AI for Financial Services: What Your Compliance Team Will Ask

The AI project that fails at the compliance review rarely fails because the answer was no. It fails because nobody could produce a clear answer to a reasonable question, and the review stalls while people go away to find out.

These are the questions compliance and risk functions in financial services actually ask, in roughly the order they ask them, and what a good answer looks like. Have these ready and the review becomes a conversation rather than an obstacle.

Where does client data go?

The first question, and the one that ends most reviews early.

What they need is not reassurance, it is a diagram: which components process client data, where each one runs, and what crosses a boundary. If the answer involves client records being transmitted to a third-party model provider, expect the conversation to stop there — and that is a reasonable place to stop.

A good answer identifies every point where data moves, names the jurisdiction of each processing location, and shows that sensitive records never leave your controlled environment. Architectures that keep data inside the perimeter exist, so an answer that treats external transmission as unavoidable is describing a product choice.

Can the system invent an answer?

Compliance teams have read about hallucination and will ask directly. The wrong response is to claim it cannot happen — that is not credible, and it signals the controls have not been thought through.

A good answer describes the controls: answers grounded in retrieved documents rather than model memory, citation on every factual claim, and explicit refusal when nothing relevant is retrieved. Then it demonstrates the refusal, because a system that declines to answer when it should is far more convincing than one that always has something to say. We have written about why this is an architectural decision rather than a feature.

Who can see what?

An assistant that reads across your systems creates an obvious risk: a user receiving information they are not entitled to, assembled from sources they could not individually access.

A good answer is that the system inherits entitlements from the source systems and applies them at retrieval — before the model sees anything — rather than filtering output afterwards. Post-hoc filtering is a weaker control and competent reviewers know it.

Expect a test. Someone will ask the assistant about a portfolio they do not cover.

What is logged, and for how long?

The supervisory question. If a banker acts on something the assistant said, can you reconstruct what it said and on what basis?

A good answer covers the question, the answer, the retrieved sources, the user, the timestamp, and the retention period — with the logs queryable rather than merely stored. "We have logs" is not sufficient when the follow-up is "show me every answer given about this client last quarter".

Who is accountable for the output?

Compliance is establishing that nobody believes the system makes decisions.

A good answer states plainly that the assistant supports a person who remains accountable, and shows how the interface reinforces this — sources visible, output presented as draft rather than conclusion, and no step where an AI output reaches a client without review. Under FinSA, MiFID and equivalent regimes, accountability sits with a person, and the system should make that obvious rather than merely true.

How do you know it still works?

The question that separates teams who have run AI in production from teams who have not. Models change, data changes, and behaviour drifts.

A good answer describes ongoing evaluation: a test set of questions with known answers, run on a schedule, with results tracked over time and a defined response when quality degrades. "We tested it before launch" is a weak answer for a system whose behaviour is not fixed.

What happens when it fails?

Not whether it fails. Reviewers want the failure modes enumerated and the response defined for each.

A good answer distinguishes between failing loudly — unavailable, refusing to answer, flagging low confidence — and failing quietly, which is the dangerous case. It explains how a wrong answer would be detected, by whom, and how quickly, and what the rollback path is.

What does the vendor see?

For bought solutions, this question is about the supplier rather than the software.

A good answer covers which vendor personnel can access your environment, under what controls, whether access is logged, whether your data is used for any training or improvement, and what happens to it on termination. The training question should have an unambiguous no.

Preparing for the review properly

Three things shorten these reviews more than anything else.

Bring compliance in before procurement, not after. A team involved at design asks the same questions, but asks them when the answers can still change the design.

Bring a data flow diagram. Most of the questions above are answered by one clear picture, and its absence is read — usually correctly — as a sign that nobody has mapped it.

Demonstrate a refusal. Show the system declining to answer something it cannot ground. Nothing else builds confidence as quickly.

Common questions

Do we need a formal AI policy before deploying? You need defined accountability, logging and review. A standalone policy document helps but is not what makes a deployment defensible.

Will compliance accept a cloud-hosted model? Frequently yes, provided client data does not reach it. The distinction that matters is not where the model runs but what it is sent.

How long does a review take? Weeks when the answers above are prepared; months when they are assembled during the review itself.

What is the most common reason for rejection? An unclear answer to the first question. Everything else is negotiable; data leaving the perimeter usually is not. Our note on why data control decides everything covers the reasoning.

Put these to us

If you are preparing a review, the fastest way to test a vendor is to ask all eight questions in one session and see which ones produce a precise answer. Book a demo and bring your compliance team.