Give me the insight, not the decision

I'm talking with a number of organizations about AI right now, and there is a panic that keeps coming up. The budget is blown, and no one can say why. There was a set figure -- based on maybe thirty percent using the tools seriously and the rest not at all -- and then it was blown past by a wide margin.

I usually ask what they think is causing it. And then it starts coming out. They don't know who is using what. They don't know whether it's being used for work or personally. The pricing models from vendors keep shifting. Someone wanted the twenty-dollar license, someone else pushed for the two-hundred-dollar tier, and no one governs it -- it just happens. In a company with thousands of employees there is no reasonable way to attribute the cost against a set budget.

What happened was that licenses were bought and it was called transformation. The C-suite said everyone gets licenses, that's how we drive AI development. But you don't buy governance by buying licenses. You can't govern what you can't see.

What I Actually Built

So I built the measurement myself. I clone all the repos locally and look at what can actually be measured. Commits per developer over the past month. Lines in and out. Spend per person against the API. It becomes an overview: who has pushed three hundred lines, who has pushed ten thousand, who is running a few hundred dollars in consumption over the past week.

That part is deterministic. It's programmatic. There is no judgment in collecting it, and so I automate it without blinking.

Where the Line Is

But then comes the question I'm waiting for. Can I get the automated report? The one where the model doesn't just gather the data but also draws the conclusion and tells me what I should do.

That's when I almost leave the room.

Gathering is one thing. Deciding is another. I won't act on a conclusion I can't validate, and the model isn't good enough for me to want to. Give me the insights. Put them on the table, gathered deterministically. Then I make the decision.

Because that's where judgment lives. Commits and lines are a rough signal, nothing more. One of the most senior people I work with writes few lines of high complexity -- sits for a day and thinks and checks in four lines that go straight to production. Three hundred lines from someone else might mean they were never really working at all. The number points. It does not decide.

The One Thing I Cannot Delegate

A language model solves what you ask it to. It cannot tell you what the number means for this particular organization, this person, this month. That interpretation is the decision. And the decision is what I'm responsible for.

If you automate the decision, you have automated away exactly what you were put there to own.

The report is the easy part. The judgment on top of it is the work. That's why it stays with a person -- and stays there.

See also: I Am Smarter Than AI (series 1) and The Checkpoint Is the Job.