Last year MIT’s Project NANDA published a number that went round the world. Roughly 95% of enterprise generative AI pilots produced no measurable profit-and-loss impact, despite an estimated $30 to $40 billion of enterprise spending.
Cue a thousand LinkedIn posts about the bubble.
McKinsey’s latest came to a similar place by a different route. In their August 2026 survey of 1,719 respondents, 80% said AI had improved their own productivity, while only 37% said their organisation saw any EBIT impact at all. Six percent qualified as high performers, and that number has not moved in a year. Gartner reckons more than 40% of agentic AI projects will be cancelled by the end of 2027. McKinsey AI 2026: 80% Productive, 37% EBIT (Aug 25) +2
So: everyone feels more productive, almost nobody can find the money.
Here is what I think is going on, and it has very little to do with the technology.
We keep buying AI like software and running it like a science experiment
Look at how a typical pilot is structured. It starts on the first of the month. It ends ninety days later. There is a steering committee, a success criteria document nobody reads twice, and a decision at the end that is essentially yes or no.
MIT’s own diagnosis of the failures is the interesting part, and it got far less attention than the headline. They point to a learning gap: most of these systems do not retain feedback, adapt to context, or improve over time.
Read that again and then look back at the ninety day pilot. You have built a measurement window that systematically captures the worst version of the thing you are testing, and then you hold a vote on it.
It gets funnier. One of the sharper critiques of the MIT study is that it defined success narrowly, as ROI measured six months after the pilot. Which means the most quoted statistic about AI failure may itself be a victim of the project mindset. We measured a long game over a short window and published the disappointment.
A project ends. A property appreciates.
Think about what you actually bought.
A project has a budget, a timeline and a completion date. Someone presents the benefits case, everyone claps, and the team moves on to the next thing. That is how we have funded software for thirty years and it works fine for software that does the same thing on day one thousand as it did on day one.
AI in a direct materials function is not that. It is closer to a property.
You buy it, and it is worth something. Then you maintain it. You fix the things that are wrong. You add to it. Your team lives in it and gradually makes it theirs. And five years later it is worth considerably more than you paid, provided you did not board up the windows in month two because the boiler was noisy.
Nobody sells a house because the first winter was expensive. Yet that is precisely what a binary pilot decision does.
Month one is meant to be disappointing
I will be blunt about our own product because it applies to everyone else’s too.
In the first few weeks of a Kodiact deployment, your category manager will disagree with us. Often. She has twenty years on cocoa and we have four weeks on your business, and she is going to be right more than we are.
That is not the system failing. That is the system working. Every correction she makes is captured, with the reasoning behind it. By month three the disagreements are sharper and rarer. By month six it is raising things nobody asked it to look at, because half a year of how your team actually decides is now inside it.
A ninety day pilot stops the clock at the precise moment the asset starts appreciating.
The partner point, which the data supports
There is one more finding worth sitting with. MIT found that pilots combining internal specialists with external expertise succeeded 67% of the time, against 22% for internal IT builds alone. k
Three times the hit rate, from not doing it alone.
So the recommendation is not “run an experiment and see”. It is: find a partner whose objectives genuinely line up with yours, who is going after a pain you can name without a workshop, and who is still going to be there in year three when the thing is actually worth something.
Start small. One category. The one that hurts. Then grow into the value as it compounds.
Where the compounding actually comes from
In our world it stacks up in four directions, and they multiply rather than add.
More decisions. Every call made, with its reasoning kept, makes the next one better informed. The ledger gets richer every week whether anyone is watching or not.
More categories. Cocoa teaches you something about sugar. Sugar teaches you something about freight. The context compounds across the portfolio rather than sitting in silos.
More patterns. Six months of how your team reads a market is a far better guide to what matters than any generic model of what should matter.
More simulations. Every position you rehearse adds to what the next rehearsal knows. You find out which assumptions were load bearing, and the next scenario starts from better ground.
None of that is visible on day thirty. All of it is obvious by month twelve.
The uncomfortable bit
If your AI business case needs a defined end date and a binary verdict to get funded, you will keep ending up in the 95%. Not because the technology does not work, but because the funding model guarantees you stop measuring before it does.
The six percent McKinsey found are not smarter than everyone else. They just stopped treating this as something with a finish line.
Judgment is an asset. Assets need looking after. And nobody ever got rich selling the house in the first winter.