The PM Who Writes Their Own SQL

The real cost of asking someone else

When you can't query the data yourself, every question has a price: a ticket, a queue, an analyst's afternoon, two days of waiting. So you ration. You only ask the questions you're confident are worth asking, which means you only ask the questions you already half know the answer to.

That's the damage. Not the delay, the self-censorship. The best product insights come from the fourth or fifth question in a chain, and you never get there if question one costs two days.

When the cost is five minutes, you follow hunches. You check the stupid thing. And the stupid thing is right often enough to justify the whole habit.

Metrics are a design decision, not a report

I picked up a lot of this running analytics for fintech clients, and it's the thing I most want teams to understand: the metric you choose determines the product you build.

Pick "applications started" and you will build a beautiful funnel entrance. Pick "businesses successfully funded" and you will suddenly care about underwriting turnaround, document upload failures, and the customers who quietly gave up on step four. Same team, same quarter, completely different product.

So I fight about metric definitions early and in public. It looks pedantic. It's the cheapest steering you will ever do.

Where AI actually helps here

The obvious use is writing the query for you, and yes, it does that well. But the more useful shift is at the other end. Analysis was always bottlenecked on the boring middle, reshaping data, checking joins, sanity-testing a number before you trust it. Collapsing that middle means more of my time goes to deciding which question matters, which was always the part only a human in the room could do.

What AI does not do is tell you the number is wrong because the events were instrumented badly in March. That still requires someone who has been close enough to the data to be suspicious of it.

The habit, stated plainly

Stay close enough to the numbers to be surprised by them. A dashboard someone else built shows you what they expected to happen. Writing the query yourself is how you find what nobody expected.