The most expensive AI project is the one that never got a no
The figure that 95 per cent of AI programmes return nothing currently carries half the market's argument. It cannot bear that weight. The conclusion it is used for is still broadly right, for a better reason.
Key insights
- The MIT NANDA figure that 95 per cent of organisations get zero return rests on 153 survey responses collected at four industry conferences, 52 interviews and a review of 300 initiatives.
- The figure concerns organisations getting zero return, not pilots failing. Those are not the same claim, and they are conflated systematically in the way the report is repeated.
- Gartner predicted at least 30 per cent abandoned after proof of concept by end-2025, and measured at least 50 per cent in January 2026, for four named reasons: data quality, risk controls, costs and unclear value.
- Omdia measures something different and lands in the same picture: in a survey of 448 enterprises in the third quarter of 2025, under a third say fewer than 5 per cent of their own PoCs reach production.
- All four reasons can be established before the build starts. That is why the early no is worth most: it costs a meeting instead of a project year.
One figure currently carries half the market's argument about AI. It comes from a report by MIT NANDA and says that 95 per cent of organisations are getting zero return. It is quoted in boardrooms, in sales material and in editorials, often with the addition that almost all pilots therefore fail.
The figure cannot bear that weight, and the reason is visible in the report's own methodology section.
What the report actually says, and what it rests on
The wording in the report is that despite 30 to 40 billion dollars of enterprise investment, 95 per cent of organisations are getting zero return. That is not the same claim as 95 per cent of pilots failing. A company can have successful pilots and still no measurable return overall, and the reverse. The two are conflated systematically in the repetition.
Then there is the evidence base. The report's own methodology section states a review of over 300 publicly disclosed AI initiatives, structured interviews with representatives of 52 organisations, and survey responses from 153 senior leaders collected across four industry conferences. The period is January to June 2025.
The survey responses were collected at four industry conferences. The report states no random sampling.
Three things follow. Survey responses collected at industry conferences are a convenience sample rather than a random one, because those who attend an AI conference differ from those who do not. Return is not defined in the quoted figure. And six months is a short window in which to judge the outcome of investments whose payback is normally counted in years.
A figure too large for its evidence makes nobody more careful. It only means that fewer people ask what their own trials produced.
A market that builds its picture of where things stand on a single figure also becomes sensitive to how that particular figure was produced. If return is never defined, and the responses are collected among those who have already come to an AI conference, the uncertainty is conceptual rather than statistical. It cannot be removed by adding respondents.
Gartner measured the outcome: at least 50 per cent, not 30
The direction is not wrong. It simply rests better on something else.
In July 2024 Gartner predicted that at least 30 per cent of GenAI projects would be abandoned after proof of concept by the end of 2025, due to poor data quality, inadequate risk controls, escalating costs or unclear business value.
That date has now passed, and Gartner has measured the outcome. In January 2026 it states that at least fifty per cent of projects were abandoned after proof of concept by the end of 2025, for the same four reasons. The forecast was therefore too optimistic, and the figure worth using is the measurement rather than the prediction.
Gartner's own evidence base is thin in turn. The forecast was presented as an analyst judgement with no sample disclosed, and the outcome is said to rest on a review of hundreds of implementations with no stated number, field period or sampling method. The difference from the MIT figure lies not in the method but in what the figure carries: a claim about how many attempts are ended is narrower than a conclusion about return across a whole market, and it does not depend on how return is defined.
Fifty per cent is a more modest figure than ninety-five, and it is more useful, because it comes with four named causes rather than a verdict.
Read the list again and notice what the four have in common. Data quality can be checked before the build starts. Risk controls can be described before the build starts. The cost picture can be estimated, and business value can be stated as a measurable claim. All four are therefore known, or knowable, before anyone writes a line of code.
Omdia measured throughput from proof of concept to production
A second way of measuring exists, and it does not produce the same number. The analyst firm Omdia asked 448 enterprises during the third quarter of 2025 what share of their own proofs of concept had moved into production. Chief analyst Eden Zoller reports the answers in AI Market Maturity, 2025 Data: forty-six per cent of enterprises say that over 10 per cent of PoC projects move forward to production, with most of these, 37 per cent, falling in the 11 to 40 per cent range, while 10 per cent of organisations report success rates above 40 per cent. Under a third of respondents say fewer than 5 per cent of PoCs make it to production, and 21 per cent put the figure at 5 to 10 per cent.
The two figures measure different things. Gartner states the share of projects abandoned after proof of concept. Omdia states the share of enterprises with low throughput from proof of concept to production. A project figure and a company figure cannot be compared directly or added together, and they do not describe the same population.
On the evidence base the advantage is Omdia's. The number of respondents is stated, the field period is given as the third quarter of 2025, and the method is a quantitative enterprise survey in which each company reports the outcome in its own portfolio, meaning what has already happened rather than what an analyst expects to happen. Gartner discloses neither a count, a field period nor a sampling method for either of its two figures.
Two different ways of measuring still land in the same picture: what gets built rarely reaches production. The spread in Omdia's answers is the most useful part of the material. A tenth of enterprises get more than four in ten attempts all the way into production, while under a third stay below five per cent. The share is therefore not a constant of nature, and the four reasons Gartner names all sit on the decision side rather than in the technology.
Why the late no becomes expensive
In the same material Gartner states that the deployment approaches come with significant costs, ranging from 5 million to 20 million dollars.
Concerns larger enterprise deployments rather than a bounded trial. The range explains why the timing of a no matters.
The range concerns larger enterprise deployments and should not be read as the price of a bounded trial. But it says something about the curve. The cost of ending a use case rises steeply with time, while the information needed for the decision barely improves after the first few weeks.
That is the only economic observation required to see why the early no is worth most.
Why the no still is not said
Rarely because the information is missing. Almost always because the project has become somebody's own.
A use case that has been given an owner, a budget and a presentation is no longer a proposal. Ending it is then read as a judgement of the person rather than of the idea, and whoever proposes it pays a social cost no calculation captures. The consequence is that projects are not ended but thinned out, which is more expensive and slower than a no.
The remedy is easy to describe and uncomfortable to adopt: set the stopping criteria before you start, and decide who is allowed to apply them.
Three decisions before the start
- What has to be true for the trial to continue? Stated as something measurable, on a bounded step. Not "that it works well".
- When is the check made? A date, set in advance. A check to be made "when we see how it goes" is never made.
- Who may say no without securing agreement first? Without that point the first two are only a wish. The mandate should be named, and said out loud while everyone is still optimistic.
The three decisions are made in the meeting where somebody would otherwise have said that we will get going and evaluate as we go, and they require someone to hold the pen when the time comes to apply them.
Common questions
On the MIT NANDA report The GenAI Divide, State of AI in Business 2025. Its methodology section states a review of over 300 publicly disclosed AI initiatives, structured interviews with representatives of 52 organisations, and survey responses from 153 senior leaders collected across four industry conferences, over January to June 2025. No random sampling is stated.
No. The report's wording is that 95 per cent of organisations are getting zero return, despite 30 to 40 billion dollars of investment. Organisations getting no return and pilots failing are two different claims, and they are conflated systematically in how the report is repeated.
Gartner predicted in July 2024 that at least 30 per cent would be abandoned by the end of 2025. In January 2026 it measured the outcome and states at least 50 per cent, for the same four reasons: poor data quality, inadequate risk controls, escalating costs or unclear business value. Gartner discloses no sample for either figure; the outcome is said to rest on a review of hundreds of implementations.
In the same material Gartner states that the deployment approaches in question come with significant costs, ranging from 5 million to 20 million dollars. The range concerns larger enterprise deployments and should not be read as the price of a bounded trial, but it explains why a late no becomes expensive.
As early as possible, and preferably before the build starts. All four reasons Gartner gives for abandonment can be established in advance: data quality can be checked, risk controls can be described, the cost picture can be estimated and business value can be stated as a measurable claim. What makes the no difficult is rarely a lack of information.
Because a project that already has an owner, a budget and a presentation behind it has become somebody's own. Ending it is then read as a judgement of the person rather than of the use case. Organisations that set the stopping criteria in advance avoid that link, because the decision has already been made by everyone together.
No. The criticism concerns what the figure can carry, not the direction. Gartner's measurement of the outcome points the same way: a large share of what is started never reaches real use. The difference is that the conclusion holds with an honest number too, and that it then leads to a different action than resignation.
Three things, all before the start. What has to be true for the trial to continue, expressed as something measurable. When that check happens, expressed as a date. And who is allowed to say no without securing agreement first. Without the third point the first two are only a wish.
If this lands on your desk, we should talk.
Ampliro Insights
New analysis, roughly weekly.
We write when the rules change and when something turns out to work in practice. One piece at a time, no sequences, and you can leave from any issue.
We store your address to send Ampliro Insights, and for nothing else. More in the privacy policy.