Failure starts before the model

Most AI projects don't fail on the technology.

They fail before the first line of code is written.

RAND studied why AI initiatives fail. The single most common root cause isn't the model, the data, or the budget — it's a miscommunication between the problem side and the solution side.

When the results don't show up, the easy explanation is an adoption problem. The honest explanation is almost always upstream: there was never a number the tool was supposed to move.

Years ago I advised a software company, over several years, whose shareholders were also its customers. The mandate was a cost programme, amid complaints about a backlog nobody could clear. It took far too long before the real sentence was on the table: the company shipped once a year, so every customer fix waited up to twelve months behind a release — and no cost programme has ever shortened a queue.

A wish: "We need to get our costs down."

A problem: "Every customer fix waits up to twelve months for the next annual release. We'd know it's solved when a fix reaches customers in the same quarter it was reported."

The second version takes twenty minutes to write. I was handed the first one on day one — and it took me far too long to ask for the second.

One move: before anyone picks a tool, ask the people who own the problem to write down — in three sentences — how they'd know it had been solved. Then ask the people who'd build it to sign that off. It fails loudly: if neither side can do it in three sentences, you don't have a problem yet. You have a wish.

That's the first gate. Next Tuesday, the second: the number that decides whether it continues.

That half of the work has no vendor — and, too often, no owner.

💬 Before your last rollout — was there one number it was meant to move? Yes or no.