AI Hallucination (Finance)

An AI hallucination is output a model states with full confidence that doesn’t hold up: a fabricated figure, a misquoted source, or a plausible number that simply doesn’t tie to the ledger. In finance the risk is sharper than in most domains: a hallucinated number often looks exactly like a real one.

Why hallucinations are harder to catch in finance

A hallucinated citation in a legal brief is checkable: the case either exists or it doesn’t. A hallucinated financial figure has no such tell. It arrives formatted correctly, in the right order of magnitude, in a table next to figures that are genuinely correct. Reviewers scanning a board pack apply reasonableness tests, and a good hallucination passes every one of them, because language models are optimized to produce output that looks like the training data, and real financial statements are the training data.

The failure modes cluster into a few shapes. The model invents a figure that was never in the source. It takes a real figure from the wrong period, entity, or scenario. It performs arithmetic in tokens rather than in a calculation engine and gets a rollup subtly wrong. Or it summarizes a document correctly but attributes the summary to a source that says something else. The last two are the most common and the least visible.

Where it bites: close, audit, and board reporting

During close, a hallucinated accrual or a misread subtotal propagates into every downstream statement before anyone reviews it. In audit, an unreproducible figure is worse than a wrong one: the finding is not just the error, it’s the absence of a control that would have caught it. In board reporting, the number is already public inside the company by the time someone rechecks it, and the correction costs more credibility than the original error cost accuracy.

What mitigation actually looks like

Confidence scores and reasoning traces show you a guess happening; they don’t stop it from shipping. The controls that work are structural: keep the model out of the arithmetic and give the sums to a deterministic calculation engine, require every stated figure to carry a citation back to a source record, and refuse to render a number whose check hasn’t passed. That last part is the difference between explainability and verifiability.

Most FP&A vendors avoid the word, talking instead about “explainability” or confidence scores. Neither stops a wrong number from shipping. Rexfin’s answer is to make the AI unable to state a figure it can’t trace back to source: sources reconcile into one governed model, calculations run deterministically, and every figure carries its citation. See why the right numbers can still add up to wrong math.