AI Answer Accuracy Is Becoming a Governance Issue for Financial Institutions

A procurement analyst at a mid-sized corporate is shortlisting banking partners. She does what most people now do before they open a tab on anyone's website: she asks a public AI assistant. "Which of these banks still offers the structured trade facility, and what does it cost?" The answer comes back in a clean paragraph, confident, well formatted, and specific on the fee. One detail is wrong. The facility she's asking about was retired eight months ago. The bank's own pages say so. The assistant does not.

She has no reason to doubt it. The answer reads exactly like the right answer would.

This is the situation a growing number of financial institutions are in without knowing it. A customer, a journalist, a regulator's staffer, a counterparty's analyst asks a public AI assistant a direct question about the institution, and gets back something fluent, plausible, and not quite true. Nobody at the institution wrote it. Nobody there can see it. And when it goes wrong, the cost still arrives at the institution's door.

AI answers travel in your name

The first thing worth getting straight is whose problem this is. The instinct in most banks and insurers is to file it under marketing, or under IT, or under "things the AI companies should fix." All three instincts are understandable and all three are wrong.

It isn't marketing's problem because the answer isn't an ad or a search result the institution can influence with copy and keywords. It isn't IT's problem because there is no system to patch; the assistant runs on someone else's infrastructure, trained on data the institution doesn't supply and can't audit. And it isn't realistically the AI provider's problem to solve case by case, because no provider can keep current on every product, price, and policy of every regulated firm in every market.

What's left is the institution itself. The answers circulate in its name. They describe its products, quote its fees, set expectations about its complaint handling and its fraud channels. To the person reading them, they are the institution speaking. The fact that the institution didn't author the words and can't edit them does not change who the customer holds responsible, and it does not change who a regulator will look to if a misled customer files a complaint.

That is the gap. Responsibility sits with the institution; control does not. Most governance failures live in exactly that kind of gap.

AI answers: a governance surface you can't see

Here is the part that makes this genuinely new, rather than an old reputational risk in fresh clothing.

Institutions already watch what's said about them. They have brand monitoring, social listening, media tracking, complaint analytics. Those tools work because the content is public — a tweet, a review, a news article, a forum post sits somewhere an analyst or a crawler can read it. You can find it, log it, respond to it.

AI answers don't behave that way. They are generated, privately, in a session between one person and one assistant, and then they're gone. There is no public post to discover. Two customers can ask the identical question an hour apart and get materially different replies, and neither reply exists anywhere you can monitor. The institution's name is being used to make specific factual claims, at scale, in conversations it has no window into.

So the surface is real, it speaks for the institution, and the existing perimeter doesn't reach it. It lives one step outside the usual line of sight — not in public where brand monitoring works, but in private sessions the institution was never set up to observe.

A useful way to hold it: this is a channel that speaks in your name, that you didn't build, can't edit, and currently can't see. Most institutions already govern every other such channel closely. This one slipped in without being named.

What AI answer errors look like

The errors aren't exotic. That's what makes them easy to miss and worth taking seriously. In its work across the sector, Lawnise has observed the same few shapes recurring.

The sharpest of them has the highest stakes. Ask where to report a suspected fraud and an assistant may name the wrong channel — an old number, a general line, a process that no longer routes the way it once did. In a fast-moving scam, the minutes a customer loses to a wrong instruction are the minutes that decide whether the money is recoverable.

The quieter shapes do their damage more slowly. A product that was discontinued months ago still gets described as available, with an assistant happily explaining how to apply for something the institution no longer sells. The right product, but quoted at the wrong price, or with a penalty or lock-in that doesn't match the current tariff — close enough to sound authoritative, off by enough to matter to someone deciding whether to borrow. A complaint or dispute timeline stated more generously than the institution's own written commitment, so a customer is told to expect a response in a window the institution never promised and may then miss through no fault of its own. And underneath all of it, a quieter tell: ask the same question of several assistants and you often get several different answers, none of them quite the institution's own. When the systems disagree, at least some of them are wrong, and the customer has no way to know which.

Our Singapore banking barometer put numbers to this. The point of this piece is the pattern beneath the numbers, which holds wherever a public assistant is asked about a regulated firm: the records can be immaculate and the representation still wrong.

How to govern AI answers you can't edit

If you can't edit the answers, what is there to govern? More than it first appears — provided the institution treats this as a managed process rather than an occasional scare.

What good looks like is a repeatable loop, run with the seriousness given to any other channel that speaks for the firm. It comes down to a handful of moves, in order.

Start by watching what the major public assistants are actually telling people about you — the real high-intent questions a customer asks before opening a product, disputing a charge, or reporting a fraud. Take each answer and check it against the institution's own published facts: the current tariff, the live policy, the published process. Where an answer drifts, work out why. A stale public source the institution can refresh is a different problem from a confident invention, and the two call for different responses.

Then rank what you find. What could genuinely mislead a customer on a price, a deadline, an eligibility rule, or a safety step belongs ahead of what is merely imprecise. And throughout, keep a dated record of what was said and what you did about it.

None of those steps requires control over the assistant. They require the institution to know what's being said in its name, to know where it diverges from the truth it does control, and to be able to show its work. That last point carries weight in a supervised industry. When a board, or a regulator, asks what the institution is doing about AI getting things wrong about it, "we don't monitor that" is a weaker answer every quarter that passes.

The governance exposure for institutions

It's worth being precise about the nature of the exposure, because overstating it helps no one. No regulator has ruled that an institution is answerable for every word a third-party model generates, and this piece doesn't claim one has. The risk is more ordinary and more durable than a single rule.

It is the risk of a customer making a decision on wrong information that carried the institution's name. It is the risk of a fair-dealing or clear-disclosure question being asked about claims the institution never made but is associated with. It is the slow erosion of trust that comes when the confident answer and the true answer keep failing to match. In markets such as Malaysia and Singapore, where supervisors hold financial firms to a high standard on disclosure, complaint handling, and treating customers fairly, those are not abstract concerns. They are the everyday texture of the obligations an institution already carries — now reaching it through a channel it hasn't yet learned to watch.

The institutions that handle this well won't be the ones that somehow force the assistants to be perfect. That isn't on offer. They'll be the ones that stopped treating AI answers as someone else's weather and started treating them as a surface they're accountable for: watched, checked, ranked, recorded.

The records were right. The representation wasn't. The difference between those two sentences is becoming a governance question, and it is one the institution, not the algorithm, will be asked to answer.