Skip to content

Guide · Buy versus build

Why most AI projects fail

MIT looked at what happened to enterprise AI pilots, and the headline number gets quoted everywhere. The number worth your time is the gap between the teams that built alone and the teams that brought in outside help.

95%

of enterprise GenAI pilots delivered no measurable P&L impact.

Source: MIT NANDA, The GenAI Divide: State of AI in Business 2025

The report behind that number (MIT NANDA, The GenAI Divide: State of AI in Business 2025) rests on 52 executive interviews, surveys of 153 leaders and analysis of 300 public AI deployments. Worth knowing before anyone quotes the figure at you, including us.

The number underneath the headline

67% vs 22%

Pilots that blended internal people with external expertise succeeded 67% of the time. Internal IT-only builds succeeded 22% of the time.

Source: MIT NANDA, The GenAI Divide: State of AI in Business 2025

Roughly three times the success rate, for the same technology.

Why they stall

The tools that failed could not retain feedback, adapt to the context of the business, or improve over time. They answered, then forgot. The 5% that worked were embedded into one high-value workflow instead of being bolted onto the side of the business as a generic chatbot.

Source: MIT NANDA, The GenAI Divide: State of AI in Business 2025

Read that again with our name on it

We get paid to build these things and we are showing you a study that says buying beats building, which means we benefit from you believing it. That is exactly the sort of claim worth checking for yourself, and you can: the report is public, and it is called “The GenAI Divide: State of AI in Business 2025”.

And here is the honest limit of it: the study looked at enterprise pilots, so the 95% belongs to them. What the equivalent number is for small businesses, nobody has measured yet, so we leave the figure exactly where the research left it.

What to take from it

Pick one workflow that costs you money every week (the enquiry nobody answers overnight, the booking that takes eight emails) and fix that one. A single named workflow beats “AI for the business” every time, because the projects that worked did one thing properly and then earned the next.

Want to find the one workflow worth fixing first?

Up to eleven questions, about three minutes. The questions cover one week in your own business, and the report names the builds your answers point at, in the order they are worth doing.

Get my free score