The model never owns a number: deterministic code vs generative AI in finance

Ask a language model to summarise a meeting and it does well. Ask it to add up a column of journal lines and you have built a system that is confidently wrong some fraction of the time — with no way to know which fraction. In finance, that property is disqualifying. Not because the error rate is high, but because it is unknowable.

The answer is not a better model or a cleverer prompt. It is a division of labour: deterministic code — SQL, Python, statistics — computes every figure and every flag, and the model is only allowed to narrate results it never produced.

A simple test for any AI finance tool

Switch the model off. If any number changes, the model owned that number — and every output it touched inherits the model's failure modes. If every number is unchanged and only the prose disappears, the architecture is sound.

This test takes one sentence to state and exposes most of the market. “AI-powered analysis” usually means the model sits somewhere in the computation path. It should sit outside it.

What the model is actually for

None of this makes the model useless — it makes it safe to use. Narrating a variance in plain language, in Arabic or English. Explaining what a flag means and what a reviewer might check first. Turning a structured exception queue into a readable briefing. These are language tasks, and language is what the model is good at.

The boundary is simple to state and simple to audit: computation is deterministic, narration is generative, and the two never trade places.

Why this matters more than accuracy benchmarks

A deterministic test either fires or it does not, and it fires the same way on the same data every time. That is what lets a controller defend a number in front of an auditor: not “the model is usually right”, but “this figure came from this query over these rows, and here they are.” Reproducibility is not a nice-to-have in finance. It is the whole game.