Why 95% of AI pilots produce nothing
MIT researchers looked at enterprise AI pilots and found that 95% delivered zero measurable impact on profit and loss. Not disappointing impact. Zero.
That number gets passed around as evidence that AI is overhyped. I don't think that's what it shows. Plenty of companies are getting real returns. The interesting question isn't whether AI works — it demonstrably does — but why the same technology produces transformation at one company and a rounding error at another.
Having watched this from inside large organizations and now with smaller ones, the pattern is consistent, and it has almost nothing to do with technology.
The pilot was designed to succeed, not to scale
Most pilots are run by the person most excited about AI, using the cleanest data available, on a task they chose because it would work. It does work. Everyone is pleased.
Then it meets the actual business — messier data, staff who didn't volunteer, a process with fourteen exceptions nobody documented — and the results evaporate.
A pilot that can't fail teaches you nothing. The useful pilot is the one designed on a representative problem with representative data and a representative user, where failure is a real possibility. Those are less fun to present and far more informative.
Nobody defined what success meant
Ask a company what their AI pilot was supposed to achieve and you'll usually get something like "see if it could help with customer service." That's a hypothesis, not a target.
Compare: reduce average first-response time on inbound support email from six hours to under one hour, without a drop in satisfaction scores, within eight weeks. Now you know what you're testing, you'll know whether it worked, and you can decide whether the result justifies rolling it out.
Vague success criteria don't just make evaluation hard. They make failure invisible, which means the organization learns nothing and repeats the exercise six months later with a different tool.
The 80% nobody wants to do
Here's the part that decides most outcomes. Roughly 80% of the work required to move an AI capability from pilot to production isn't AI work at all. It's data engineering, process documentation, workflow integration, access control, and measurement infrastructure.
It is unglamorous, invisible in a demo, and where nearly every stalled initiative is actually stuck. The model isn't the hard part — the models are extraordinary and getting better weekly. Connecting one to the specific, undocumented, exception-riddled way your company actually operates is the hard part.
Companies that succeed accept this early and budget for it. Companies that fail keep looking for a tool that will let them skip it.
Adoption was assumed
A tool that nobody uses produces exactly nothing, and usage is far lower than leaders believe. When a company tells me "everyone's using it," the actual number is usually a third of the team using it occasionally for low-stakes tasks.
This isn't resistance. It's rational. Learning a new tool costs real time, the benefit is uncertain, and the existing way works well enough. Unless someone makes the case, provides training, and follows up, most people quietly continue as before.
Adoption is change management. It requires the same deliberate effort as any other behavioral change in an organization, and it's routinely budgeted at zero.
No one owned it
This is the root cause underneath the other four.
In most companies between 20 and 200 people, AI is assigned to whoever seems most interested — often someone in IT or operations who already has a full-time job. They get no additional time, no budget authority, and no mandate to change how other departments work.
They do their best. What they cannot do is make the whole organization take it seriously, because they lack the standing to. So the pilot happens, produces something modestly interesting, and dies quietly when its champion gets pulled onto something urgent.
Large companies mostly avoid this failure mode because they have someone senior whose job it is. Not because they're smarter — because the accountability exists.
The pattern
None of the five reasons is technical. Not one is solved by a better model or a different vendor.
Which is good news, in a sense: these are ordinary management problems, and companies solve ordinary management problems all the time. Define the outcome. Fund the unglamorous work. Treat adoption as change management. Give someone the authority and the accountability.
Do those four things and you're operating like the 5%. Skip them and it won't matter which tool you bought.
Related tool. The AI Opportunity Model builds the realization and adoption discounts described here directly into its math — which is why it produces smaller numbers than most calculators, and more defensible ones.