Field guide · Proof
The 95% failure number is being read backwards.
The most quoted statistic in enterprise AI is used to argue that buying AI is risky. The report it comes from argues something closer to the opposite.
You have heard the number. Ninety-five percent of enterprise generative AI pilots fail. It arrives in board packs, procurement objections and roughly one in three conference keynotes, and it has become the default reason to postpone a decision.
It comes from a real piece of work. MIT’s NANDA initiative published The GenAI Divide: State of AI in Business 2025, and Fortune’s interview with the lead author reported the headline finding: about 5% of AI pilot programmes achieved rapid revenue acceleration, while the rest delivered little or no measurable impact on profit and loss. The research drew on 150 interviews with leaders, a survey of 350 employees and an analysis of 300 public deployments.
Three things about that statistic almost never travel with it, and each one changes what it means for anyone selling or buying enterprise AI.
One: the report blames integration, not the technology
The authors were explicit that model quality was not the constraint. The report attributes the gap to what it calls a learning gap, in tools and in organisations, and the lead author’s summary to Fortune was that generic tools work for individuals because they are flexible, then stall in enterprises because they do not learn from or adapt to specific workflows.
That is a finding about deployment design. It says that products which sit outside a workflow tend to die there, which is an argument about how AI gets bought and implemented rather than an argument against buying it.
Two: the report’s own data favours buying over building
This is the part that gets removed in transit. A close reading of the report notes that in the same sample, external partnerships reached deployment about 67% of the time, against the dismal internal build rate that produces the headline.
So the statistic that procurement quotes at vendors to justify caution is drawn from a study whose internal evidence suggests that working with an outside specialist is the more reliable route. The number is being deployed against the very conclusion it supports.
If you sell AI into large organisations, that is the most useful paragraph on this page. When the 95% arrives in a deal as an objection, you have a legitimate, sourced answer rather than a defensive one.
Three: consider who benefits from the framing
Worth stating carefully, because it is an argument about incentives rather than an accusation of bad faith. NANDA builds infrastructure for AI agents. The report identifies the core enterprise problem as a lack of memory and learning, then describes systems that remember and learn as the fix. The same analysis makes the point directly, noting that the report ties its conclusion to a product feature and quoting an Oxford lecturer describing the structure.
There is no evidence anybody cooked the data. There is a well-understood pattern where the diagnosis and the cure come from the same building, and readers ought to notice it. The same caution applies to the many consultancies that have since republished the figure at 90%, 85% or 80%, each with a service offering attached to the end of the failure story.
What the honest version of the picture looks like
None of this means enterprise AI is going well. Several independent sources find something similar, and they predate the MIT report.
Older research put the share of AI pilots that never reached production in a comparable range, and RAND’s work has placed AI project failure above 80%, which is roughly double the failure rate of conventional enterprise IT projects. Enterprise software rollouts are already famously bad, so doubling that is a meaningful finding.
The more current framing comes from Forrester, which reported in April 2026, on the back of a survey of 1,500 AI decision-makers, that most enterprises are still struggling to convert growing AI adoption and investment into measurable business impact. Its named barriers are worth listing: low AI aptitude among staff, an overemphasis on productivity use cases, difficulty measuring impact, and adoption siloed inside individual functions. Model capability does not appear.
Put those together and you get a defensible summary. Most organisations are not getting value from AI. The reason is overwhelmingly organisational rather than technical. And buying from a specialist tends to work better than building internally, which is the opposite of what the headline number is used to argue.
How to use this if you sell
Do not fight the statistic. Adopt it.
The strongest positioning move available to an AI vendor right now is to agree that most deployments fail, name the specific reasons in your category, and then show your deployment method as the answer to those reasons. That is a more credible conversation than another accuracy benchmark, and it maps directly onto what the research actually found.
Concretely, that means publishing your implementation model with the same care you publish your product. What the first 90 days look like. Which internal roles have to exist on the customer side. What you refuse to start without. Where you have seen it go wrong before.
Vendors treat that material as an afterthought owned by professional services. Given that most buyers now start from the assumption that these projects fail, it is the most persuasive thing you own.
How to use this if you buy
Ask any vendor quoting a failure statistic at you what their own deployment-to-production rate is, measured how, across how many customers. Most will not have the number. The ones who do have just told you something more useful than any benchmark.
And check the provenance of a statistic before it enters a board paper. This one has been circulating for a year with its caveats stripped off, and it has almost certainly delayed decisions that should have gone ahead.