Do 95% of AI pilots fail? What the MIT report says

The short version

In August 2025 a report from MIT's Project NANDA was widely reported as showing that 95% of generative AI pilots fail. The report's executive summary uses wording close to that, but its detailed figures are narrower, and it words the 95% in several different ways.

It words the 95%, and the matching 5%, at least five ways with different bases: 'organizations' getting 'zero return', 'integrated AI pilots', 'custom enterprise AI tools' that reach production, a 'failure rate for enterprise AI solutions', and 'enterprises' with AI tools 'integrated in workflows at scale'.

It rests on 52 interviews, 153 conference survey responses and a review of over 300 public initiatives. It calls its figures 'directionally accurate'. It measured success six months after a pilot, and warns that this may understate success.

For general-purpose tools such as ChatGPT and Copilot, the report says over 80% of organisations have explored or piloted them and nearly 40% report deployment.

None of this means AI pilots always work. In our view, it means the 95% is not a number to plan around.

Where the 95% comes from

The figure comes from 'The GenAI Divide: State of AI in Business 2025', a report from MIT's Project NANDA dated July 2025. On 18 August 2025 Fortune's headline put it this way: 'MIT report: 95% of generative AI pilots at companies are failing'. The word 'pilots' is also in the report's executive summary: 'Just 5% of integrated AI pilots are extracting millions in value'.

The same page says '95% of organizations are getting zero return'. A few pages later the report says 'only 5% of custom enterprise AI tools reach production', 'The 95% failure rate for enterprise AI solutions', and 'Only 5% of enterprises have AI tools integrated in workflows at scale'. Those are different claims with different bases: organisations, pilots, tools and enterprises.

The chart behind the 5% has three steps for task-specific tools: 60% investigated, 20% piloted and 5% successfully implemented. The summary describes the 60% as a share of organisations and does not say what the 20% and 5% are shares of. On the same page it also calls the 5% a share of 'integrated AI pilots'.

If the three steps share one base, 5 of the 20 that piloted is 25%, a fall of about 75% between pilot and implementation, not 95%. That is our arithmetic, not MIT's, and the chart does not state its base.

What it measured, and for how long

The report says it is 'based on a multi-method research design that includes a systematic review of over 300 publicly disclosed AI initiatives, structured interviews with representatives from 52 organizations, and survey responses from 153 senior leaders collected across four major industry conferences'. The research period was January to June 2025.

In its methodology note the report says success means 'deployment beyond pilot phase with measurable KPIs', with 'ROI impact measured 6 months post-pilot'. Elsewhere it defines 'successfully implemented' for task-specific tools as 'ones users or executives have remarked as causing a marked and sustained productivity and/or P&L impact'. It does not say how the two definitions relate. It never defines what a pilot is.

It does not name the four conferences. It does not break its samples down by country, sector or size. It says it calculated confidence intervals 'where applicable' but prints none. It does not state where respondents are based, and it gives no UK-specific figures.

The authors are open about the limits. They call the figures 'directionally accurate based on individual interviews rather than official company reporting', and add that 'success definitions may differ across organizations'. They say the six-month window may be insufficient, 'potentially understating success rates for longer-term implementations'. The report is labelled 'Preliminary Findings'.

What the 95% does not say

It is not a failure rate for AI in general. The 5% in the main chart is for what the report calls 'enterprise-grade systems, custom or vendor-sold'. That describes the type of tool, not the size of the buyer. Separately, the report defines enterprises as 'firms with over $100 million in annual revenue', and its summary says its buyers include 'enterprises, mid-market, SMBs'.

For general-purpose tools the picture is different. The report says: 'Over 80 percent of organizations have explored or piloted them, and nearly 40 percent report deployment.' It adds that these tools 'primarily enhance individual productivity, not P&L performance'.

The report gives two chatbot figures that are easy to confuse. Its chart shows 50% piloted and 40% successfully implemented, which is 80% of those that piloted. Its text says generic chatbots 'appear to show high pilot-to-implementation rates (~83%)', then adds that this 'masks a deeper split in perceived value'. We could not find how the 83% was derived.

Details in early coverage that are not in the report

Fortune described the research as 'based on 150 interviews with leaders, a survey of 350 employees, and an analysis of 300 public AI deployments'. The report says 52 organisations and 153 senior leaders. We searched the report and found no '150' or '350'.

We found no source for those figures. Fortune quotes the lead author, so they may come from his remarks, but its page does not say so. We found no correction or editor's note on either Fortune page as at 10 November 2026.

Fortune's second article, on 21 August, noted that the market had read the report as an indictment of AI as a whole 'even though that's not exactly what the research said'.

The build-or-buy finding

A widely quoted finding is that 'In our sample, external partnerships with learning-capable, customized tools reached deployment ~67% of the time, compared to ~33% for internally built tools'. The authors add that the figures 'reflect self-reported outcomes', that differences may reflect 'organizational capabilities rather than implementation approach alone', and that the correlation 'does not necessarily prove causation'.

The chart for this finding is titled '% of Deployments', which describes a share. The text describes how often each approach 'reached deployment', which is a success rate. The report does not explain how the two relate.

BespokeWorks builds bespoke software. In 'Why SaaS isn't dead' we argue that most businesses should keep most of what they rent and build only what it cannot do. Read our view of this finding with that interest in mind. The report's own caveats are the safer guide.

What the UK evidence measures instead

The UK government's 2026 AI adoption survey asks a different question. DSIT commissioned BMG Research to interview 3,503 UK private-sector businesses with at least five employees by telephone between 7 January and 21 April 2026. It excludes public administration and defence, education, human health and social work, households as employers and extraterritorial bodies. It found 32% using at least one AI technology, up from 16% in 2025. Its measure of use includes piloting.

Among 1,224 businesses that use AI (an unweighted count), 71% reported improved workforce productivity. Half reported an overall increase in productivity, which the report describes as generally modest. For operating costs, 64% reported no impact, 24% a reduction and 9% an increase. These effects are self-reported.

That is a survey of use and self-reported effects. It reports no failure rate and no return in pounds. It cannot be compared with the MIT figures, and neither report gives the other's number.

Gartner's 40%

Another number you will see is Gartner's. On 25 June 2025 it predicted that 'Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls'.

That is a forecast, not a measurement. It covers projects built around AI agents, not every generative AI pilot. Its end date has not arrived, and the press release gives no method and no base number of projects. We checked its wording on Gartner's own page and in an Internet Archive capture of 22 October 2026.

The report's status

The report does not say it was peer reviewed. It lists one reviewer, from Project NANDA itself.

MIT's PDF was available on 18 August 2025, returned 'not found' that afternoon, and by 21 August the site offered reports through a request form instead. Today the old link redirects to an overview page that does not list the report. We found no statement on MIT's own pages about why, and no correction or follow-up.

We read a copy hosted elsewhere. It is identical, by checksum, to four Internet Archive captures of MIT's own file taken on 18 August 2025.

What to do with these numbers

Our suggestion: do not plan around the 95%. Use it as a prompt to ask better questions before you start a pilot.

Before you start, write down:

• the one workflow you are testing, and who owns it
• what counts as success, as a number you can check
• how long you will wait before judging it
• what you will do if it does not work, including stopping

In our view the most useful line in the report is in its executive summary, not its headline. It says the divide 'does not seem to be driven by model quality or regulation, but seems to be determined by approach'. The report's reading is that how a tool is chosen and fitted into the work seems to matter more than the tool itself. Its evidence is interviews and a conference survey, not a controlled comparison, and it lists regulation among 'External Factors Not Fully Addressed'.

What we could not establish

Whether the report was peer reviewed. It does not say.

What the report counts as a pilot. It does not define the term.

What the percentages in its main chart are a share of. The chart does not label its base.

How the report derived its figure of about 83% for chatbots.

Where the 150 interviews and 350 employees in some coverage came from. They are not in the report.

Why MIT stopped serving the PDF. We found no statement.

Whether MIT has described the report as unpublished or not peer reviewed. A secondary source says it has. We could not verify that.

Whether any respondents were based in the UK. The report does not say.

How Gartner reached its 40% forecast. The press release does not say.

How it was researched

We read the full 26-page report, the two Fortune articles that popularised the figure, Gartner's press release and the UK government's 2026 AI adoption survey, all on 10 November 2026. We could not read the Reuters report of the Gartner forecast. We used no blog or social post for any number. We drafted this article with AI assistance. A second AI reviewer, working separately, then checked each claim against the sources on 10 November 2026, and we changed the article where it found errors. BespokeWorks sells AI consulting and bespoke software, so we have an interest in this topic. We have tried to say only what the sources say.

Sources

Links open the source. Page dates are those shown on each page. All were read on 10 November 2026.

The MIT report

1. MIT NANDA, The GenAI Divide: State of AI in Business 2025 (July 2025), copy hosted by MLQ
2. Internet Archive capture of MIT's own file, 18 August 2025
3. MIT Media Lab, NANDA overview

Press coverage

1. Fortune, MIT report: 95% of generative AI pilots at companies are failing (18 August 2025)
2. Fortune, An MIT report that 95% of AI pilots fail spooked investors (21 August 2025)

The Gartner forecast

1. Gartner press release, 25 June 2025
2. Internet Archive capture of the Gartner press release, 22 October 2026

UK evidence

1. UK government, AI adoption survey 2026 (published 29 October 2026)
2. AI adoption survey report 2026: executive summary
3. Artificial intelligence adoption research 2026: full report