AI Academy · Book
Executives & Directors · Module 01 · Chapter 005

AI Leaders vs AI Followers

The best AI models are now within a few points of one another, and almost any company can use them. What separates leaders from followers is a loop of behaviors: choose a real problem, own it, change the work, and prove the result before scaling or stopping. Most organizations lead in some functions and follow in others, and following is a position, not a verdict.

≈ 15 min read

After this chapter you can

  • Explain why access to capable AI models no longer separates leaders from followers, using the convergence of model quality.
  • Describe the eight leader behaviors as a loop of four moves - choose, own, change and prove.
  • Cite evidence that management behavior, not technology, explains different results from the same technology.
  • Recognize leader and follower as patterns that coexist inside one company, and diagnose them function by function.
  • Name the one behavior gap whose closure would create the most business impact in twelve months, and its owner.

In August 2024, the best AI model that a company could only rent outperformed the best model it could download and run itself by just 0.5 percent on Arena, a public leaderboard where people compare answers without knowing which model wrote them. A year of fast releases followed, and by March 2026 the top closed model’s lead had grown to 3.3 percent. In the same ratings, models from six developers on two continents shared the top tier1. The gap, measured as the percentage difference between the two models’ Arena scores, moves in both directions, but it stays within a few points.

The lead of the best closed model over the best open-weight model was 0.5 percent in August 2024 and 3.3 percent in March 2026; it stays within a few points, so access no longer separates competitors.0.5%Closed vs open model gapAugust 20243.3%Closed vs open model gapMarch 2026Source: Stanford AI Index 2026, Arena · Mar 2026
Figure 1.5.1 The gap moves but stays within a few points. Access to a capable model is no longer something a competitor lacks.

For an executive, the leaderboard settles a strategic question before it is asked. Whatever model your organization uses, your strongest competitor can use the same one, or one almost as good, within months. Why AI, Why Now? showed how quickly the price of using such models has fallen. If access is equal on both sides of a market, access cannot be what puts one side ahead.

Yet companies with the same access are getting very different results. AI Is Changing Everything described how wide that gap between use and value has become. The question here is what the organizations on the right side of it do differently.

Same technology, different management

The answer is older than generative AI. In the decade after 1995, American productivity accelerated while Europe’s did not, mostly in industries that used information technology heavily. Nicholas Bloom, Raffaella Sadun and John Van Reenen asked whether the difference lay in the technology or in the firms. They studied establishments in Europe owned by American multinationals and compared them with establishments owned by other multinationals, operating in the same countries and investing in the same kinds of IT. The American-owned firms got more productivity out of the same IT, and the authors traced the advantage mainly to how they managed people: how they hired, promoted, paid, kept and removed staff2.

Same technology, different management: that is the whole argument in one study. Earlier firm-level work on computers found the same thing. Companies that changed their work practices along with their systems got far more from the same spending than companies that automated what they already did3.

The AI evidence points the same way. A 2020 survey of more than 3,000 managers by MIT Sloan Management Review and BCG found that having the right data, technology and talent was not enough: only 20 percent of companies with those fundamentals reported significant financial benefits from AI. The share rose to 73 percent among those that also learned with AI, changing how people worked as they went4. These are survey answers reported by managers, and the data predate generative AI, so they show a pattern, not a cause. But the pattern lines up with three decades of research on general-purpose technology: the tool is available to everyone, and the behavior around it is not. The step from the first group to the last was not a better model. It was a behavior.

So it helps to define the terms with care. An AI leader is not a company with the most licenses or the newest model. It is an organization whose habits turn widely available AI into measured changes in how work is done. An AI follower has the same tools but starts with them, spreads its effort thin and counts activity. Both are patterns of behavior, and both can be found inside the same company.

The loop that leaders close

The habits that separate the two groups are not exotic. Taken together, they form a loop with four moves, and each turn of it teaches the organization something it did not know before.

Leaders run a loop of four moves - choose, own, change, prove - and learn faster than competitors with every turn; followers break the loop at one of the moves.ChooseA real problem,few prioritiesOwnA senior sponsor,controls built inChangeRedesigned work,equipped peopleProveOutcomes decide:scale or stopLearn fasterthan your competitors
Figure 1.5.2 Eight behaviors in four moves. Leaders close the loop; followers break it somewhere.

Choose means starting with a business problem and keeping the list short. Own means a senior sponsor who can fund the work and stop it, with controls designed in from the start. Change means redesigning the work and equipping the people who do it. Prove means judging the result against what came before, then scaling what works and stopping what does not. The lesson from the last turn makes the next choice better.

Followers are rarely lazy or foolish. They are usually busy, fragmented or early, and they break the loop at one point: the work starts with a tool, nobody senior owns it, the old process stays, or pilots run on because nobody decides. Each move rests on two behaviors, eight in all: a real problem and few priorities; a senior sponsor and controls built in; redesigned work and equipped people; outcome measures and the discipline to scale or stop. Each break has the same cost. The organization stops learning, while a competitor with the same model keeps going.

Choose: a problem before a product

Followers tend to begin with a technology question: we have a new model, so what can we do with it? That question produces demonstrations. Leaders begin with a business question: which important problem should we solve? “Where can we use a chatbot?” sends a team shopping. “Why do our trucks still break down between services when the warning signs are already in our systems?” sends it to the process. A simple test works in most steering meetings. If the first sentence of an AI program names a product, the program is probably following. If it names a business pain, it may be leading.

Choosing also means narrowing. BCG’s 2024 survey of 1,000 senior executives found that the companies it classed as AI leaders pursued, on average, about half as many AI opportunities as their peers and scaled more than twice as many AI products and services5.

BCG found that AI leaders pursued about half as many opportunities as other companies and scaled more than twice as many, so focus, not volume, produced results.About ½Opportunities pursuedLeaders vs other companiesMore than 2×Products and services scaledLeaders vs other companiesSource: BCG survey of 1,000 executives · 2024
Figure 1.5.3 Leaders do fewer things and finish more of them.

The reason is attention. Disconnected pilots compete for the same executives, data specialists, security reviewers and managers, and none of them gets enough to finish. Richard Rumelt’s warning about strategy applies directly: a long list of goals is not a strategy, because a strategy chooses6. Fifty ideas are a brainstorm. A handful of sponsored problems is a program. How to score and rank candidates is the subject of Prioritizing the AI Portfolio in Module 09, and how to pick a first project is covered in Where Should We Start With AI? later in this module. The behavior here is the willingness to choose, and to leave good ideas on the table.

Own: a senior owner, with controls built in

Grassroots experimentation is valuable, and it is fragile. People will try new tools on their own. Without a senior owner, their best experiments stay in personal habits and never change how the organization works. McKinsey’s 2026 global survey found that respondents at the companies it calls AI high performers were about twice as likely as others to say that their senior leaders demonstrate commitment to AI initiatives7. The answers are self-reported, but they point the same way as the older research on management.

Ownership has a practical meaning. It is not the chief executive choosing a model or a keynote that mentions AI. It is a named senior leader who can fund work past the pilot, remove the barriers that stop it, and close it when it fails. The test is simple: who can stop a bad pilot, and who can fund a good one into production? If the answer to both is nobody, the organization is following, whatever its slides say.

Ownership on its own is not enough. It has to come with controls.

Leaders combine senior ownership with controls built into the work, which gives responsible speed; ownership alone is exposed, controls alone stall, and neither leaves scattered pilots.HighLowSeniorownershipLowControls built into the work · HighFast but exposedOne leak halts the programResponsible speedOwned and governedScattered pilotsExperiments stay in drawersPolicy, no progressQueues and quiet workarounds
Figure 1.5.4 Speed that lasts comes from ownership and controls together, designed in from the first experiment.

Each quadrant is familiar. With neither ownership nor controls, enthusiasm produces scattered pilots. Controls without an owner produce policies and approval queues; nothing is funded, and people quietly use public tools instead. An owner without controls moves fast until a data leak or a biased answer forces a freeze. Leaders sit in the top right: privacy, security, intellectual property and human oversight are part of the first experiment, not a review that arrives after the damage. If a program cannot scale because it has no controls, it is not fast. It is unfinished. Committees and approval gates belong to The AI Governance Operating Model in Module 07; who decides what belongs to AI Decision Rights and Accountability in Module 04.

Change: redesign the work, equip the people

The third move is where much of the value is made and many programs stall. Followers add AI to the old process: a new button on the old screen, with the same steps, the same targets and the same handoffs. One step gets faster, and the delay stays where it was. Leaders ask a harder question: if we designed this process today, with AI available from the first day, how would it work? Then they change who does the first pass, who handles exceptions, what gets measured and where a person still owns the outcome. This is the transformation level from AI Is Changing Everything, seen as a habit rather than a goal, and the bend from faster people to a changed process that The AI Adoption Curve described.

Followers bolt AI onto the old process and speed up one step while the delay stays; leaders redesign the work and train people in the real job, so the work itself changes.BOLTED ONAI added to the old processOne step faster, same delayREDESIGNEDNew first pass, new roles,people trained in the real jobThe work itself changesvs
Figure 1.5.5 A login is not a redesign. Leaders change the work, not only the screen.

Redesign fails without the people who must work in the new way. Bloom, Sadun and Van Reenen’s American multinationals did not win on hardware; they won on how they managed people2. A license is not adoption, and a town hall is not training. Useful training is practice in the real job, with the real data rules, under a manager who knows what good work looks like. Followers buy tools and hope enthusiasm does the rest. How to build capability at scale is the subject of The AI Talent Gap later in this module. In the European Union, organizations that deploy AI also have a legal duty, as of October 2026, to take measures to support AI literacy among their staff, as The AI Adoption Curve explains.

Prove: scale with evidence, stop with a lesson

Followers report activity: pilots launched, licenses issued, demonstrations given. Leaders compare outcomes with a baseline, such as cycle time, error rate or customer effort before and after. Without a baseline, an improvement is a story. How to build those measures is the subject of Measuring AI Business Value; the behavior that matters here is what leaders do with the answer.

Leaders decide on evidence - work that beats its baseline is scaled with a standard workflow, controls and coaching; work that does not is stopped, the lesson recorded and the people moved to what works.Did it beatits baseline?YesScale itStandard workflowand controlsManagers who can coach itNoStop itWrite down the lessonMove people to what works
Figure 1.5.6 Proof matters only when it forces a decision.

Scaling is not more users on the same demonstration. It is a standard workflow, controls that travel with it, measures that still hold at volume and managers who can coach it. Stopping matters just as much, and it is where followers often struggle. In 2024 Gartner predicted that at least 30 percent of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, rising costs or unclear business value8. That is a forecast, not a measurement, but the distinction it points to is real. Abandonment is what happens when nobody decides: a project fades, its people drift and nothing is learned. Stopping is a decision. The project ends on a set date, the team writes down why it did not work, and its people move to something that does. Leaders make stopping respectable. Followers keep every pilot alive because ending one feels like failure, and so they never free the attention that the next good problem needs.

Both columns in one company

It would be convenient if organizations sorted neatly into leaders and followers. They do not. A US Census Bureau survey of management practices at about 35,000 manufacturing plants found that 40 percent of the variation in those practices occurs within the same firm, and that the practices accounted for more than 20 percent of the variation in productivity9.

The study predates generative AI and covers factories, not AI programs, but the lesson transfers. One company can lead in one function and follow in the next. The honest diagnosis is mixed, function by function, and it should name a strength as well as a gap. It also removes an excuse. If leadership is a set of behaviors, an organization that is behind today can start the loop on one problem tomorrow. Behind is a position. It is not a personality.

Story: a day with a mining company’s operations chief

The chief operating officer of a mining company runs copper mines on three continents, from the pit to the port. Her day shows both columns inside one company.

Over one day the operations chief chooses a problem over a demo, sponsors a pilot, sees redesigned maintenance planning, scales one pilot and stops another, finds follower habits in procurement and reports a mixed diagnosis.08:00A vendordemoShe asks whichproblem it solves09:30A sponsordecidesFunds themaintenancepilot's next phase11:00The planningdeskAI drafts,planners judge14:00The pilotreviewScale one,stop one16:00ProcurementManyexperiments andno owner18:30A note tothe CEOWe lead inmaintenance,follow inprocurement
Figure 1.5.7 One executive, one day, both columns. Every move she makes is a behavior; none needs a better model.

At eight o’clock a vendor presents a newer model and a long list of possible uses. She asks one question: which of our problems does this solve better than what we have? Nobody can say, so it waits. Her program has a handful of named problems, each with an owner, and a longer list she has chosen not to fund.

At half past nine the maintenance-planning pilot, which worked at one mine, needs money and data access to go further. As its sponsor she decides that morning rather than at next quarter’s committee. The decision is easy because the safety and security leads joined the team in its first week. Rules for site data, and a human sign-off on every job that touches safety-critical equipment, were designed in from the start.

At eleven she joins a call with the planning desk. When a haul truck’s sensors flag a problem, the assistant pulls the alert, the truck’s repair history and the equipment manual, and drafts a work order with the likely job and the parts it needs. A maintenance planner checks it, and anything safety-critical goes to a senior reliability engineer. The team lead now coaches the quality of the plans rather than the number of work orders raised. The work has changed, not just the screen.

At two o’clock comes the pilot review. The planning assistant beat the baseline measured before it started: work orders go out faster and need fewer corrections. It scales to the next mine with the same controls and measures. A second pilot, which summarizes shift reports, shows plenty of use and no change in any measure. She stops it, asks for a one-page lesson and moves its engineers to the planning rollout.

At four she meets procurement, and the picture changes: many disconnected experiments, no owner and a dashboard of licenses issued. She does not scold anyone. She asks the head of procurement to choose one problem and own it.

At half past six she writes three lines to the chief executive. We lead in maintenance planning. We follow in procurement. Next quarter, procurement gets one sponsored problem with a baseline. Nothing she did that day required a better model.

What this means for leaders

The leaderboard will keep shifting, but the best models stay close, so the source of advantage keeps moving away from access and toward behavior. That gives executives four responsibilities. Make the choice explicit, with a short list of named problems and a longer list of what you will not do. Give each problem a senior owner who can fund it and stop it, with controls designed in from the first experiment. Insist that the work and the people change, not only the tools. And demand a decision from every pilot review: scale, stop, or a dated reason to continue.

Underneath all four is a fifth: tell the truth about where you stand. The most useful sentence in the operations chief’s day was “we follow in procurement”. An organization that can say that has already started to lead.

Check yourself

  1. Because the best models are now close in quality, the companies that pick the best model will lead.
  2. In BCG’s 2024 survey, AI leaders pursued more opportunities than other companies.
  3. US multinationals in Europe got more productivity from the same IT than other multinationals.
  4. AI leaders move fastest because they take the biggest risks.
  5. One function can lead on AI while the wider company follows.
  6. If a company is behind on AI today, it cannot become a leader.

Reflection: where does your loop break?

What comes next

Behavior explains why companies with the same models pull apart. It leaves a harder question for strategy. If a competitor can use the same model tomorrow, and can copy a good process in time, what can it not copy? The next chapter, The AI Competitive Advantage, looks at where lasting advantage can still come from.

Laws referenced

Not legal advice. Laws change; verify before relying on this, and consult counsel for decisions.

EU AI Act · EU

Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744

Risk-based rules. Prohibited practices include social scoring, untargeted scraping of facial images, and emotion recognition in workplaces and schools (with narrow exceptions). High-risk systems (Annex III: biometrics, safety components of critical infrastructure such as energy, water and traffic, employment and worker management, credit, education, essential services, law enforcement, migration, justice) need risk management, data governance, documentation, logging, human oversight, human oversight that keeps people able to understand the system, notice automation bias (over-reliance on its output), override it or stop it (Art. 14(4)), appropriate accuracy, robustness and cybersecurity (Art. 15), automatic logging of events (Art. 12), a provider quality-management system (Art. 17) and conformity assessment. An Annex III system is not high-risk if it poses no significant risk of harm, for example a narrow procedural or preparatory task that does not replace human assessment; systems that profile people are always high-risk, and a provider relying on this exception must document it and register (Art. 6(3)). Deployers of high-risk AI must use it as instructed, assign competent human oversight, monitor its operation, keep logs for at least six months and report serious incidents (Art. 26); employers must inform workers' representatives (Art. 26(7)). Public bodies, private providers of public services, and deployers of credit-scoring or life and health insurance pricing systems must carry out a fundamental-rights impact assessment before first use (Art. 27). Providers must run post-market monitoring (Art. 72). A deployer that puts its name on a high-risk system, substantially modifies it, or changes its purpose so that it becomes high-risk takes on the provider's obligations (Art. 25(1)). A substantial modification (Art. 3(23)) of a high-risk system needs a new conformity assessment, unless the change was pre-determined and documented at the first assessment, as with planned continuous learning (Art. 43(4)). Providers of general-purpose AI models (from 2 Aug 2025) must keep technical documentation, have a policy to comply with EU copyright law including text-and-data-mining opt-outs, and publish a sufficiently detailed summary of training content (Art. 53). Research, testing and development before a system is placed on the market or put into service is outside the Act, except testing in real-world conditions (Art. 2(8)). Since the 2026 Omnibus, the Art. 4 AI-literacy duty is an obligation of effort (take measures to support literacy), not of result. Fines reach EUR 35 million or 7% of global turnover for prohibited practices.

  • 2024-08-01 — Entered into force
  • 2025-02-02 — Prohibited practices (Art. 5) and the AI-literacy duty (Art. 4) apply
  • 2026-07-27 — Omnibus softens Art. 4: providers and deployers must take measures to support AI literacy; no specific level must be guaranteed
  • 2025-08-02 — General-purpose AI model obligations apply; governance and penalties regime in place
  • 2026-08-02 — Transparency duties (Art. 50) apply: disclose AI interaction, label synthetic and deepfake content (marking for generative systems already on the market: 2 Dec 2026)
  • 2027-12-02 — High-risk obligations for Annex III systems (e.g. hiring, credit, education, essential services) - moved from 2 Aug 2026 by the 2026 Omnibus
  • 2028-08-02 — High-risk obligations for AI in products regulated under Annex I

Last verified 2026-10-06 · official text

References

  1. Stanford Institute for Human-Centered AI (HAI). AI Index Report 2026, Chapter 2: Technical Performance. Stanford University. 2026.
  2. Nicholas Bloom, Raffaella Sadun and John Van Reenen. Americans Do IT Better: US Multinationals and the Productivity Miracle. American Economic Review 102(1), 167-201. 2012.
  3. Erik Brynjolfsson and Lorin M. Hitt. Beyond Computation: Information Technology, Organizational Transformation and Business Performance. Journal of Economic Perspectives 14(4), 23-48. 2000.
  4. Sam Ransbotham, Shervin Khodabandeh, David Kiron, François Candelon, Michael Chu and Burt LaFountain. Expanding AI's Impact With Organizational Learning. MIT Sloan Management Review and Boston Consulting Group. 2020.
  5. Boston Consulting Group. Where's the Value in AI?. Boston Consulting Group. 2024.
  6. Richard Rumelt. Good Strategy/Bad Strategy: The Difference and Why It Matters. Crown Business. 2011.
  7. McKinsey & Company (QuantumBlack). The state of AI in 2026: On the road to ROI. McKinsey & Company. 2026.
  8. Gartner. Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025. Gartner press release, 29 July 2024. 2024.
  9. Nicholas Bloom, Erik Brynjolfsson, Lucia Foster, Ron Jarmin, Megha Patnaik, Itay Saporta-Eksten and John Van Reenen. What Drives Differences in Management Practices?. American Economic Review 109(5), 1648-1683. 2019.

Further reading

Sources last verified 2026-10-09.