Why 95% of AI pilots fail — and what the 5% do differently

What the numbers actually say

In August 2025, MIT's NANDA initiative published "The GenAI Divide: State of AI in Business 2025," built on 150 leadership interviews, a 350-person employee survey, and an analysis of 300 public AI deployments. Its headline finding: about 95% of enterprise GenAI pilots deliver no measurable impact on profit and loss. Only around 5% reach production with value the business can see.

One honest caveat before using that number in a board deck: the study counts a pilot as failing if it shows no measurable P&L impact within roughly six months. That's a strict bar, and critics have noted it. But even read generously, the direction matches every other serious dataset we have.

McKinsey's 2025 State of AI survey makes the same point from the other side: 88% of organizations now use AI in at least one business function, yet only 39% report any EBIT impact at all — and most of those put it under 5% of EBIT. The gap between using AI and earning anything from it is the defining feature of enterprise AI right now.

Why pilots stall: the learning gap

When pilots die, executives tend to blame regulation or model quality. MIT's research points somewhere less comfortable: integration. Generic tools get dropped into specific workflows, never learn the context of the work, and quietly fall out of use. The researchers call it a learning gap — on both sides, since the organization doesn't adapt the workflow and the tool doesn't adapt to the organization.

Budgets make the gap worse. The study found more than half of GenAI spend pointed at sales and marketing tools, while the most measurable returns showed up in unglamorous back-office automation — cutting outsourced process work, agency fees, and manual operations.

The build-versus-buy data is just as lopsided: tools purchased from specialized vendors or built with partners succeeded about 67% of the time, while internally built tools succeeded roughly a third as often. Most failed pilots, in other words, were structural failures — scoped as technology demos rather than as changes to how a specific team works.

What the 5% do differently

The pilots that survive look boring on paper. Across the reporting, the same operating habits keep showing up:

• They pick one workflow with a named owner — not a platform rollout in search of use cases.
• They write the baseline down before the pilot starts: hours spent, cost, error rate, cycle time.
• They ship into a real team's week, not a sandbox nobody is paid to visit.
• They measure behavior change weekly — usage that doesn't survive week four won't survive procurement.
• They buy or partner where the problem isn't differentiating, and build only where it is.

None of that requires better models. All of it requires operating discipline — which is why the divide keeps widening between organizations that have it and organizations buying more software hoping it appears.

A 90-day arc that beats a 12-month roadmap

Diagnose (weeks 1–3). Baseline where your people actually are — capability, weekly usage, momentum — instead of guessing. A structured AI proficiency assessment takes minutes and gives you a defensible starting line.

Plan (weeks 4–6). Score candidate workflows on frequency, pain, data readiness, blast radius, and ownership, then commit to two or three. (We published the scoring rubric we use — steal it.)

Activate (weeks 7–13). A named squad, protected hours, weekly demos to leadership, and a kill-or-scale decision on a date you set in advance. That cadence is the whole product of AI transformation planning: a sequence your team can actually run, with numbers attached.