07// ADVISE

A year of AI spend, no number next to it

We have been at this a year and I still cannot put a number next to any of it.

// The diagnosis

There was never a definition of what good looked like, so nothing was ever measured against anything, and now every conversation about value is a conversation about impressions. The pattern is familiar: budgets are not being cut, but the money has moved from a curiosity gate to an evidence gate, and the programmes that cannot produce evidence are the ones quietly wound down at the next planning round.

// ADVISE

Define what good means before anything else, in a way your business would recognise. Read and group the failures you already have. Build a graded set from your own cases rather than generic examples. Then a score the business can read without an engineer translating it, and a check that it stays still when nothing has changed.

// Already built

I built this for a commodities trading firm: an evaluation framework for an AI advisory product, so the business could tell whether the model was right rather than just fluent.

// The outcome

A number you can take to the board, and the confidence that it moves for reasons you can explain.

// Before you call

Ask whoever owns it to show you, with evidence, that it is better than it was three months ago. If nobody can, that is the finding, and it is the same question your next planning round will ask.

Not this if nothing is in production yet. Build first.

// If this one is yours

It goes straight to my inbox. No sequence, no sales team.