Why 95% of AI pilots never reach the P&L — and what the other 5% do differently

MIT's NANDA initiative published a number in 2025 that should be taped to the wall of every executive conference room: after studying hundreds of enterprise GenAI deployments against tens of billions of dollars in spending, they found that roughly 95% of pilots deliver no measurable P&L impact. Only about one in twenty reaches production with value you can point to in the financials.

Read that carefully. It doesn't say AI doesn't work. It says most deployments of AI don't work — which is a very different diagnosis, with a very different cure.

The wrong autopsy

When a pilot dies, the internal story is almost always about the technology. The model hallucinated. The tool wasn't ready. Maybe next year.

The MIT researchers found something else. The pilots didn't stall because the models were weak — they stalled because of flawed workflow integration. The tool arrived, but the work didn't change. The pilot ran next to the business instead of inside it.

You've probably seen the pattern up close:

• A tool gets adopted by one enthusiast, who becomes its only user — and its single point of failure.
• The pilot has no baseline, so nobody can say whether it helped. Six months later, "did it work?" is answered with anecdotes.
• The workflow it was supposed to improve still exists in its old form, so people quietly revert the moment the pilot hits friction.
• Nobody owns it. Not the intern who set it up, not the manager who approved it, not the exec who mentioned it at the offsite.

None of those are technology failures. They're organizational failures — and that's actually good news, because organizational failures are fixable without waiting for a better model.

What the 5% do differently

Study the deployments that do reach the P&L and a consistent playbook emerges:

They pick one workflow, not a "strategy." The winners don't roll out AI everywhere. They choose one process with real hours in it — intake, reporting, follow-ups, quoting — and go deep.

They capture a baseline before building anything. How many hours does this take today? What's the error rate? What does it cost per instance? Without this number, your pilot cannot succeed — not because it won't help, but because you'll never be able to prove it helped.

They build with the operators, not around them. The people who run the workflow are in the room from day one. This isn't a courtesy; it's the mechanism by which the new way becomes the default way. Software imposed on a team gets abandoned at the first inconvenience.

They measure in three tiers. Usage first (is anyone touching it?), workflow efficiency second (time and cost against the baseline), P&L impact third. Measuring only usage produces what we call vanity adoption — dashboards full of logins, financials full of nothing.

They end with a verdict. A pilot should terminate in a written go/no-go. Expand it, run it as-is, or kill it. The 95% mostly never end at all — they just fade, which is how a company accumulates a graveyard of half-dead subscriptions.

The uncomfortable question

Here's the test we'd put to any leadership team: if your most AI-fluent person resigned tomorrow, what would still be running a month later?

If the honest answer is "almost nothing," you don't have an AI capability — you have an AI dependency on one or two individuals. The difference between the 95% and the 5% is exactly this: whether AI lives in people's browser tabs or in the company's processes.

That's the entire premise behind how we work. An audit to find where AI actually pays off (and where it doesn't — the skip list matters as much as the roadmap), one pilot measured against a real baseline, then embedding: process, training, documentation, ownership. Not because it's fashionable, but because it's what the successful minority actually does.

---

Source: MIT NANDA, "The GenAI Divide: State of AI in Business 2025." If you want to know which of your workflows would survive contact with this playbook, the AI Ownership Scorecard takes four minutes — or skip straight to the AI Opportunity Audit.