95% OF AI PILOTS RETURN NOTHING.

THE NUMBER NOBODY FRAMES

MIT's NANDA initiative reported in 2025 that about 95% of generative-AI pilots produce no measurable P&L impact. The number made headlines as "AI doesn't work." The full reading is more useful: AI works fine — pilots are structured to fail.

THE AUTOPSY

• The model is rarely the problem. Across Stanford's documented enterprise deployments, 77% of the hardest problems were process, data and change management. In 42% of deployments the model itself was fully interchangeable — 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. In the Stanford evidence base, value only registered once systems ran inside live workflows with real users. Organizations that kept manufacturing proof-of-concepts captured almost none of it.
• Nobody defines the finish line. A pilot without pre-agreed KPIs cannot succeed, because there is nothing to succeed against. The measurement has to be designed before the build, not reconstructed after it.

WHAT THE 5% DO DIFFERENTLY

They run the numbers first and gate the budget on them. They fix the process before wiring AI into it. They ship one narrow workflow into production instead of piloting six in a lab. And they treat a failed first attempt as tuition: 61% of successful deployments had a failed attempt behind them. The second run is where it pays — if the second run changes the method, not just the vendor.

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. That discipline is the entire difference between the 95% and the 5%.