Executives & Directors · Module 01 · Chapter 014

Where Should We Start With AI?

"Where can we use AI?" has the same answer in every company: almost everywhere. The question that decides results is where to start, because a first project teaches the whole organization what AI is for. Start where the work matters, the feedback comes in weeks and the mistakes can be seen and undone.

≈ 16 min read

After this chapter you can

Explain why "where do we start?", not the list of ideas, is the executive decision, and why capacity rather than ideas is the constraint.
Recognize four common wrong starts - easiest first, everything at once, biggest first and copy the headline.
Explain why a first project needs fast feedback and recoverable mistakes, with applicant screening as the example of what not to start with.
Judge a first win by the value it delivers and by what it teaches the organization.

Suppose you are leaving a long board meeting when the chair stops you at the door. Everyone is talking about AI, she says, so where are we starting? In your bag is the output of last month’s leadership workshop: thirty-eight ideas, from every function, each with an enthusiastic owner. Your budget, your best people and your own attention can carry two of them well. Perhaps three. You have about thirty seconds.

Many leaders answer with whatever is newest, loudest or easiest. That is understandable, and it is how a great many AI programs begin in the wrong place. Researchers at RAND interviewed 65 experienced data scientists and engineers about why AI projects fail. They cite estimates that more than 80 percent of AI projects fail, about twice the rate of IT projects without AI. Of the five root causes they found, four concern what a project chooses to work on and how well that choice is understood: a misunderstood problem, missing data, a focus on the technology rather than the users’ problem, and problems too difficult for AI to solve¹. The choice in the doorway matters more than it looks.

Where you start decides what you learn

The possibility question, “where can we use AI?”, is useful for an afternoon. In principle, language models and prediction can touch almost every process that involves documents, decisions or data, so the honest answer is “almost everywhere”. That answer produces a list. It does not produce a direction.

The leadership question is narrower: where do we start? It matters because a first project does two jobs at once. It delivers something, and it teaches the whole organization what AI is for. People watch the first project closely, and they draw conclusions from it that last far longer than the project.

Figure 1.14.1 A first project is a lesson as well as a delivery. The second lesson is the slowest to unlearn.

If the first project is trivial, people learn that AI is a toy. If it collapses, they learn that AI does not work here, and that lesson spreads quickly and fades slowly. If it works on something that matters, they learn how to do it again. Choosing the start is therefore not a technology decision. It is a decision about what the organization will believe about AI for the next several years, and that makes it a leadership decision.

This chapter stays at why the choice matters. The methods for finding and shaping candidates are taught in AI Use-Case Discovery and Design (Module 05), and the methods for ranking and balancing them in Prioritizing the AI Portfolio (Module 09).

Ideas are cheap; capacity is not

Why not start ten things and keep whatever works? Because the idea list shows only ideas. It hides what every start spends.

Figure 1.14.2 Ideas are the cheapest input to an AI program. The scarce inputs sit below the waterline.

Each start consumes a senior sponsor’s attention, the few specialists who understand your data, the weeks it takes to get that data out of the systems that hold it, managers’ time to change how a team works, and employees’ willingness to try something new. You can collect thirty-eight ideas in an afternoon. You cannot find thirty-eight sponsors.

The scarcest input is patience. A first project that drags on for a year without a visible result spends the patience the second project needed. Richard Rumelt’s study of strategy makes the general point: the core of strategy is choosing where to concentrate limited effort, and a long list of things to do is not a strategy at all². Strategy begins when you decide what not to do yet. How AI costs add up is a separate subject, taught in Understanding AI Total Cost of Ownership in Module 08; here the constraint is people, attention and time.

Four wrong starts

When organizations start in the wrong place, they often do it in one of four ways. Each looks reasonable in the meeting where it is chosen.

Figure 1.14.3 Four ways to start in the wrong place. Naming them makes them harder to choose by accident.

Easiest first. Meeting notes for everyone, or a general assistant on the intranet. These are fine as experiments. The trap is when the easy project uses up the only sponsorship available, no customer notices any change, and leadership concludes that it has “tried AI”. The wider version of this trap, where quick wins crowd out everything else, is the subject of Quick Wins, Strategic Bets and Transformation Initiatives in Module 09.

Everything at once. A pilot in every function, so that nobody feels left behind. The result is dozens of pilots, few real owners, duplicated effort and nothing strong enough to scale. Gartner predicted in 2024 that at least 30 percent of generative AI projects would be abandoned after the proof of concept by the end of 2025, citing poor data, weak risk controls, rising costs and unclear business value³. It was a forecast, not a measurement, but every one of those causes grows when attention is spread thin.

Biggest first. The most ambitious idea on the list, often chosen because it makes the best announcement. It needs data, skills and habits the organization has not built yet. This is RAND’s “problem too difficult” and “missing data” in executive form¹.

Copy the headline. A competitor announces something, and you follow, without its data, its people or its starting point. This is RAND’s “technology over the problem”: the project begins with someone else’s solution rather than your own need.

Start where feedback is fast

There is a quieter test that catches many wrong starts: how soon will you know whether it is working?

Daniel Kahneman and Gary Klein, who spent careers on opposite sides of the debate about expert judgment, agreed on the conditions under which real skill develops. The environment must be regular enough to learn, and people must have an adequate opportunity to learn it, which depends on how quickly and clearly they find out whether they were right⁴. Their subject was individual experts, but the logic carries over to organizations learning a new technology. An organization learns how to use AI from feedback, and it cannot learn from feedback that arrives in four years.

Figure 1.14.4 A first project should tell you within months whether it works. Some valuable projects cannot, and they belong later.

Some of the most valuable uses of AI have slow feedback by nature: forecasts of outcomes years away, research that is confirmed only by later experiments, models of long customer lifetimes. They are not bad ideas. They are poor first ideas, because for most of their life nobody can tell a working project from a failing one. That ambiguity is where patience runs out and where the conclusion “AI does not work here” takes hold.

A first project should produce a result someone can check within months, and ideally a signal within weeks: drafts accepted or rewritten, time saved on a defined task, a backlog that shrinks.

Start where mistakes can be seen and undone

Not all AI mistakes are equal, and the difference matters most at the start, when the organization has not yet learned how to watch AI at work.

Figure 1.14.5 Acceptable risk is not zero risk. It is risk you can see, undo and own.

Some mistakes are recoverable. A meeting summary gets a name wrong and someone corrects it. A first draft has a weak paragraph and the specialist checking it rewrites it. The error is visible, cheap and fixed. Other mistakes are hard to reverse. An automated message reaches a customer before anyone can stop it. A credit or clinical decision is acted on before anyone checks it. A job applicant is rejected automatically and never learns why, so nobody inside the organization sees the error either.

A first project belongs on the recoverable side. That is not timidity. A person who checks every output before it counts is how the organization learns what AI gets right and wrong in its own work, which is the knowledge every later project needs. Making large, irreversible decisions under uncertainty is treated more fully in Preparing for AI Uncertainty and Strategic Change, in Module 10.

What not to start with: screening job applicants

The clearest example of a hard-to-reverse first project appears on many first-project lists: AI that screens job applicants. It looks attractive. It promises speed on a high-volume task, and a person still makes the final call.

It is a poor first project for two reasons. The first is invisibility. A candidate who is wrongly ranked low rarely comes back to say so, so the errors never reach anyone who could fix them. The second is the law. Under Annex III of the EU AI Act, AI systems intended for the recruitment or selection of people, including those that filter applications and evaluate candidates, are high-risk⁵. In New York City, Local Law 144 has regulated automated hiring tools since July 2023⁶.

Figure 1.14.6 As of October 2026, AI that screens applicants is regulated in the EU and in New York City. Confirm with counsel.

The lawHiring AI is regulated (as of October 2026)

Under the EU AI Act (Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744), AI used for recruitment and selection is high-risk under Annex III. High-risk systems need risk management, data governance, documentation, logging, human oversight that keeps people able to understand, override or stop the system, and appropriate accuracy and robustness. Deployers must use the system as instructed, assign competent human oversight, keep logs for at least six months and inform workers’ representatives. The 2026 Digital Omnibus moved these obligations for Annex III systems to 2 December 2027⁷. In New York City, Local Law 144 requires an automated tool that substantially assists hiring or promotion decisions to have an independent bias audit within the past year, a published summary of results, and notice to candidates at least ten business days before use; penalties run from 500 to 1,500 US dollars per violation. A final human sign-off does not by itself take a screening tool outside these rules. This is orientation, not legal advice. Confirm with counsel.

None of this means AI has no place in HR. It means applicant screening is a poor place for an organization to learn how to use AI. A recoverable HR start looks different: drafting answers to employees’ policy questions for an HR specialist to check before they are sent, for example. How bias arises and is tested is the subject of Bias and Fairness, in Module 06.

Judge a first win by value and learning

What does a good first move look like? Judge it on two things at once: the value it delivers and what it teaches.

Figure 1.14.7 A first win matters, can be finished, teaches the organization and earns the trust that funds the next project.

It matters. It works on a problem a leader already cares about, described in words a customer, a colleague or the finance team would recognize. It can be finished. It is bounded to one site, one team or one kind of document, it has a named owner, and it has a clear way to stop. It teaches. On the way, the organization learns where its data really lives, how to check AI’s work and how people take up a new routine. It earns trust. An honest, visible result buys the credibility a bigger project will need.

This idea is older than AI. The organizational psychologist Karl Weick argued that problems framed at too large a scale overwhelm the people facing them and block action, and that progress comes from small wins, which he defined as concrete, complete, implemented outcomes of moderate importance⁸. John Kotter’s study of corporate change made generating short-term wins one of the eight stages he saw in successful transformations, because visible results disarm cynics and keep sponsors committed⁹.

That is why a smaller first win can be the right choice and a giant one the wrong one. The small win is not the destination. It is how the organization earns the skills and the standing to attempt the big one.

Story: the driverless bet that went first

In August 2016 Ford announced that it intended to have a high-volume, fully autonomous vehicle in commercial operation in 2021, in a ride-hailing or ride-sharing service. The vehicle would have no steering wheel and no accelerator or brake pedals. Its chief executive, Mark Fields, said autonomous vehicles could have as significant an impact on society as Ford’s moving assembly line had a century earlier¹⁰, the redesign The Four Industrial Revolutions described. Six months later Ford said it would invest 1 billion dollars over five years in Argo AI, a new company founded by former Google and Uber leaders, to build the virtual driver for that vehicle¹¹.

Figure 1.14.8 Ford’s biggest AI bet went first, and its decisive feedback could arrive only after years.

In October 2022 the bet was written down. Ford recorded a 2.7 billion dollar non-cash, pretax impairment on its investment in Argo AI and a net loss of 827 million dollars for the quarter. It said it had concluded that large-scale, profitable commercialization of Level 4 systems, the kind its 2016 vehicle needed, would be further out than originally anticipated, and it shifted its spending to driver-assistance technology it developed itself¹². Argo AI, which Volkswagen also backed, shut down, and its two backers absorbed parts of its operations¹³. Jim Farley, by then Ford’s chief executive, said profitable, fully autonomous vehicles at scale were “a long way off”¹².

Ford did not describe its decision in this chapter’s terms; its stated reason was timing. Read against this chapter, though, the 2016 bet had the marks of a hard start. It was the biggest idea available: a vehicle with no controls, in public traffic. Its decisive feedback, whether a driverless service could run profitably at scale, could arrive only after years. Its mistakes, on public roads, could not easily be undone. And it carried a public date. None of that makes it foolish. It makes it a bet that needed more patience than most programs have.

Figure 1.14.9 What Ford turned to could be shipped, used and measured every week.

What Ford turned to looks different. Its BlueCruise system allows hands-free driving on pre-mapped highways, while an in-cabin camera checks that the driver’s eyes stay on the road¹⁴. By October 2022 more than 83,000 owners had enrolled in it or its Lincoln equivalent, and they had logged more than 21 million hands-free miles in just over a year¹². That is a feedback loop measured in weeks: a product people use, a person supervising at every moment and data from every trip.

The contrast does not mean the smaller product is free of risk. Driving is a domain where some mistakes cannot be undone. In April 2024 the US National Highway Traffic Safety Administration opened an investigation into BlueCruise after two fatal crashes, and in January 2025 it upgraded the investigation to an engineering analysis covering about 129,000 vehicles [@techcrunch-bluecruise-probe-2025; @reuters-bluecruise-probe-2025]. That is one reason this chapter advises most organizations to start much further from that edge. The lesson of the case is order. The biggest idea went first, its proof was years away, and when the company changed course it moved to work whose results it could see every week. The ambition was not dropped: Ford said it remained optimistic about Level 4, although it might not build that technology itself¹².

What this means for leaders

The decision about where to start belongs to the executives who own the outcomes, not to the most enthusiastic team or the most persuasive vendor. Technologists must tell you what is feasible; which outcome, which risk and which owner are leadership choices. Make the choice deliberately, explain it, and keep a visible list of the ideas that are waiting rather than rejected, so that the people behind them stay engaged.

Then protect the first win. Give it a named owner, a bounded scope, a person who checks the output, and a date on which you will decide to expand it or stop it. Report what it taught as well as what it delivered. And resist the pressure to announce it as the start of a transformation before it has produced a result.

Check yourself

Start with the easiest project, because any win builds momentum.
A project whose results can only be proven years later is a poor first AI project, even if it is valuable.
Screening job applicants is a safe first project as long as a person makes the final decision.
Giving every department its own pilot is the fastest way to find what works.
A smaller first win can be a better choice than a bigger one.
Choosing where to start is mainly a technology decision.

Reflection: test your own start

What comes next

Choosing where to start is half of the decision. A few months later the board will ask for the other half: how do we know it is working? Answering that honestly is harder than it sounds, because usage is easy to count and value is not. The next chapter, Measuring AI Business Value, takes up that question.

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

EU Digital Omnibus on AI · EU

Regulation (EU) 2026/1744

First amendment to the AI Act. Defers high-risk obligations (Annex III to 2 Dec 2027, Annex I to 2 Aug 2028), adds two prohibited categories, softens the Art. 4 AI-literacy duty to "take measures to support", and simplifies some compliance duties. Art. 50 transparency duties still apply from 2 Aug 2026, with one transition (new Art. 111(4)): providers of generative AI systems placed on the market before 2 Aug 2026 must meet the Art. 50(2) marking duty by 2 Dec 2026.

2026-07-24 — Published in the Official Journal
2026-07-27 — Entered into force
2026-12-02 — Grace period ends for safeguards against two new prohibited uses (non-consensual intimate imagery, child sexual abuse material)
2026-12-02 — Art. 50(2) marking duty applies to generative AI systems placed on the market before 2 Aug 2026 (Art. 111(4))

Last verified 2026-10-10 · official text

NYC Local Law 144 (automated employment decision tools) · US - New York City

NYC Local Law 144 of 2021; DCWP rules

An automated tool that substantially assists hiring or promotion decisions needs an independent bias audit within the past year, a published summary of results, and notice to candidates at least ten business days before use. Penalties USD 500 to 1,500 per violation.

2023-07-05 — Enforcement began