what the AI era asks of finance

Models are cheap now; judgment isn’t. Notes on where the leverage in finance actually sits.

For most of its history, the constraint in finance was analysis. Building the model, reading the filings, reconciling the numbers — that was the work, and the people who could do it fast were the leverage. That constraint is gone. A competent model of almost anything now costs minutes and cents.

What didn’t get cheaper is deciding whether the model is right, whether the number can be signed, and whether the answer holds up when someone asks why. The scarce thing moved from producing analysis to standing behind it.

Watch where the time goes on a desk that has adopted AI seriously. It isn’t spent typing — it’s spent reviewing. The analyst’s job shifts from assembling the memo to interrogating it: does this figure match the filing, does this assumption survive contact with the last three deals, would I defend this line in front of the committee.

That review work is judgment, and judgment doesn’t scale the way generation does. It’s earned per firm, per desk, per person. Which is why the firms getting real leverage from AI aren’t the ones with the best prompts. They’re the ones who encoded how their own people decide — and made the machine work inside that.

Finance runs on accountability. Every report has a name on it, and the name matters more than the report. So the real requirement for AI in this industry isn’t accuracy in the abstract — it’s provenance. Where did this number come from, what did it pass through, and who touched it on the way.

This is a design constraint, not a compliance afterthought. An agent whose output can’t be traced will be checked line by line, which erases the time it saved. An agent whose every figure links back to a filing, a feed, or a decision gets reviewed at the level of judgment — which is where the human belongs anyway.

The judgment a firm runs on is mostly undocumented. It lives in how a sector head marks up a draft, which covenants a treasurer actually worries about, what the last committee rejected and why. None of that is in a foundation model, and none of it arrives with an API key.

Capturing it is slow, specific work: sitting with the desk, mapping the workflows where judgment repeats, building a graph of the firm’s own documents and decisions. That’s the unglamorous part of the AI era, and it’s the part that compounds. Models will keep getting better for everyone. The context is yours alone.

Data gives the machine something to stand on. Craft makes the output fit the way the firm actually works. Judgment decides what ships. Get all three and you have the only thing this industry ultimately pays for — trust.