AI risks for mortgage advice firms

Most risk conversations about AI in advice firms circle the same two topics: client data and inaccurate output. Both are real. Neither is the one that catches firms out, because both are obvious enough that someone raises them early.

The risks that actually bite are duller. They are about who is using what without telling anyone, what happens when a supplier changes something, and whether your insurance and your permissions still line up with how you now work.

Here is a fuller register, with the controls that go with each entry.

Shadow usage

The most common exposure in intermediary firms today is not a badly chosen tool. It is an unchosen one.

An adviser installs a note-taking app on a personal phone. An administrator pastes a client's email into a free chat account to make it sound better. Someone uses a browser extension that summarises documents and, incidentally, transmits them. None of this is malicious and all of it is invisible until something goes wrong.

The control is not a prohibition nobody enforces. It is asking, without penalty, what people are already using, then approving a small number of tools that do the job properly so there is no reason to go around the process. Firms that lead with discipline get concealment. Firms that lead with a usable alternative get disclosure.

Follow it with a written list of approved systems, reviewed quarterly, and a named person who has to sign off additions.

Supplier dependence and sudden change

Software that thinks is less stable than software that files. Vendors update models, change behaviour, alter pricing, get acquired, and occasionally shut down. A tool that behaved one way in March can behave differently in June without anyone telling you.

That matters more when the output goes into client files. If your note quality quietly changes because an underlying model was swapped, you will find out at the next file check rather than at the moment of change.

Controls: keep periodic samples of output so you can tell when quality shifts. Ask suppliers what notice they give of material changes. Establish how you would export everything and continue operating if the service stopped tomorrow, and confirm the format of that export before you need it. Where a tool has become load-bearing, make sure at least two people in the firm know how to run the manual version of the process.

Deskilling

This is a slow risk and it is genuinely underrated.

If new advisers learn the job in an environment where the notes write themselves and the explanations are drafted for them, they may not develop the habit of thinking a case through in writing. Writing is how most advisers work out what they actually think about a case, and a firm that removes it removes some of its own training.

The control is deliberately keeping some work manual for less experienced staff, and treating drafted output as something they must be able to critique rather than accept. If an adviser cannot explain why a drafted rationale is right, they should not be sending it.

Over-reliance in the moment

Related but different. An experienced adviser under time pressure approves a summary because it looks right and the day is long.

You cannot fix this with a policy that says be careful. You fix it structurally: by making review require an action rather than a tick, by sampling approved output independently, and by not setting productivity expectations that assume the review takes no time.

If your capacity planning has already banked the time saving, the review will be the thing that gets squeezed.

Insurance and permissions

Two questions worth putting in writing to the relevant parties.

Ask your professional indemnity insurer how they view AI-assisted work in the advice process. Policies vary, disclosure obligations vary, and the moment to discover a problem is not after a claim. Answer honestly on renewal forms.

If you are an appointed representative, ask your principal before deploying anything that touches client files or client communications. Networks are increasingly setting expectations here, and a principal that discovers your usage during a supervisory visit will respond less generously than one you consulted.

Directly authorised firms should record the decision at governance level, with a named senior individual accountable for it, in the same way they would record any other material change to how the firm operates.

Data protection, briefly but properly

Covered widely elsewhere, so the short version. You remain the controller. Every vendor is a processor and needs an agreement that says so. You need to know where processing happens, what is retained, and whether inputs feed model training. Where processing is likely to be high risk, the ICO expects a data protection impact assessment, and bulk handling of financial and health data usually qualifies.

The specific control most often missing is deletion. Transcripts, drafts, chat histories and uploaded documents accumulate in tools nobody has assigned a retention period to. Assign one and make it someone's job.

Marketing claims about your own use of it

A smaller risk, and an easy one to trip over. Telling clients your service is powered by AI, or implying an automated system delivers something it does not, is a communication that has to be fair, clear and not misleading like any other.

Be plain about what is automated and what is not, particularly if a client might otherwise believe an adviser reviewed something that no adviser saw.

Making the register real

A list like this is worthless unless it has three columns filled in: what could go wrong, what we do about it, and who owns it.

Review it when something changes rather than annually, because the technology is moving faster than most governance cycles. And be honest in it. A register that says all outputs are reviewed by an adviser, in a firm where that is aspirational, is worse than no register at all, because it documents a control you are not operating.