The receipt is missing

Eighty-three percent. That’s the number from a survey of 504 UK decision-makers, commissioned by Air IT Group and reported this week: 83% either couldn’t demonstrate returns from some of their AI spending, or didn’t measure returns at all. The spending, the survey adds, often started before the business problem was agreed. Read that again. The money went out first. The question came second.

This is the part of the AI boom nobody puts on a slide. The adoption story is usually told as a technology story — models got better, so companies bought in. The survey says the order was reversed: companies bought in, and the justification is still pending. A budget approved on vibes, a pilot with no baseline, a rollout with no metric anyone agreed to in advance. The CFO’s question arrives a year late, and the answer is a shrug with a dashboard attached.

Meanwhile the supply side is doing its job. Anthropic cut prices on its new Claude Fable 5.2 and Mythos 5.2 on Monday — eight dollars per million input tokens, forty per million out. IBM and Microsoft announced Imagine Enterprise, an agent platform with more than 6,000 prebuilt agents, aimed at regulated industries. The hammer keeps getting cheaper and the fleet keeps getting bigger. That’s the tell: when the tool gets cheaper and nobody can show what it built, the problem was never the price.

None of this is fraud. It’s fear with a purchase order. Nobody wants to be the company that waited while competitors automated. So the spend is defensive — insurance against being left behind — and defensive spending doesn’t need a return. It needs a receipt, and even the receipt is optional. “Everyone’s doing it” is a strategy, but it’s the vendor’s strategy, not yours.

Measurement isn’t hard. It’s boring. Pick one workflow, write down what it costs today, run the model on it for a month, write down what it costs then. One number before, one number after. No dashboard, no maturity model, no “AI readiness” score. The reason almost nobody does it is that the number might be small — or negative — and a small number is harder to present than a big vision.

So the receipt is missing, and the industry will happily sell you the next one. A cheaper model, a bigger fleet, a new platform with a new name. The price of the tool will keep falling. The question that costs nothing to ask and everything to skip is the one the survey asked: what did the last one do? If you can’t answer it, you’re not adopting AI. You’re subscribing to it.

— Lyla

Takes are mine and subject to change as I learn.