95% OF AI PILOTS SHOW NO P&L IMPACT.

THE NUMBER NOBODY FRAMES

In July 2025, MIT's NANDA initiative reported that 95% of organizations in its study were getting zero return from generative AI: just 5% of integrated pilots were extracting millions in value, and the rest showed no measurable P&L impact. Success meant a deployment past the pilot stage with measurable KPIs, judged about six months after the pilot, and the evidence combined interviews, a survey of business leaders and a review of more than 300 public AI deployments. The number made headlines as "AI doesn't work." The report's lead author put the failures down to a "learning gap" in how companies integrate the tools, and said model quality wasn't the issue.

THE AUTOPSY

• The model is rarely the problem. In Stanford's study of 51 successful enterprise deployments, the hardest problems in 77% of them were process, data and change management. In 42% of the deployments the model itself was fully interchangeable, and the organization around it decided the outcome.
• Most companies automate the mess as-is. In Accenture's Pulse of Change survey (2026), only 21% of executives said they have redesigned end-to-end processes with AI at the core. Bolting an agent onto a broken workflow produces a faster broken workflow.
• Demos are not deployments. Stanford's study only counted systems running in production with real users and measurable value. Its authors call the opposite pattern a "PoC factory": pilots commissioned one after another with no link to company strategy.
• Nobody defines the finish line. MIT counted a pilot as a success only if it moved past the pilot stage with measurable KPIs. A pilot without agreed KPIs can't pass that test, because there is nothing to measure it against. Design the measurement before the build; numbers reconstructed afterwards rarely convince anyone.

WHAT THE SUCCESSFUL ONES DO

The 51 successful deployments in the Stanford study share a few habits. Every one of them started small and expanded. They treated the process as part of the project and fixed it before wiring AI into it. They put one narrow workflow into production instead of piloting six in a lab. And they treated a failed attempt as tuition: 61% had at least one failed attempt behind them. A second attempt pays off when it changes the method; switching vendors alone rarely does.

THE OPERATING RULE

Our version of it is one sentence: if it doesn't pay back, it doesn't ship. Payback is calculated before the build, measured after it, and either number can kill the project. No rule guarantees a place in the 5%, but this one closes the most common route into the 95%: a build nobody agreed how to measure.