AI in Finance — Frequently Asked Questions
Where AI fits in finance and where it must not — why the model must never be the source of a number, the reconciliation problem underneath, document extraction, anomaly detection and audit expectations.
We have not built a financial system. See AI Finance for the boundary we would draw.
The boundary#
What is the single rule?#
The system may find, structure, explain and draft. It must not be the source of a number.
Every figure reaching a report should trace to a system of record, and the arithmetic should be done by something deterministic. A model produces a plausible number as readily as a correct one, and in finance a plausible wrong number is the failure.
Can a model do the calculation if it is careful?#
Do not let it. The reason is not capability, it is verifiability — you cannot unit-test a model's mental arithmetic, and you cannot reproduce it either, since the same question can be answered two ways on two runs.
Arithmetic belongs somewhere a test can pin it down. Ask the model which numbers matter and why; compute them elsewhere.
Why does a wrong figure matter more here than elsewhere?#
Because of how far it travels before anyone checks. Financial reporting is a chain of summaries, and each stage trusts the one beneath it rather than re-deriving from source.
By the time a figure is questioned it may have been repeated in three documents and acted on once. Correcting it is then a restatement rather than an edit.
Where it helps#
What is the strongest use?#
Explaining a variance. The number comes from the ledger; the question is why. A system that retrieves the relevant transactions, contracts and notes and drafts an explanation saves real time — and the explanation is checkable against the sources it cites.
What about reading invoices and receipts?#
Genuinely useful, with two conditions. A confidence threshold with human review below it. And awareness that extraction errors are silent — a transposed figure or a wrong date looks exactly like a right one.
The operational consequence: an accuracy figure measured once, at selection time, decays without telling you. Suppliers change their invoice layouts, a new scanner produces different artefacts, a currency format shifts. Pull a small random sample every month and check it by hand.
Is anomaly detection worth it?#
As a queue for humans, yes — with the reason shown, not just a score. The practical difficulty is tuning: too sensitive and the queue is ignored within a fortnight, too blunt and it finds only what a simple threshold would have caught.
Start deliberately narrow, measure how many flags turn out to be worth investigating, and widen from that number rather than from intuition.
What else is safe?#
Contract and obligation review — finding payment terms, renewal dates, penalty clauses across agreements nobody has read since signature. Drafting management commentary from figures the finance system produced. Helping structure a scenario model while the arithmetic stays in the spreadsheet.
Finding is a good fit. Concluding is not.
The real obstacle#
Why do finance AI projects stall?#
Reconciliation, almost always. The systems disagree for legitimate reasons — different cut-off times, different scope, different definitions of the same word — and resolving that is finance work, not model work.
The symptom is distinctive: the system answers correctly on Tuesday and differently on Wednesday, because it read a different source. Teams spend weeks tuning prompts against what is really a data ownership problem.
What is the right order?#
Reconcile first, automate second. The reverse is usually attempted because it demonstrates faster, and it produces numbers nobody trusts — which is worse than no automation, because trust is hard to rebuild.
See data engineering for the layer underneath.
Controls#
What would you require before a figure reached a report?#
Traceability to a system of record with the source shown. Deterministic arithmetic. Sources reconciled with disagreements surfaced rather than silently resolved. Confidence thresholds on extraction with human review below them. A named person accountable. An audit trail of what was read, produced and approved. And tested behaviour at period end, when it is under load and the deadline is real.
Does automation change who is accountable?#
No. Whoever owned the number before still owns it. Automation changes how the work is done, not who answers for it — and that should be stated explicitly, because "the system produced it" is a response that satisfies nobody.
What about audit and compliance?#
Requirements apply unchanged, and controls frameworks increasingly ask specific questions about automated processing in financial reporting. Anything touching statutory reporting needs a controls review before it starts.
Confirm your obligations with your auditors and counsel; this is general information, not advice.
Could AI replace a financial analyst?#
It shifts the ratio of time spent gathering to time spent thinking. That is a genuine improvement and a much smaller claim than the one usually made.
The part that does not move is deciding what deserves attention and saying so to people who may not want to hear it. An analyst's value concentrates in exactly the moments where being wrong is expensive and being confident is easy — which is where a probabilistic system is least suited to stand alone.