Most AI pilots do not fail. That is the problem. They finish, they produce a slide, somebody says "interesting", and nothing changes. Six months later the same organisation is running a different pilot with a different tool, and the leadership team is no closer to a decision than it was before the first one started.
I hear this story more than any other. Every week, replies to my newsletter arrive from leaders in schools, trusts and companies describing an AI pilot that went well and then quietly went nowhere. The pattern is too consistent to be bad luck. It is a set of mistakes, and most of them are made before the pilot begins.
Here are the seven I see most often: why each one happens, what it does to you, and what to do instead.
The pilot was chosen because it was safe, not because it was decisive
An AI pilot chosen because it cannot fail is a pilot that cannot teach you anything.
A leadership team that is nervous about AI picks the use case with the least at stake: an admin task in the back office, an internal question-and-answer bot, a tool that summarises meeting notes. Nobody objects, nobody is affected, and nobody much cares whether it works. The pilot succeeds in the narrow sense and answers no question the leadership team actually has.
Why it happens: the choice is made in the risk conversation rather than the strategy conversation, and safe pilots are easy to approve.
What it does to you: you learn that the tool functions, which the vendor could have told you, and nothing about whether your organisation can change around it.
What to do instead: choose a pilot which, if it works, would force a decision somebody would rather not make. If the honest answer to "what would we change if this works?" is "not much", pick a different pilot.
Nobody decided what "it worked" would mean before it started
A pilot without a written definition of success ends with the word "interesting", and "interesting" is where pilots go to die.
The evidence on this is uncomfortable. MIT's Project NANDA published its GenAI Divide report in August 2025, drawing on 52 interviews, 153 senior leader surveys and a review of more than 300 public AI initiatives. More than 80 per cent of the organisations studied had explored or piloted generative AI. Only 5 per cent of custom enterprise tools ever reached production. The researchers did not blame the models. They blamed tools that never fitted the way work was actually done, in organisations that had never said what fitting would look like.
Why it happens: defining success in advance feels like prejudging the result. It also exposes something awkward, which is that the leadership team has not agreed what AI is for here. That is the first question of the Readiness Test, and a pilot cannot answer it on the team's behalf.
What it does to you: at the end, every participant holds a different view of whether it worked, and the loudest view wins. Usually the loudest view belongs to whoever ran it.
What to do instead: ask what I have come to call the Exit Question. The Exit Question is the one question I put to a leadership team before any AI pilot starts: if this works, what will change, who has already agreed to it, and on what date will you decide? A pilot without an answer is not a pilot. It is an experiment with no exit.
The pilot ran with volunteers and the rollout was planned for everyone
A pilot staffed by enthusiasts measures enthusiasm, not adoption.
The Tony Blair Institute's Generation Ready report (January 2026) found that 91 per cent of teachers using AI in England are entirely self-taught, and that only 8 per cent of school leaders had made any organisational change in response to AI, with a third having no plans to. Read those two figures together and you have the volunteer problem. The people who put their hands up were already using the tools at home. The pilot confirms what the self-taught can do, and then the rollout meets everyone else.
Why it happens: volunteers say yes, sceptics say no, and a leadership team under pressure takes the path of least resistance.
What it does to you: results that vanish on contact with the wider organisation, and sceptics who feel vindicated because nobody asked them. As I argued in why AI adoption is really a leadership problem, the middle layer decides whether AI is expected rather than merely allowed, and the middle layer is exactly who the volunteer pilot leaves out.
What to do instead: recruit at least one open sceptic and one middle leader whose team would have to change if the pilot worked. Measure what the reluctant do with the tool, not what the keen do. The keen will be fine either way.
It tested the tool and not the work around it
The tool is almost never the thing that fails. The work around the tool is.
The cleanest example I know comes from education. The Education Endowment Foundation's trial of ChatGPT in lesson preparation, evaluated by NFER and published in December 2024, involved 259 teachers across 68 English secondary schools. Teachers given ChatGPT and a short guide cut their weekly planning time by 31 per cent, about 25 minutes a week, with no measurable drop in the quality of what they produced. As pilot results go, that is a good one.
Now ask the leadership question it leaves behind: what happened to the 25 minutes? If nobody decided, the time dissolved back into the same working week. A time saving that no one redirects is not an outcome. It is a rounding error.
Why it happens: it is far easier to measure whether a tool works than to redesign the process it sits inside, so leaders measure the tool.
What it does to you: a pilot that proves the tool and changes nothing. The MIT researchers saw the same from the other direction: the organisations that got value changed a workflow and put the tool inside it.
What to do instead: pilot the process change, with the tool as one part of it. Which step disappears? Who signs off now? What does the person do with the time they get back? If those three questions have no answers, you are testing software, not your organisation.
Success had no owner and failure had no date
A pilot with no end date is not a pilot. It is a habit.
Gartner predicted in June 2025 that more than 40 per cent of agentic AI projects will be cancelled by the end of 2027, citing unclear business value and inadequate risk controls, and describing most current projects as "early stage experiments or proofs of concept that are mostly driven by hype". I would put it more plainly: pilots drift because ending one is a decision, decisions have owners, and nobody volunteered.
Why it happens: nobody wants to be the person who "stopped AI", so the pilot is kept at a low simmer, with two or three people still using it, and it never has to be judged.
What it does to you: pilot purgatory. Across the organisations I work with it is common to find three or four pilots running at once, none with an end date, each fragmenting the attention that any one of them would need to earn a decision.
What to do instead: name one person, not a team, who owns the outcome. Name the date. On that date there are three permitted results: scale it, stop it, or extend it once with the reason written down. Extend it twice and you have stopped without admitting it.
The pilot was kept quiet
A pilot that nobody outside the pilot knows about cannot be scaled, because scaling is a story before it is a system.
Why it happens: fear of scrutiny. Governors, trustees, unions, parents, customers and the board all have views about AI, and it feels safer to run the pilot under the radar and present a finished result.
What it does to you: there is no story to scale, because nobody was brought along, and the governance never gets built. Grant Thornton's 2026 AI Impact Survey of 950 senior leaders, conducted in February and March 2026, found that only 7 per cent of organisations still at the piloting stage were confident they could pass an independent AI governance audit within 90 days, against 74 per cent of those with AI fully integrated. Governance is built by the act of scaling in the open. The quiet pilot never builds it, which is one reason the quiet pilot never scales.
What to do instead: tell the people affected what you are testing, why, and when you will decide. That is one of the seven behaviours in what responsible AI adoption looks like in practice, and it is the one that pilots most often skip.
The leadership team treated the pilot as the decision
Running a pilot is not the same as deciding about AI. Quite often it is a way of postponing the decision while appearing to make one.
Deloitte's State of AI in the Enterprise 2026 report, which surveyed 3,235 senior leaders across 24 countries, found that only around a third described themselves as deeply transforming how they work, while 37 per cent admitted they were using AI at the surface. Pilots live in that gap. They generate activity without commitment, and activity is easy to mistake for progress.
Why it happens: a pilot lets a leadership team say "we are doing something about AI" without answering what AI is for here, who may use it, and who decides when it goes wrong.
What it does to you: a year of pilots and no position. When somebody finally asks the board-level question, the honest answer is "we have tried a few things".
What to do instead: a pilot should answer a question the strategy has already asked. If there is no strategy, then the pilot is the strategy, and you will find that out when it ends.
The Mistake I See Most Often
Underneath all seven is one mistake: leaders run pilots to find out whether AI works, when the question they need answered is whether their organisation can change.
Whether the tools work has been answered, publicly and repeatedly. Whether your organisation can redesign a process, redirect the time, bring the sceptics with it and decide on a date is a question only your organisation can answer, and it is the only question worth a pilot.
When I sit down with a leadership team for a strategy session, I usually start by asking about the last pilot. Tell me what happened. Then: what changed because of it? The silence that follows is not embarrassment. It is the moment the team realises the pilot was never designed to change anything. The list we build next, of what would have had to be true for it to change something, is where the real strategy work starts.
I have sat on the other side of this too. As Director of Digital Strategy in a further education college I was the one running the pilots, and I recognise several of these mistakes from the inside. The safe use case, the volunteers, the missing end date: each felt like prudence at the time. Prudence is what a stalled pilot looks like from within.
What a pilot that goes somewhere looks like
A pilot that goes somewhere is designed backwards from the decision it exists to force.
The use case was chosen because a positive result would oblige somebody to change something. Success was written down before the first login, and the Exit Question had an answer. A sceptic and a middle leader were inside it. The process was redesigned with the tool inside it. One person owned the outcome and there was a date in the diary. The people affected were told at the start, not shown a result at the end. And on the date, the leadership team met and chose one of three outcomes, in the open.
None of that requires a better tool. All of it requires a leadership team willing to decide.
What to do next
Start with the pilots you already have, not the one you are planning.
List every AI pilot currently running in your organisation and put the Exit Question to each one: if this works, what will change, who has agreed, and when will you decide? Any pilot without an answer gets one of two treatments this month: an owner, a date and a definition of success, or a clean stop that everyone hears about. Then, and only then, design the next pilot to force a decision you would rather not make, and put a sceptic inside it.
If your leadership team is working through pilots that went well and went nowhere, this is the kind of work I support through AI strategy sessions and implementation advisory: turning experiments into decisions, and decisions into a rollout the whole organisation can follow. The place to start is here.
Dan Fitzpatrick is the founder of The AI Educator and works with leadership teams in schools and organisations on AI strategy, adoption and change. More about Dan.
