What a banking P&L teaches you about AI

Eight years with credit authority and a revenue book of around USD 100M turn out to be oddly precise training for building with language models. Three lessons transferred whole.

1 · Nothing exists without provenance

A credit file is not a story; it is a chain of custody. Every figure traces to a statement, an auditor, a date — because someone with signing authority will be asked, years later, why they believed it. Language models produce the opposite by default: fluent figures with no custody at all. The fix isn't better prompting; it is architecture. In the systems I build, a model may fetch, structure, explain and verify — it may never mint a number. Every quantity carries value, units, source, grade and date, or it does not ship. This single rule separates systems a counterparty can rely on from demos that impress in a meeting and die in diligence.

2 · The power to say no is the product

A bank's real product is not money; it is underwriting — the discipline of declining. A model that answers everything is a salesman. The systems worth building refuse: push an assumption past what evidence supports and you get the binding constraint, named, in your own units. In front of sophisticated audiences this lands almost strangely well — the refusal is the credibility. I have watched a computed "no" convince a room that every "yes" before it had been earned.

3 · Size the material to the decision-maker

I have sat on the receiving end of data rooms measured in the thousands of pages — corpora no committee could digest in the time a decision allowed, from issuers whose numbers later proved the point. The lesson cuts both ways: overwhelming your reader is indistinguishable, in outcome, from deceiving them. So the systems I build are layered by reader: the committee gets statements and sensitivities in their own format; the engineer gets full depth; the generalist gets three minutes that leave them genuinely informed — all rendered from the same underlying truth, so the layers cannot drift apart. AI makes this cheap for the first time. That, more than any model capability, is what should worry incumbent process — and delight anyone who has ever had to actually read the annexes.

The convergence

The deeper point: AI systems are becoming counterparties — advising, screening, underwriting. The standards we will demand of them are the ones banking already invented: auditability, stated limits, disclosed assumptions, someone accountable for the model's conduct. People who hold both languages — credit and code — are going to be busy. It is not a crowded intersection.

— Oscar · Stockholm, July 2026