"95% of AI pilots fail": what the MIT report actually measured

Since August 2025 I have seen one number in more steering committee decks than any other: 95% of AI pilots fail. Sometimes it is the reason to slow down, sometimes the reason a vendor's approach is different. It comes from a report by MIT's Project NANDA, The GenAI Divide: State of AI in Business 2025, and it is usually quoted without anyone having read it.

The report is 26 pages long. This post goes through what it measured, what the 95% refers to, and what you can and cannot use it for.

The short answer: it is a preliminary, non-peer-reviewed report based on 52 interviews, 153 survey responses collected at conferences and a review of public announcements. The 95% is applied to four different things in different places, and success is defined two different ways. Its direction matches other surveys, which find that most organisations do not yet see large, measurable value from generative AI. But no other survey comes close to 95%, and the number should not be used as a failure rate.

Where the number comes from

The report, by Aditya Challapally, Chris Pease, Ramesh Raskar and Pradyumna Chari, is labelled preliminary findings and is dated July 2025 (MIT NANDA, archived copy). It names one reviewer, who is also a co-author, and it does not claim peer review. The PDF disappeared from MIT's website in August 2025, shortly after the press coverage, and the old address now redirects to the research group's page; the Internet Archive keeps the original. Its evidence base, as the report describes it:

Four different 95%s

The figure that travels is "95%", but the report attaches it, or its inverse 5%, to different units in different places:

Success is defined twice, too. On page 7 a tool counts as successfully implemented when users or executives described it as causing a marked and sustained productivity or P&L impact: a self-reported judgement. On page 24, the methodology defines success as deployment beyond the pilot with measurable KPIs, which is a lower bar. The funnel behind the 5% looks like this:

There are smaller problems as well. The headline $30 to 40 billion of enterprise investment has no source. One page says half of generative AI budgets go to sales and marketing, another says about 70%. And the report closes by promoting the NANDA framework as part of the answer, an interest worth noting.

What the report gets right

It would be a mistake to dismiss the whole report. Three of its observations match what I see in client work, and are worth keeping, with their caveats:

• Shadow AI is everywhere. Only 40% of companies said they had bought an official LLM subscription, while workers at over 90% of them used personal AI tools for work. People adopt general tools faster than organisations adopt systems.
• The pilot is not the hard part. General tools spread; systems embedded in a workflow, which need integration, data access and change in how people work, mostly stall. That is an architecture and operating-model problem, the subject of Agentic AI from demo to production.
• Buying with a partner reached deployment more often than building alone: about two thirds against one third in the interviews. The report itself says this is self-reported, may reflect differences in the organisations rather than the approach, and measures reaching deployment, not P&L impact.

What other surveys find

If you need a number for how many organisations get value from AI, the larger surveys give a range, each measuring something different:

The consistent message across all five is that measurable, large value from AI is still concentrated in a minority of organisations, somewhere between one in eight and one in three depending on the question. BCG's 2025 follow-up put only 5% of companies in its top "future-built" group (BCG, September 2025). That is a sobering finding on its own, without the 95%.

How to use the number, and how not to

1. Do not quote 95% as a failure rate. It is an inverse of a self-reported success share from 52 interviews, defined two ways.
2. Quote the direction instead. Most organisations report little measurable P&L impact from generative AI so far; surveys put clear value at roughly 13% to 37% depending on the definition.
3. Define success before the pilot starts. A metric, a baseline and a threshold for going further, as in A GraphRAG business case, done the evidence-based way.
4. Measure outcomes, not opinions. Self-reports diverge from measurement; in one controlled study developers believed AI made them faster while it made them slower (Does AI make developers faster?).
5. Read the source. A 26-page PDF takes half an hour. It is cheaper than a strategy built on a headline.

Company case studies need the same reading. The most quoted one, Klarna, tells a steadier story in its SEC filings than in its headlines: see Klarna's AI customer service: the 700-agent claim and the U-turn.

How much weight this evidence deserves

Sources

• Challapally, Pease, Raskar and Chari (July 2025). The GenAI Divide: State of AI in Business 2025. MIT NANDA, preliminary findings. Internet Archive copy of the original MIT file.
• Gartner (October 2026). Survey finds only 13% of CIOs say their enterprise has realized significant or transformational value from AI.
• IBM Institute for Business Value (May 2025). CEO Study 2025.
• BCG (October 2024). Where's the Value in AI?
• BCG (September 2025). The Widening AI Value Gap.
• McKinsey (August 2026). The state of AI in 2026: On the road to ROI.