A 2025 MIT Media Lab study found only about 5% of the enterprise generative-AI pilots it examined delivered measurable P&L impact. It was a preliminary report from the Lab’s Project NANDA, based on 52 interviews, 153 survey answers and a review of public announcements, and the number was softer than it sounded. Still, it gave every leadership team a scare and a reason to get serious. A year on, the surveys are in on what they did with it.
Mostly, they got organised. McKinsey’s global survey, published in August with 1,719 respondents, finds 44% of companies scaling AI across the business, up from 38% a year earlier, while 40% of respondents at large companies say they are scaling AI agents, up from 27%. IBM, which sells AI services itself, counts a Chief AI Officer at 76% of the companies in its CEO study, up from 26%. The org chart has embraced AI with real enthusiasm.
The profit line has not. The share of companies that attribute any operating profit to AI is 37% in McKinsey’s survey, about the same as last year. Its high performers, who credit AI with at least 5% of operating profit, are still about 6% of respondents. A year of new titles and new agents, and the group making serious money from AI has not grown.
Gartner has found a habit that marks out the leaders, and it is an unglamorous one. It surveyed 2,002 CIOs and technology executives this summer and found that only 31% have a rule for when an AI pilot should not move into production. Only 17% have one for switching off AI that is already live and not paying its way. CIOs with the first rule are 2.3 times more likely to be in Gartner’s best-performing group, and those with the second, 2.6 times. Presenting the figures in Orlando last month, Gartner’s Alicia Mullery described “a reckoning for organizations that lack fundamental disciplines, such as defining use-case value ahead of time and establishing criteria for when to walk away.”
A stop rule sounds pessimistic. In practice it is what makes success possible. Without one, every pilot ends as “promising”, and promising pilots get another quarter and another steering group.
With clients, the one number I ask them to track is the share of real work, such as tickets or invoices, that the AI finishes without a person redoing it. We measure it on live work and check it at day 60. If it sits below the level written into the business case, usually around 50%, the pilot stops, however good the demo looked.
A number like that only moves if the work moves with it. Bain, the consultancy, finds generative AI delivering about 38% of its expected value, against 52% for older automation that had two decades of process mapping behind it. In McKinsey’s survey, nearly three quarters of high performers rebuilt the workflow around AI; about a quarter of everyone else did.
So before the pilot starts, write down how the work will run if it succeeds and which number will show it by which date. The stop rule is that number, missed on that date.
Somewhere this week, a steering group is reviewing an AI pilot in its fourteenth month. The update calls it promising. It has called it promising since February.
If someone in your company signs off AI pilots, send them this one. And if it was forwarded to you, you can subscribe at vanguards.media.
David
Sources: MIT NANDA, The GenAI Divide: State of AI in Business 2025 (preliminary, July 2025). McKinsey, The state of AI in 2026, 25 Aug 2026 (1,719 respondents). IBM Institute for Business Value, CEO Study, 4 May 2026 (2,000 CEOs). Gartner, press release, 19 Oct 2026 (2,002 CIOs and technology executives). Bain & Company, Realization Rate: The Metric Reshaping Enterprise AI, 8 Oct 2026 (951 companies).