Most big companies have an AI pilot running somewhere, and usually several. Ask how many have made it into the way the business actually runs and the room goes quiet.
There's data behind that silence. A report from MIT's NANDA initiative, covered by Fortune, found only about 5% of AI pilot programs achieve rapid revenue acceleration. The vast majority stall, with little or no measurable impact on the P&L. The research drew on 150 interviews with leaders, a survey of 350 employees and an analysis of 300 public AI deployments.
The authors didn't blame the models. They pointed to what they call a learning gap, in the tools and in the organisations using them. Generic tools work well for individuals, then stall inside companies because they don't adapt to how the work actually flows.
Pilots built to survive, never to decide
Here's my read on why so many stall. Most AI pilots are set up to keep going. They have a sponsor, a budget and a vague goal like "explore use cases". Nobody writes down what result would end it, so nothing ever does.
A startup can't afford that. It runs a bet with a number to hit by a set date. If the number doesn't move by the date, the bet dies and the team moves on. That discipline is the whole point of lean startup, and it's exactly what most corporate pilots are missing. I've written about how a lean startup workshop for corporate teams builds it.
Pick the boring problem
The report found more than half of generative AI budgets go to sales and marketing tools. The biggest returns showed up in back-office work instead, like cutting outsourced processes and external agency costs.
That tracks. The flashy customer-facing demo wins the budget. The dull internal process, the one somebody does by hand every Friday, is where the tool would actually pay for itself. If you're choosing your next pilot, start with the work people hate doing and can measure easily.
Buy before you build
Another finding should sting a few IT departments. Buying AI tools from specialised vendors and building partnerships succeeded about 67% of the time. Internal builds succeeded only a third as often. Almost everywhere the researchers went, companies were trying to build their own anyway.
Building can still make sense. But if your pilot starts with a six-month internal build before anyone tests the idea, you've already made the most expensive choice before learning anything.
Put line managers in charge
The report also flagged empowering line managers, rather than only central AI labs, as a factor in success. That's the person who owns the workflow and will notice on day two whether the tool helps. They also need a team willing to say when it doesn't, which is why psychological safety is the AI adoption bottleneck in so many rollouts.
Central teams are good at platforms and governance. They're usually too far from the work to know whether a pilot is changing anything. Give the line manager the question and the kill date, and give the central team the job of making it safe.
How I'd run the next one
Write a one-line hypothesis. "If the claims team uses this tool on first drafts, we'll cut handling time by a fifth within eight weeks." Agree the number before you start, and agree what happens if you miss it.
Run it small and fast on real work. Check the number at the date. Then make a call to scale it or stop it. Either answer is useful. A pilot still running a year later with no decision has taught you nothing.
Our lean startup and product sprint workshop teaches teams to set up bets this way, with a hypothesis and a kill date the sponsor signs off before day one. Bring a pilot that's stalled and we'll rebuild it as a bet.
