The 95% AI pilot number says less than you think

A year on, the MIT figure still turns up in board packs and sales decks. Ninety-five percent of companies get no measurable return from generative AI. One person quotes it to kill a project. Another quotes it to sell a rescue. We have heard it used both ways in the same week.

It deserves a closer read, because what the report says about the cause is more useful than the number.

What did the report measure?

The study is "The GenAI Divide: State of AI in Business 2025" from MIT's NANDA initiative, released in the summer of 2025. Per the coverage in Fortune and Virtualization Review, it drew on a review of more than 300 public AI initiatives, 52 interviews and surveys of 153 leaders. The headline finding: about 95% of organisations saw no measurable effect on profit and loss, and only about 5% of custom enterprise tools reached production.

That is a modest sample for a number that now travels like a law of physics. It is a fair warning sign. It is a weak basis for "AI does not work."

What did it blame?

This is the part that gets cut when the number is repeated. The authors did not put the failures down to weak models. They called it a learning gap. Most of the systems they looked at "do not retain feedback, adapt to context, or improve over time."

Read that again if you run a firm. The tool that drafted your proposal on Monday knows nothing on Tuesday about the correction your operations lead made to it. The pricing rule your finance manager explained in a chat lives in that chat. Next week a colleague opens a fresh session and the AI makes the same mistake with the same confidence.

The people interviewed for the report said it in plainer words. Fine for a first draft. Not trusted for anything that matters, because it forgets.

Why is that a memory problem?

A pilot that cannot remember what it was told cannot get better. A pilot that cannot get better never earns a place in a real process, so people stop opening it. Six months later it shows up in a survey as "no measurable return."

The fix we run on ourselves is unglamorous. We keep a written company memory: plain files holding our rules, the facts we have checked, and the corrections we have made, each with the reason it was made. Our AI sessions read it before they start work. When someone corrects an output, the correction goes into the memory, so the next session starts from it instead of from zero.

Two details carry most of the weight.

1. The memory belongs to the company, not to one person's chat history. A new colleague's session inherits what the last person taught it.
2. Every correction carries its reason. "Never quote a delivery date before the warehouse confirms stock" is a rule with a why attached, and the why is what stops someone deleting it six months later.

On top of that sits an approval gate. Drafted proposals and client deliverables land in a review queue, and a person releases them. An action only moves to running on its own after it has proved itself. Some never do.

What will memory not fix?

We would be overselling if we stopped there. The learning gap is one cause among several. Gartner, forecasting in June 2025 that over 40% of agentic AI projects would be cancelled by the end of 2027, pointed at rising costs and unclear business value. A pilot with no business owner fails with or without memory. A pilot aimed at a task nobody needed can fail faster with memory, because it gets good at the wrong thing.

Memory also does not make a model reason better. It gives the model your facts. It cannot hand it judgment.

One thing to do this week

Pick the AI tool in your business that people have quietly stopped using. Ask its two heaviest users one question: what did you have to tell it more than once? Write their answers in one shared document. If the list runs past ten items, you have found your learning gap, and you have the first page of a company memory.

To see where your firm stands on the other causes, try the AI readiness check. Book 20 minutes at pgco.world/book.