Part of our guide: AI strategy
Somebody in your business has put a number in a board paper this quarter, and it says that ninety-five per cent of AI pilots fail. Other numbers in circulation say that forty per cent of agentic projects will be cancelled and that data preparation is most of the cost. Each one of them arrives with an institution's name attached, and it is the name that is doing the work.
We went and read the primary sources, meaning the actual report rather than the press coverage or the LinkedIn post. Four of the six most-quoted numbers in our client conversations do not survive that trip, and two of them are wrong in ways that would change a decision.
This is not an argument that AI is working better than people think, and some of the sceptical numbers are sound, so we cite them below. It is an argument that a business making capital allocation decisions on borrowed statistics is exposed, and that checking is cheap.
"MIT found 95 per cent of AI pilots fail"
This is the most repeated enterprise AI statistic of the past year, and it comes from The GenAI Divide: State of AI in Business 2025, published by MIT Project NANDA in July 2025.
The report is real and the figure is in it, and four things about that figure are not widely known.
It was labelled preliminary findings and was never peer reviewed, drawing its evidence from over 300 publicly disclosed AI initiatives. It also drew on structured interviews with representatives of 52 organisations and survey responses from 153 senior leaders gathered at four industry conferences. The report says its own figures are "directionally accurate based on individual interviews rather than official company reporting".
The 5 per cent sits in a funnel chart measuring one narrow category, that of task-specific AI tools reaching production. The same chart shows general-purpose tools such as ChatGPT and Copilot reaching implementation around 83 per cent of the time. A reader who takes 95 per cent as the failure rate for all enterprise AI has taken a number about one category and applied it to everything.
Success was defined as marked and sustained productivity or profit impact, so a deployment producing a modest, real return was counted as a failure.
The number has also grown in transit, and Fortune described the methodology as 150 interviews and a survey of 350 employees, roughly three times the actual sample. The most-cited AI statistic of the year has been circulating with a sample description substantially larger than the one in the report.
Kevin Werbach, who directs Wharton's Accountable AI Lab, could not locate the basis for the figure and wrote that MIT should release the supporting data or retract the report. The original MIT-hosted link now redirects to a NANDA overview page listing no publications, and that change is fairly described as quietly de-linked. It should not be confused with a retraction, and nor should it be confused with MIT's separate disavowal of an unrelated paper in 2025.
What to say instead. If you want a sourced statement that most AI pilots do not pay, use McKinsey's November 2025 survey of 1,993 respondents across 105 countries. Around a third have begun scaling, 7 per cent report full scaling and 39 per cent attribute any enterprise-level EBIT impact to AI. Most of them cite under 5 per cent of EBIT, and that is a properly sampled survey supporting the same conclusion without the fragility.
Wharton's own 2025 adoption report is worth holding alongside it, and that report, surveying leaders at large firms, found three in four reporting positive returns. Both cannot be describing the same thing and they are not, and the real lesson is that they are asking different questions and defining success differently.
"Gartner says 40 per cent of agentic AI projects will be cancelled"
The quote is accurate, and Gartner does say that over 40 per cent of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls.
Two qualifications sit behind that, and the first is that it is a prediction published on 25 June 2025, not a measurement and not 2026 research. Press coverage regularly presents it as current findings, and the second qualification is that the underlying data is a poll of 3,412 webinar attendees, a self-selecting audience rather than a sampled survey.
Gartner's more useful 2026 finding is rarely quoted: only 17 per cent of organisations have deployed AI agents, while over 60 per cent expect to within two years. Gartner places agentic AI at the Peak of Inflated Expectations and states that fully autonomous agents are not ready for the majority of enterprise use cases.
That gap between 17 and 60 is the number worth putting in a board paper, and it describes the decision your business is actually facing.
"Data preparation is 60 to 80 per cent of the cost"
We have used a version of this ourselves, and we were wrong to do so.
Every instance we could trace leads back to a 2014 New York Times article based on interviews and expert estimates, describing the share of data scientists' time spent on data preparation. It is not a cost share, it was never a study and it predates the technology it is now used to explain.
There is no current, sourced figure for data preparation as a proportion of AI project cost, and we looked for one. Every citation resolves to marketing content, each quoting a different range and none of them carrying a sample or a method.
What is actually measured. Dun and Bradstreet's July 2026 survey of 10,000 businesses found only 6 per cent say their data is fully ready to support AI at scale, against 34 per cent claiming to be scaling AI into production. That is the real finding, and it is better than the unsourced one because it makes the same point from measurement: the constraint is data, and most organisations are further from ready than their programme plan assumes.
The same trap catches BCG's 10-20-70 rule, often quoted as 70 per cent of the budget going to people and process. BCG said 70 per cent of the effort, not the cost, and the article is from February 2022, about classical machine learning.
"Seventy-five per cent of UK financial services firms use AI"
That was true in November 2024, when the Bank of England and FCA published it, and it is still being quoted in 2026 as current.
The fourth edition of that survey closed on 31 July 2026 with results expected at the end of the year. Until then, anyone citing the 75 per cent is citing a two-year-old measurement of the fastest-moving thing in their business.
The current UK number comes from the ONS, published 20 July 2026, and it is a better statistic anyway, so we wrote about what it shows in our look at the UK adoption data.
"EU AI Act compliance costs 400,000 euros per high-risk system"
This one is not merely stale, and the authors of the underlying CEPS study have said the figure was produced incorrectly. They say their setup and annual maintenance numbers were added together, and that the figure cites page numbers that do not exist in the Commission's impact assessment.
The study's actual figures are roughly 193,000 to 330,000 euros of setup cost plus about 71,400 euros a year, applying to providers of high-risk systems that do not already have a quality management system. Most UK mid-market firms are deployers rather than providers, and neither number is the right one for them to plan against.
The high-risk regime has also moved, and that is worth knowing given how often the AI Act is used to create urgency. The Digital Omnibus entered into force on 27 July 2026 and pushed high-risk obligations under Annex III to 2 December 2027, though transparency obligations did begin on 2 August 2026.
The two numbers that do survive
Both of them are peer reviewed, and peer review is rare in this field.
Brynjolfsson, Li and Raymond published in the Quarterly Journal of Economics in May 2025, studying 5,179 customer support agents. They found that AI assistance raised issues resolved per hour by 14 per cent on average, with the gains concentrated almost entirely in novice and lower-skilled workers, who improved by 34 per cent. The most experienced staff saw minimal impact.
Dell'Acqua and colleagues published in Organization Science in March 2026, finding that inside the tasks AI is suited to, consultants completed 12.2 per cent more tasks. They did it 25.1 per cent faster and at significantly higher quality, and outside that boundary AI users were 19 per cent less likely to produce a correct answer.
Put together, they say something more useful than any of the headline numbers: AI produces large, real gains within a boundary and it degrades performance outside it. It also compresses the gap between your best and weakest people rather than lifting everyone equally. Every one of those findings has an operating model consequence, and none of them is captured by a single failure rate.
What to do with this
Three habits will do the job, and none of them is expensive.
Ask for the primary source before the number enters a paper. The report is what counts rather than the article, and if nobody can produce it in ten minutes, the number does not go in.
Check the fieldwork date, not the publication date. Deloitte's State of AI in the Enterprise 2026 was fielded in August and September 2025, and Stanford's AI Index 2026 republishes McKinsey's mid-2025 survey for its headline enterprise figures. A 2026 label frequently sits on 2025 data, and in a field that moves this quickly a year is a long time.
Watch for the sample growing in transit. When a statistic passes through three outlets, the method description tends to inflate, and if your source is a summary of a summary, you do not know what you are quoting.
None of this makes the case for AI weaker, and it makes your case for it defensible, and that matters more when you are the person who signed the business case.
We do this kind of work as part of a Breathe discovery sprint, where the first job is usually separating what a business believes about AI from what is actually evidenced. If a number in your board pack is doing a lot of load-bearing, we are happy to look at it with you.