AI doesn't fail on the model. It fails on everything you didn't write down.

I gave a talk in London recently on AI adoption, and the argument I kept coming back to is one I'd been circling for months at work: the return on AI is decided by how you prepare the organisation, not by which model you buy.

Start with the number everyone has seen. MIT's Project NANDA studied enterprise GenAI deployments and found roughly 95% of pilots produced no measurable impact on profit or loss, against tens of billions in spend. That statistic got read as "AI doesn't work." That's not what it says. The researchers were measuring whether integrated pilots moved the P&L, and most didn't. It's a finding about how companies deploy, not about whether the technology functions. One summary of it put the conclusion better than I could: they didn't find that AI doesn't work. They found that enterprises don't work, at least not in the way AI needs them to.

That matches what I see. Intelligence arrives in organisations whose knowledge was never written down in a form anything but a human could read. The process lives in someone's head. The standard is "ask Kwame, he knows." The real decision history is in a chat thread nobody can find. Then a model shows up, and we're surprised it can't reason about a context that was never actually externalised.

So the hard part of AI adoption isn't the part anyone budgets for. It isn't the licence or the integration, or even the training. It's the unglamorous work of getting your organisation's knowledge into a shape that something other than a human can use. Your data, yes, but also your standards, your decision records, your definitions of good. The stuff that was always implicit, made explicit.

And here's what makes this so easy to get wrong: that work has no demo. You can't show it in a board meeting. Buying a model feels like progress. Writing down how your organisation actually works feels like admin. So the money goes to the visible thing and the pilot quietly joins the 95%.

The uncomfortable implication is that most of us have been grading ourselves on the wrong exam. The question isn't "which model should we use?" Models are becoming a commodity and they improve without you. The question is whether your context, the layer underneath, is yours, deliberate, and fit for a machine to read. Own the layer, rent the model.

If you're about to start an AI programme, I'd spend the first month not on tools at all. I'd spend it finding out how much of what your organisation knows exists only in people's heads, and starting the patient work of writing it down. Less exciting. Far more likely to be in the 5%.