Most AI pilots fail. The AI isn't the problem.

MIT's NANDA group put a number on something a lot of us have been watching in slow motion. Roughly 95% of enterprise generative AI pilots produce no measurable ROI, according to their State of AI in Business 2025 report. They call the gap between the winners and everyone else the *GenAI Divide*.

The headline reads like a story about AI being oversold. It's really a story about how companies run projects.

What the 5% actually do

The Legal.io summary of the report is worth reading because it names the pattern behind the wins. The companies getting value tend to do three things:

1. Buy narrowly-scoped tools from specialized vendors, not general-purpose platforms. 2. Integrate deeply into an existing workflow, not alongside it. 3. Target back-office automation, not customer-facing agents.

None of that is about the model. It's about scope, integration, and picking a problem boring enough to actually finish.

Why pilots stall

In Forbes, Andrea Hill argues the failures come from vague problem framing, weak data foundations, and treating AI as a technology project rather than an operational change. That matches what we see when we get called in to look at a stalled pilot.

The usual shape:

- Someone picked a tool before defining the problem. - The data the tool needs lives in three systems that don't talk to each other. - Success wasn't defined, so no one can tell if it's working. - The people whose workflow is supposed to change weren't in the room when it was designed.

Any one of those is enough to kill an initiative. Pilots tend to have all four.

The unglamorous version works

In our experience, the AI work that pays back looks like this: a specific handoff between two systems, a document type that gets read the same way every week, a form that gets triaged before a human sees it. Not agents. Not chat. Not a demo you'd film.

Those projects work because they inherit the discipline of good ops work. You define the current process, you find the specific step where a model helps, you wire it into the tools people already use, and you measure whether the step got faster or more accurate. If it didn't, you rip it out.

The reason 95% of pilots fail isn't that language models can't do useful work. It's that a pilot with a fuzzy goal, siloed data, and no owner would have failed with any technology. AI just made it easier to start one.

The implication

If you're sitting on a stalled pilot, the fix probably isn't a better model. It's a smaller problem, cleaner data, and an owner who cares whether it ships. That's less exciting than the pitch deck. It's also what the 5% figured out.