The Big Thinkers Shaping Enterprise AI Frontier Programs

Why This Map Matters

Enterprise AI strategy is not a field with one consensus view. It is a genuine intellectual debate — conducted across business schools, economics departments, practitioner blogs, and boardrooms — about questions that don't yet have settled answers: Is AI primarily an economic disruption or an organizational design challenge? Should you automate decisions or augment the humans who make them? Is the bottleneck technology, data, or leadership? Does scale win, or does judgment?

The people below are not merely writing about this space — they are actively shaping how practitioners think about it. Understanding their arguments, their disagreements, and where they're moving toward common ground is one of the fastest ways to develop a coherent strategic position of your own.

The Thinkers: An Overview Map

The Thinkers in Depth

1. Marco Iansiti & Karim R. Lakhani

The Operating Model Thesis

Core argument: Traditional firms are constrained by scale (you need more people to do more), scope (it's hard to straddle industries), and learning (knowledge is siloed). AI removes all three constraints simultaneously. When AI runs the core processes, firms can scale without adding headcount, operate across industries, and continuously improve without human bottlenecks. The competitive implication is that AI-native firms aren't just more efficient — they operate under a different set of rules. Traditional firms that layer AI on top of existing operating models will be disrupted by firms that rearchitect around AI from the inside out.

Key framework: The "AI factory" — the concept of a firm in which data and analytics pipelines replace traditional human-run processes at the core. This is the Harvard Business School origin of the term that has since entered mainstream enterprise vocabulary.

What they get right for enterprise programs: Their framing of operating model rearchitecture — not tool adoption — as the central challenge is arguably the most important single idea in enterprise AI strategy. Their empirical work on "collisions" between AI-native and traditional firms explains why incumbents are losing ground even when they spend heavily on AI.

Where they are challenged: Critics argue the Iansiti/Lakhani thesis was written primarily with AI-native firms (Alibaba, Amazon, Ocado) as the reference case, and that the prescription of "rearchitecting around AI" is vastly easier for greenfield companies than for incumbents with 50-year-old processes and legacy infrastructure. The operating model transformation they describe takes 5–10 years for a large enterprise — which the brisk pace of their prose somewhat obscures.

Essential reading:

• Competing in the Age of AI: Strategy and Leadership When Algorithms and Networks Run the World (Harvard Business Review Press, 2020) — the primary text

• "Competing in the Age of AI," Harvard Business Review, January–February 2020 — the condensed article version

• "Managing Our Hub Economy," HBR, 2017 — the earlier framing of platform power dynamics

2. Erik Brynjolfsson

The Productivity J-Curve Thesis

Core argument: Brynjolfsson's contribution is simultaneously the most hopeful and the most cautionary in this field. His "productivity paradox" — first articulated for IT in the 1990s and revived for AI in 2017 — observes that transformative technologies consistently fail to show up in productivity statistics for years or decades after widespread adoption. The reason is not that the technology is ineffective: it is that capturing value from a general purpose technology requires massive complementary investments in organizational redesign, skills, and new processes. His "Productivity J-Curve" framework models how this works: measured productivity initially declines as organizations invest in intangible capital (retraining, process redesign, governance) before eventually producing returns that dramatically exceed the initial investment.

For enterprise AI programs, this is the theoretical foundation for why the pilot purgatory problem exists — and also why organizations that are investing in the unglamorous 70% (people, process, workflow redesign) will eventually emerge with advantages that are very hard to replicate quickly.

2025 update — the GenAI divide: Brynjolfsson's 2025 co-authored work (including the Stanford Digital Economy Lab's enterprise AI playbook with Pereira and Graylin) empirically confirms what the J-curve predicts: a sharp divergence between the ~5% of organizations that are generating significant AI value and the ~95% that are not, despite equivalent access to the technology. The differentiator is always organizational, never technological.

What he gets right: The J-curve is the most rigorous explanation available for why AI investment is not yet showing up at scale in macroeconomic productivity statistics, and it gives enterprise programs a framework for explaining to impatient boards why the returns to early investment are slow and why they will eventually accelerate.

Where he is challenged: Optimists argue that generative AI is different from prior general purpose technologies in important ways — notably that its implementation barriers are lower (you don't need to retool a factory, you need to write a prompt) and that the complementary investments required are smaller. Whether this changes the shape of the J-curve, or merely compresses its timeline, is an open empirical question.

Essential reading:

• "AI and the Modern Productivity Paradox: A Clash of Expectations and Statistics," NBER Working Paper 24001, 2017/2019 — the foundational paper

• "The Productivity J-Curve: How Intangibles Complement General Purpose Technologies," NBER Working Paper 25148, 2019 — the mechanism

• The Enterprise AI Playbook: Lessons from 51 Successful Deployments (Stanford Digital Economy Lab, March 2026) — the most recent empirical field work, co-authored with Pereira and Graylin

• Race Against the Machine (with Andrew McAfee, 2011) — the earlier and more accessible framing of technology-labor complementarity

3. Ajay Agrawal, Joshua Gans & Avi Goldfarb

The Prediction Machine / Decision Architecture Thesis

Core argument: This trio from the University of Toronto's Rotman School has produced the most rigorous economic framework for thinking about what AI actually does and why it disrupts. Their first book, Prediction Machines (2018), argued that AI is fundamentally a technology for making prediction cheap — and that this one insight, followed through carefully, explains most of what AI will and won't be able to do. Their second book, Power and Prediction (2022), extended the argument: the key unit of analysis is not the task or the job but the decision. Every decision has two ingredients — prediction (provided by AI) and judgment (remaining with humans). As AI improves prediction dramatically, it decouples prediction from judgment in decisions that were previously fused, creating what they call "the Between Times" — the period before organizations have redesigned their decision architectures to take advantage of separated prediction and judgment.

Their framework generates a powerful taxonomy for enterprise use case prioritization:

• Point solutions — AI improves a single task; lowest disruption, easiest adoption

• Application solutions — AI enables a new process without changing the surrounding system

• System solutions — AI requires redesigning the entire decision-making system; highest disruption, highest value, most complex to achieve

What they get right: The prediction/judgment framework is one of the most practically useful analytical tools available for AI leads thinking about which use cases to prioritize and why. It also explains why system-level AI redesigns take so much longer than point solutions — they require renegotiating organizational power structures, not just deploying technology.

Where they are challenged: Critics argue the prediction/judgment distinction may break down for generative AI, which can increasingly perform reasoning that looks more like judgment than prediction. The framework was developed primarily with supervised machine learning in mind; how it maps onto large language models and agentic systems is an area of active theoretical development.

Essential reading:

• Prediction Machines: The Simple Economics of Artificial Intelligence (Harvard Business Review Press, 2018)

• Power and Prediction: The Disruptive Economics of Artificial Intelligence (Harvard Business Review Press, 2022)

• "How to Win with Machine Learning," Harvard Business Review, September–October 2020

• "The Economics of Transformative AI" (with Brynjolfsson), University of Chicago Press, December 2025

4. Andrew Ng

The Practitioner Playbook Thesis

Core argument: Ng's contribution to this field is less theoretical and more practical than the academics above — but it has arguably had the broadest direct influence on enterprise AI program design. His framing is optimistic and operationally specific: AI transformation is not a mystery, it is a repeatable set of steps that any large enterprise can follow. The steps he identified — pilot projects to build momentum, build an in-house AI team, provide broad AI training, develop an AI strategy, develop internal and external communications — have been absorbed into the operating vocabulary of enterprise AI programs worldwide.

His metaphor that AI is "the new electricity" — a general purpose technology that will transform every industry the way electricity did in the early 20th century — has also been widely adopted, and usefully positions AI as infrastructure rather than application. The practical implication he draws from this: just as every company eventually needed an electricity strategy (it would be absurd to think of some companies "having electricity" and others not), every company will eventually need an AI strategy that is integrated into its core operations.

Ng's current work: Through Landing AI and the AI Fund, Ng has been particularly focused on industrial AI — applying machine learning to manufacturing, quality control, and supply chain in sectors that are not primarily software companies. This has shaped his emphasis on the importance of bridging domain expertise and AI expertise, which he identifies as the single most important talent bottleneck.

What he gets right: The democratization argument — that AI transformation is achievable for most enterprises if approached with discipline and realism — is an important corrective to the "only digital natives can do this" fatalism that paralyzes many incumbent organizations. His emphasis on building internal capability rather than outsourcing AI is well-grounded in the evidence.

Where he is challenged: Critics note that the Ng playbook was developed primarily in the supervised learning era, and that the generative AI and agentic AI environment introduces governance, risk, and operating model complexities that the original playbook did not address. His optimism about the replicability of AI transformation is also challenged by the persistence of the pilot purgatory problem despite widespread adoption of his framework.

Essential reading:

• AI Transformation Playbook: How to Lead Your Company into the AI Era (Landing AI, free PDF, 2019) — the definitive practitioner reference

• AI for Everyone (Coursera course, ongoing) — the most widely distributed executive-level AI literacy resource

• deeplearning.ai newsletters and blog — active commentary on current developments

5. Ethan Mollick

The Co-Intelligence / Human-AI Partnership Thesis

Core argument: Mollick's distinctive contribution is a behavioral and management lens on what AI adoption actually requires at the human level. Where most thinkers in this space write from a strategy or economics perspective, Mollick writes from the perspective of someone who has spent years in the classroom running live experiments with AI — watching how real people actually interact with these systems and what that means for how organizations should think about adoption.

His central concept is the "jagged frontier" — the observation that AI capability is deeply uneven and counterintuitive. AI can write a compelling legal brief and fail at elementary arithmetic. It can pass a medical licensing exam and confidently hallucinate citations. This unevenness means that the standard "automate the routine, protect the creative" heuristic frequently gets it wrong. The tasks AI is best at are not always the ones that feel routine; the tasks it is worst at are not always the ones that feel complex. The implication for enterprise programs is that you cannot determine AI's appropriate role in a workflow without actually experimenting with it in that specific context.

His four practical rules for human-AI collaboration — Always Invite AI to the Table, Be the Human in the Loop, Treat AI Like a (Smart but Alien) Person, Assume It's the Worst AI You'll Ever Use — are among the most practically applicable guidance for individual workers and AI leads trying to develop appropriate working relationships with AI systems.

His current blog: One Useful Thing (Substack) — arguably the most actively read ongoing commentary on AI and work, read by hundreds of thousands of practitioners. Mollick publishes field experiments, research syntheses, and practical guidance in real time, making it a genuinely useful feed for AI leads tracking the fast-moving capability frontier.

What he gets right: The jagged frontier framing is the most useful single concept for explaining to executives why AI pilots produce surprising results, why blanket policies ("use AI for X, not Y") fail, and why the only real answer is broad experimentation. His behavioral observation that "frontier workers" — those who push AI use far beyond the median — are the drivers of organizational AI value is well-supported by the OpenAI enterprise data.

Where he is challenged: Some enterprise practitioners find Mollick's individual-level framing (one person working with AI) less directly applicable to the system-level and governance challenges of large enterprise deployment. His optimistic tone about individual experimentation can also sit in tension with the reality that uncoordinated individual AI use creates the shadow AI problem at scale.

Essential reading:

• Co-Intelligence: Living and Working with AI (Portfolio/Penguin, 2024) — the primary text

• One Useful Thing Substack (substack.com/ethanmollick) — ongoing field experiments and research synthesis

• "Cyborgs, Centaurs and Self-Automators: The Three Modes of Human-GenAI Knowledge Work" (with Lakhani et al., HBS, 2025) — the research paper underpinning the human-AI work typology

6. David De Cremer

The Leadership Abdication Thesis

Core argument: De Cremer's argument is the most uncomfortable one in this field for enterprise executives, which is probably why it is also the most necessary. His diagnosis: the single biggest cause of enterprise AI failure is not bad technology, bad data, or bad operating models — it is leadership abdication. Executives, intimidated by the technical complexity of AI, have systematically ceded ownership of AI strategy to data science teams, IT departments, and external vendors who do not have a full view of the organization's goals, culture, or stakeholders. The result is that AI initiatives are designed and deployed without genuine leadership engagement, and then fail to achieve adoption or business impact because nobody with real authority has committed to making them work.

His prescription — nine leadership behaviors for AI-savvy executives — is grounded in behavioral science rather than strategy theory. The behaviors are deliberately not technical: they are about vision-setting, communication, stakeholder management, and accountability structures. His central thesis is that the skills required to lead AI transformation are the same as the skills required to lead any major organizational transformation — but that leaders have convinced themselves AI requires special deference to technologists that they would never show in, say, a major M&A integration.

His concept of "complementary intelligence" is also notable: the argument that as AI automates the predictable and the codifiable, the specifically human skills that remain valuable — empathy, ethical reasoning, contextual judgment, political navigation — become more important, not less. The appropriate leadership response to AI is to invest in developing those human capabilities, not to compete with AI on its own terms.

What he gets right: The leadership abdication diagnosis is strongly supported by the empirical literature. BCG's finding that active executive sponsorship makes firms 1.8× more likely to scale AI effectively, and McKinsey's finding that only 21% of organizations have redesigned any workflows around AI, both point to the same root cause De Cremer identifies. His behavioral science framing also provides a more actionable leadership development agenda than strategy texts typically offer.

Where he is challenged: Some practitioners find the emphasis on human leadership somewhat reactive — it is most useful as a corrective to over-technologized approaches, but it can underweight the genuine technical challenges of data quality, platform architecture, and model governance that cannot simply be solved by better leadership communication.

Essential reading:

• The AI-Savvy Leader: Nine Ways to Take Back Control and Make AI Work (Harvard Business Review Press, 2024) — the primary text

• Leadership by Algorithm: Who Leads and Who Follows in the AI Era (2020) — the earlier framing

• Named one of Forbes' 10 Must Read Tech Books of 2024

7. The Stanford Digital Economy Lab (Brynjolfsson, Pereira, Graylin)

The Field Evidence Thesis

Core argument: The March 2026 publication of The Enterprise AI Playbook: Lessons from 51 Successful Deployments represents the most recent and most empirically grounded contribution to this field. Based on 51 deep case studies of successful enterprise AI deployments, 150 executive interviews, and 350 employee surveys, the central finding is stark: "Same technology, same use cases, vastly different outcomes. The difference was never the AI model. It was always the organization."

The research identifies specific organizational factors that separate successful deployments from failures: the presence of leaders who understand both domain processes and AI capabilities simultaneously (the "bridge" talent problem); sponsor continuity through failure (in every case where a failed pilot eventually produced a successful deployment, the same executive was involved in both); and the decisive role of cross-functional ownership — the point at which departmental AI initiatives must escalate to corporate OKRs to break through cross-functional resistance.

What makes this contribution distinctive: Where most of the thinkers above reason from theory or individual observations, the Stanford DEL work is grounded in systematic empirical evidence across a large number of cases. It provides the evidentiary foundation for claims that have often been asserted without rigorous support.

Essential reading:

• The Enterprise AI Playbook: Lessons from 51 Successful Deployments (Stanford Digital Economy Lab, March 2026) — available free at digitaleconomy.stanford.edu

Where the Thinkers Converge

Despite different lenses and methodologies, a remarkably consistent picture emerges across these perspectives.

1. The technology is not the constraint. Every thinker in this field, from the most economics-oriented (Brynjolfsson) to the most practitioner-oriented (Ng), reaches the same conclusion: the constraint on AI value creation is organizational, not technological. Models are accessible. Infrastructure is available. The bottlenecks are operating model design, data quality, governance architecture, leadership engagement, and human capability.

2. Complementarity, not substitution. From Brynjolfsson's productivity paradox (value requires complementary organizational investment) to De Cremer's human leadership thesis (human capabilities become more valuable as AI automates prediction) to Mollick's co-intelligence framing (AI as partner, not replacement) to Agrawal/Gans/Goldfarb's prediction/judgment separation (AI handles prediction, humans retain judgment) — all of these thinkers are arguing, in different vocabularies, for the same underlying point: the most value comes from AI and human capabilities working together, not from AI displacing human capabilities.

3. System redesign, not tool adoption. Iansiti/Lakhani's operating model rearchitecture, Agrawal/Gans/Goldfarb's "system solutions," and Ng's organizational transformation playbook all arrive at the same prescription: the organizations that capture significant AI value are the ones that redesign processes and decision-making systems around AI, not the ones that bolt AI onto existing processes.

4. Implementation lags are real and long. Brynjolfsson's J-curve, Ng's "it will be hard," Mollick's "ten hours to get it," and the Stanford DEL's "transformation measured in weeks and others measured in years" all point to the same empirical reality: significant AI value takes longer to realize than pilots suggest.

Where the Thinkers Diverge

The pace question. Brynjolfsson's J-curve implies a relatively patient view — the benefits are real but they take time to materialize as organizations build complementary capabilities. Ng's playbook implies a more aggressive timeline — disciplined execution of a known process should produce results within 12–18 months. The empirical evidence from the Stanford DEL suggests both are right: individual deployments can move very quickly, but enterprise-wide transformation takes years.

The human displacement question. De Cremer and Mollick both argue for a strongly human-centered view — AI augments rather than replaces, and the most valuable human capabilities become more important. Iansiti and Lakhani's framework is more agnostic — they describe firms where "humans are moved off to the side" as the AI factory runs core processes, without treating this as necessarily problematic. Brynjolfsson's productivity research suggests task reallocation (some roles disappear, new ones emerge) rather than net replacement, but is careful not to predict the balance.

The strategy vs. execution question. Iansiti/Lakhani and Agrawal/Gans/Goldfarb are primarily concerned with strategic positioning — which firms win, how competitive advantage shifts. Ng, Mollick, and De Cremer are primarily concerned with organizational execution — how specific programs, teams, and leaders make AI work in practice. These are complementary rather than conflicting, but practitioners often find the strategy frameworks under-specific about how and the execution frameworks under-ambitious about the magnitude of change required.

The governance question. De Cremer is the most insistent that governance and ethical AI are not afterthoughts — they are central leadership responsibilities that must be designed in from the start. The economists tend to treat governance as a constraint on deployment rather than as a capability to build. Enterprise practitioners in regulated industries tend to side with De Cremer on this point, and the EU AI Act (August 2026) has made his argument considerably more urgent.

The Reference Library

Essential Books

Essential Articles and Research Papers

Essential Ongoing Reading

A Reading Path by Role

For the executive who needs strategic framing fast: Start with Iansiti & Lakhani's HBR article (2020, 7 pages), then read the Agrawal/Gans/Goldfarb QRCA summary of Power and Prediction, then the Deloitte and McKinsey State of AI reports for current data.

For the AI lead who needs operational depth: Start with Ng's AI Transformation Playbook (11 pages, free), then Mollick's Co-Intelligence for the behavioral layer, then the Stanford DEL Enterprise AI Playbook (March 2026) for empirical evidence on what works.

For the executive thinking about governance and leadership culture: Start with De Cremer's The AI-Savvy Leader, then Mollick's co-intelligence framework for the workforce engagement layer, then the EU AI Act compliance guidance for the regulatory dimension.

For the strategist thinking about competitive positioning: Read Iansiti & Lakhani's full book, then Power and Prediction for the decision architecture framework, then Brynjolfsson's J-curve work for the honest timeline picture.

For ongoing monitoring: Subscribe to Mollick's One Useful Thing Substack. Read the OpenAI State of Enterprise AI and McKinsey State of AI reports annually. Track Stanford DEL publications for new empirical evidence.

April 2026 | This guide covers publicly available work. All thinkers' positions evolve — check their current publications and blogs for the most recent views.

Francisco Marques da Silva Director-level AI & Data practitioner, 20+ years building and leading transformation programmes across industries and borders. Publishing under the aiOps3000 brand, on aiops3000.com · Hashnode · LinkedIn