There are two stories about AI returns in the enterprise and both are true at once. That is precisely why boardrooms are so confused.
The four numbers that matter
- 88% of organisations have adopted AI, per the 2026 AI Index from Stanford HAI.
- 80% of McKinsey’s respondents report individual productivity gains.
- 37% attribute some EBIT impact to AI use — roughly the same share as last year.
- ~6% are high performers: they attribute at least 5% of EBIT to AI. That figure has been flat for several cycles.
And the most-quoted datapoint of all: The GenAI Divide, from MIT’s NANDA project, found that 95% of generative AI pilots produced no measurable impact on the P&L, across 52 executive interviews, 153 surveys and analysis of 300 public deployments.
Why productivity doesn’t reach EBIT
If 80% feel they work better and only 37% see it in the accounts, value is leaking somewhere. In the projects we audit, it leaks in three places:
1. The saving is spread in crumbs. Twenty minutes a day for each of fifty people is 4,000 hours a year on paper. In practice those twenty minutes get reabsorbed: into other tasks, into meetings, or simply into doing the same work under less pressure. For the saving to reach the P&L it has to convert into something concrete — more volume on the same headcount, less outsourcing, a lead time that lets you charge more.
2. The process wasn’t redesigned. Dropping AI into a workflow built for humans gives you a slightly faster human workflow. The step change only comes when you change who does what and in what order. That is the expensive part, and almost nobody budgets for it.
3. Nobody measured the before. Without a baseline, any ROI discussion is a discussion of opinions. And since the opinion of whoever paid for the project carries most weight, the project “works” until someone asks for the number.
The pattern in the 5%
The interesting thing about the MIT study is not the 95% that fail, it is the 5% that don’t. That group extracts real value, and it is not distinguished by having better models. It is distinguished by integrating AI into a specific workflow instead of deploying a generic tool and waiting.
It matches what we see: the projects that move the needle are surprisingly narrow in scope and surprisingly deep in integration. One process, end to end, with an owner, a metric and a budget.
The ones that don’t move it tend to be the opposite — broad, shallow and ownerless.
How to build the business case without lying
- Pick a metric that already exists on your dashboard. If you have to invent the metric to justify the project, the project isn’t justified.
- Measure the baseline before touching anything. Two weeks of data on the current process. It is the highest-return investment in the whole project.
- Decide in advance what you will do with the freed capacity. With no answer, the saving will not exist in accounting terms.
- Set a review date and an abandonment criterion. A project without an abandonment criterion is not an investment, it is a subscription.
The honest reading
AI is not delivering the return the market promised, and that does not mean it delivers none. It means the return is concentrated in few hands, and that the practices producing it are known, boring and unglamorous.
The 6% of high performers has been 6% for years. What has changed is that we now know reasonably well what they do differently.
Sources
- The 2026 AI Index Report — Stanford HAI
- The State of AI: Global Survey 2026 — McKinsey
- The GenAI Divide: State of AI in Business 2025 — MIT NANDA (report PDF)
- MIT report: 95% of generative AI pilots at companies are failing — Fortune
- MIT Finds 95% Of GenAI Pilots Fail Because Companies Avoid Friction — Forbes