer and object experience in the internet of things: An assemblage theory approach. Journal of Consumer Research, 44 (6), 1178–1204.

Huang, M. H., & Rust, R. T. (2017). Technology-driven service strategy. Journal of the Academy of Marketing Science, 45 (6), 906–924.

Huang, M. H., & Rust, R. T. (2018). Artificial intelligence in service. Journal of Service Research, 21 (2), 155–172.

Huang, M. H., & Rust, R. T. (2020). Engaged to a robot? The role of AI in service. Journal of Service Research , 109467052090226. https://doi.org/10.1177/1094670520902266 .

Huang, M. H., Rust, R. T., & Maksimovic, V. (2019). The feeling economy: Managing in the next generation of artificial intelligence (AI). California Management Review, 61 (4), 43–65.

Humphreys, A., & Wang, R. (2018). Automated text analysis for consumer research. Journal of Consumer Research, 44 (6), 1274–1306.

Kelly, S. D. (2019). What computers can’t create. MIT Technology Review, 122 (2), 68–75.

Kim, S. Y., Schmitt, B. H., & Thalmann, N. M. (2019). Eliza in the uncanny valley: Anthropomorphizing consumer robots increases their perceived warmth but decreases liking. Marketing Letters, 30 (1), 1–12.

Kirkpatrick, K. (2020). Tracking shoppers. Communications of the ACM, 63 (2), 19–21.

Kotler, P., & Keller, K. L. (2006). Marketing Management . Pearson Prentice Hall: Upper Saddle River.

Kumar, V., Rajan, B., Venkatesan, R., & Lecinski, J. (2019). Understanding the role of artificial intelligence in personalized engagement marketing. California Management Review, 61 (4), 135–155.

Lambrecht, A., & Tucker, C. (2019). Algorithmic bias? An empirical study of apparent gender-based discrimination in the display of STEM career ads. Management Science, 65 (7), 2966–2981.

Lauterborn, B. (1990). New marketing litany: Four Ps passé: C-words take over. Advertising Age, 61 (41), 26.

Lee, D., Hosanagar, K., & Nair, H. S. (2018). Advertising content and consumer engagement on social media: Evidence from Facebook. Management Science, 64 (11), 5105–5131.

Lehmann, D. R. (2020). The evolving world of research in marketing and the blending of theory and data. International Journal of Research in Marketing, 37 (1), 27–42.

Leung, E., Paolacci, G., & Puntoni, S. (2018). Human versus machine: Resisting automation in identity-based consumer behavior. Journal of Marketing Research, 55 (6), 818–831.

Lewis, T. G., & Denning, P. J. (2018). Learning machine learning. Communications of the ACM, 61 (12), 24–27.

Liebman, E., Saar-Tsechansky, M., & Stone, P. (2019). The right music at the right time: Adaptive personalized playlists based on sequence modeling. MIS Quarterly, 43 (3), 765–786.

Liu, X. (2020). De-targeting to signal quality. International Journal of Research in Marketing, 37 (2), 386–404.

Liu, X., Singh, P. V., & Srinivasan, K. (2016). A structured analysis of unstructured big data by leveraging cloud computing. Marketing Science, 35 (3), 363–388.

Longoni, C., Bonezzi, A., & Morewedge, C. K. (2019). Resistance to medical artificial intelligence. Journal of Consumer Research, 46 , 629–650.

Luo, X., Tong, S., Fang, Z., & Qu, Z. (2019). Frontiers: Machines versus humans: The impact of AI chatbot disclosure on customer purchases. Marketing Science, 38 (6), 937–947.

Ma, L., & Sun, B. (2020). Machine learning and AI in marketing – Connecting computing power to human insights. International Journal of Research in Marketing . https://doi.org/10.1016/j.ijresmar.2020.04.005 .

McDuff, D., & Czerwinski, M. (2018). Designing emotionally sentient agents. Communications of the ACM, 61 (12), 74–83.

Mende, M., Scott, M. L., van Doorn, J., Grewal, D., & Shanks, I. (2019). Service robots rising: How humanoid robots influence service experiences and elicit compensatory consumer responses. Journal of Marketing Research, 56 (4), 535–556.

Misra, K., Schwartz, E. M., & Abernethy, J. (2019). Dynamic online pricing with incomplete information using multiarmed bandit experiments. Marketing Science, 38 (2), 226–252.

Montes, R., Sand-Zantman, W., & Valletti, T. (2019). The value of personal information in online markets with endogenous privacy. Management Science, 65 (3), 1342–1362.

Netzer, O., Lemaire, A., & Herzenstein, M. (2019). When words sweat: Identifying signals for loan default in the text of loan applications. Journal of Marketing Research, 56 (6), 960–980.

Neumann, N., Tucker, C. E., & Whitfield, T. (2019). Frontiers: How effective is third-party consumer profiling? Evidence from field studies. Marketing Science, 38 (6), 918–926.

Ng, I. C. L., & Wakenshaw, S. Y. L. (2017). The internet-of-things: Review and research directions. International Journal of Research in Marketing, 34 (1), 3–21.

Novak, T. P., & Hoffman, D. L. (2019). Relationship journeys in the internet of things: A new framework for understanding interactions between consumers and smart objects. Journal of the Academy of Marketing Science, 47 , 216–237.

Ordenes, F. W., Ludwig, S., De Ruyter, K., & Grewal, D. (2017). Unveiling what is written in the stars: Analyzing explicit, implicit, and discourse patterns of sentiment in social media. Journal of Consumer Research, 43 (6), 875–894.

Pitt, C. S., Bal, A. S., & Plangger, K. (2020). New approaches to psychographic consumer segmentation: Exploring fine art collectors using artificial intelligence, automated text analysis and correspondence analysis. European Journal of Marketing . https://doi.org/10.1108/EJM-01-2019-0083 .

Power, B. (2017). How Harley-Davidson used artificial intelligence to increase New York sales leads by 2,930%. Harvard Business Review digital article, (may 30), https://hbr.org/2017/05/how-harley-davidson-used-predictive-analytics-to-increase-new-york-sales-leads-by-2930 .

Pulles, N. J., & Hartman, P. (2017). Likeability and its effect on outcomes of interpersonal interaction. Industrial Marketing Management, 66 , 56–63.

Rai, A. (2020). Explainable AI: from black box to glass box. Journal of the Academy of Marketing Science, 48 (1), 137–141.

Rogers, E. M. (1962). Diffusion of innovations (1st ed.). New York: Free Press of Glencoe.

Rust, R. T. (1997). The dawn of computer behavior: Interactive service marketers will find their customer is not human. Marketing Management , 6(fall), 31-34.

Rust, R. T. (2020). The future of marketing. International Journal of Research in Marketing, 37 (1), 15–26.

Rust, R. T., & Huang, M. H. (2020). The feeling economy: How artificial intelligence is creating the era of empathy. Palgrave-Macmillan.

Rust, R.T., Kamakura, W.A., & Alpert, M.I. (1992). Viewer preference segmentation and viewing choice models for network television. Journal of Advertising , 21 (1), 1–8.

Rust, R. T., Rand, W., Huang, M. H., Stephen, A. T., Brooks, G., & Chabuk, T. (2020). Real-time brand reputation tracking using social media. Working paper.

Schneider, M. J., Jagpal, S., Gupta, S., Li, S., & Yu, Y. (2017). Protesting customer privacy when marketing with second-party data. International Journal of Research in Marketing, 34 (3), 593–603.

Schneider, M. J., Jagpal, S., Gupta, S., Li, S., & Yu, Y. (2018). A flexible method for protecting marketing data: An application to point-of-sale data. Marketing Science, 37 (1), 153–171.

Schoenick, C., Clark, P., Tafjord, O., Turney, P., & Etzioni, O. (2017). Moving beyond the Turing test with the Allen AI science. Communications of the ACM, 60 (9), 60–64.

Simester, D., Timoshenko, A., & Zoumpoulis, S. I. (2020). Targeting prospective customers: Robustness of machine-learning methods to typical data challenges. Management Science, 66 (6), 2495–2522.

Singh, J., Nambisan, S., Bridge, R. G., & Brock, J. (2020). One-voice strategy for customer engagement. Journal of Service Research , 1–24. https://doi.org/10.1177/1094670520910267 .

Soleymanian, M., Weinberg, C. B., & Zhu, T. (2019). Sensor data and behavioral tracking: Does usage-based auto insurance benefit drivers? Marketing Science, 38 (1), 21–43.

Srinivasan, S., Rutz, O. J., & Pauwels, K. (2016). Paths to and off purchase: Quantifying the impact of traditional marketing and online consumer activity. Journal of the Academy of Marketing Science, 44 (4), 440–453.

Sutton, D. (2018). How AI helped one retailer reach new customers. Harvard Business Review , (may 28), https://hbr.org/2018/05/how-ai-helped-one-retailer-reach-new-customers .

Thomaz, F., Salge, C., Karahanna, E., & Hulland, J. (2020). Learning from the dark web: Leveraging conversational agents in the era of hyper-privacy to enhance marketing. Journal of the Academy of Marketing Science, 48 (2), 43–63.

Timoshenko, A., & Hauser, J. R. (2019). Identifying customer needs from user-generated content. Marketing Science, 38 (1), 1–20.

Tong, S., Luo, X., & Xu, B. (2020). Personalized mobile marketing strategies. Journal of the Academy of Marketing Science, 48 (2), 64–78.

Treacy, M., & Wiersema, F. (1997). The Discipline of Market Leaders . Reading, MA: Perseus.

Ukanwa, K., & Rust, R. T. (2020). Discrimination in service. Working paper.

Valls, A., Gibert, K., Orellana, A., & Anton-Clave, S. (2018). Using ontology-based clustering to understand the push and pull factors for British tourists visiting a Mediterranean coastal destination. Information & Management, 55 , 145–159.

van Doorn, J., Mende, M., Noble, S., Hulland, J., Ostrom, A., Grewal, D., & Petersen, A. (2017). Domo arigato Mr. Roboto: Emergence of automated social presence in organizational frontlines and customers’ service experiences. Journal of Service Research, 20 (1), 43–58.

Varki, S., Cooil, B., & Rust, R. T. (2000). Modeling fuzzy data in qualitative marketing research. Journal of Marketing Research, 37 (4), 480–489.

Wang, Q., Li, B., & Singh, P. V. (2018). Copycats vs. original mobile apps: A machine learning copycat-detection method and empirical analysis. Information Systems Research, 29 (2), 273–291.

Wang, Y. X., Ramanan, D., & Hebert, M. (2017). Learning to model the tail. 31st conference on neural information processing systems (NIPS).

Wedel, M., & Kannan, P. K. (2016). Marketing analytics for data-rich environments. Journal of Marketing, 80 (6), 97–121.

Wilson, H. J., & Daugherty, P. R. (2018). Collaborative intelligence: Humans and AI are joining forces. Harvard Business Review, July–August, 114–123.

Wirtz, J., Patterson, P. G., Kunz, W. H., Gruber, T., Lu, V. N., Paluch, S., & Martins, A. (2018). Brave new world: Service robots in the frontline. Journal of Service Management, 29 (5), 907–931.

Yadav, M. S., & Pavlou, P. A. (2020). Technology-enabled interactions in digital environments: A conceptual foundation for current and future research. Journal of the Academy of Marketing Science, 48 (2), 132–136.

Download references

This research was supported by grants (106-2410-H-002-056-MY3 and 107–2410-H-002-115-MY3) from the Ministry of Science and Technology, Taiwan.

Author information

Authors and affiliations.

Department of Information Management, College of Management, National Taiwan University, 1, Sec. 4, Roosevelt Rd, Taipei, 10617, Taiwan

Ming-Hui Huang

Center for Excellence in Service at the Robert H. Smith School of Business, University of Maryland, 3451 Van Munching Hall, College Park, MD, 20742, USA

Roland T. Rust

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Ming-Hui Huang .

Additional information

Publisher’s note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

John Hulland and Mark Houston served as Editors for this article.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Huang, MH., Rust, R.T. A strategic framework for artificial intelligence in marketing. J. of the Acad. Mark. Sci. 49 , 30–50 (2021). https://doi.org/10.1007/s11747-020-00749-9

Download citation

Received : 03 March 2020

Accepted : 29 September 2020

Published : 04 November 2020

Issue Date : January 2021

DOI : https://doi.org/10.1007/s11747-020-00749-9

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Artificial intelligence
• Machine learning
• Mechanical AI
• Thinking AI
• Strategic AI marketing
• Standardization
• Personalization
• Relationalization
• Find a journal
• Publish with us
• Track your research

• Search for: Toggle Search

‘You Transformed the World,’ NVIDIA CEO Tells Researchers Behind Landmark AI Paper

Of GTC ’s 900+ sessions, the most wildly popular was a conversation hosted by NVIDIA founder and CEO Jensen Huang with seven of the authors of the legendary research paper that introduced the aptly named transformer — a neural network architecture that went on to change the deep learning landscape and enable today’s era of generative AI.

“Everything that we’re enjoying today can be traced back to that moment,” Huang said to a packed room with hundreds of attendees, who heard him speak with the authors of “ Attention Is All You Need .”

Sharing the stage for the first time, the research luminaries reflected on the factors that led to their original paper, which has been cited more than 100,000 times since it was first published and presented at the NeurIPS AI conference. They also discussed their latest projects and offered insights into future directions for the field of generative AI.

While they started as Google researchers, the collaborators are now spread across the industry, most as founders of their own AI companies.

“We have a whole industry that is grateful for the work that you guys did,” Huang said.

Origins of the Transformer Model

The research team initially sought to overcome the limitations of recurrent neural networks , or RNNs, which were then the state of the art for processing language data.

Noam Shazeer, cofounder and CEO of Character.AI, compared RNNs to the steam engine and transformers to the improved efficiency of internal combustion.

“We could have done the industrial revolution on the steam engine, but it would just have been a pain,” he said. “Things went way, way better with internal combustion.”

“Now we’re just waiting for the fusion,” quipped Illia Polosukhin, cofounder of blockchain company NEAR Protocol.

The paper’s title came from a realization that attention mechanisms — an element of neural networks that enable them to determine the relationship between different parts of input data — were the most critical component of their model’s performance.

“We had very recently started throwing bits of the model away, just to see how much worse it would get. And to our surprise it started getting better,” said Llion Jones, cofounder and chief technology officer at Sakana AI.

Having a name as general as “transformers” spoke to the team’s ambitions to build AI models that could process and transform every data type — including text, images, audio, tensors and biological data.

“That North Star, it was there on day zero, and so it’s been really exciting and gratifying to watch that come to fruition,” said Aidan Gomez, cofounder and CEO of Cohere. “We’re actually seeing it happen now.”

Envisioning the Road Ahead 

Adaptive computation, where a model adjusts how much computing power is used based on the complexity of a given problem, is a key factor the researchers see improving in future AI models.

“It’s really about spending the right amount of effort and ultimately energy on a given problem,” said Jakob Uszkoreit, cofounder and CEO of biological software company Inceptive. “You don’t want to spend too much on a problem that’s easy or too little on a problem that’s hard.”

A math problem like two plus two, for example, shouldn’t be run through a trillion-parameter transformer model — it should run on a basic calculator, the group agreed.

They’re also looking forward to the next generation of AI models.

“I think the world needs something better than the transformer,” said Gomez. “I think all of us here hope it gets succeeded by something that will carry us to a new plateau of performance.”

“You don’t want to miss these next 10 years,” Huang said. “Unbelievable new capabilities will be invented.”

The conversation concluded with Huang presenting each researcher with a framed cover plate of the NVIDIA DGX-1 AI supercomputer, signed with the message, “You transformed the world.”

There’s still time to catch the session replay by registering for a virtual GTC pass — it’s free.

To discover the latest in generative AI, watch Huang’s GTC keynote address:

NVIDIA websites use cookies to deliver and improve the website experience. See our cookie policy for further details on how we use cookies and how to change your cookie settings.

Share on Mastodon