AI CRM for mortgage brokers: what is useful and what is risky

Every CRM sold into the UK intermediary market now has an AI story. Some of it is substantial, some is a summarise button, and the difference is not visible from a feature list. What makes AI inside a CRM a distinct question — rather than just another AI tool — is what the CRM already holds: your entire client base, your fact-finds, your bank statements, your file notes and your advice records.

That concentration is what makes the capability useful. It is also what makes the questions sharper than they would be for a standalone assistant you paste text into.

What you will actually be offered

The AI features appearing in broker platforms cluster into a handful of shapes.

• Turning a conversation or a set of answers into a structured fact-find or case note.
• Summarising a long case history for whoever picks it up next.
• Classifying uploaded documents and flagging what appears to be missing or out of date.
• Drafting client emails and progress updates from case data.
• Producing first drafts of suitability wording.
• Suggesting a next action, a review opportunity or a protection gap.
• Extracting figures from statements and payslips into fields.

The first four are administrative and are where most firms should start. The last three touch advice, judgement or accuracy in ways that need more care.

The useful test

Ask what happens if this feature is wrong and nobody notices.

A miscategorised document costs a minute. A summary that omits something the client said can mislead the adviser who reads it instead of the file. A suitability paragraph that describes a rationale the adviser did not hold is a defective record. A next-best-action prompt that pushes a client towards a product because the pattern matched is edging into territory a qualified person is supposed to own.

Sort the features by that consequence and set your approval requirements accordingly. Cheap mistakes can be corrected in flight. Expensive ones need a human sign-off that actually happens.

Questions to put to a CRM vendor

• What client data is sent to a model, and is any of it sent outside your own infrastructure?
• Is client data used to train models, by default or at all, and can that be turned off in the contract?
• Where is processing carried out and which sub-processors are involved?
• Are prompts and outputs logged, and can a reviewer see them later?
• Is the approved version stored on the case record, distinguishable from the draft?
• Can an adviser be required to approve before anything reaches a client?
• Can the firm disable individual features rather than the whole capability?
• What happens when the model has insufficient information — does it flag, or fill the gap?

If the answers are vague, the feature is not ready for a regulated advice file. That is a reasonable thing to say to a supplier.

Where the regulatory weight sits

There is no separate AI rulebook to comply with. The FCA's approach has been to apply existing obligations rather than write a new set, which means the responsibility lands where it already was: on the firm, through senior management arrangements, systems and controls, the standards for communications, the requirements around suitability, and Consumer Duty.

Practically, that means you cannot point at a supplier when an output goes wrong. If a client receives a misleading communication, it is your communication. If a file cannot be defended, it is your file.

Data protection sits alongside it. A CRM processing client personal data through AI features is doing so under UK GDPR, and the ICO's guidance on AI and data protection is the relevant reference point. Know your lawful basis, your retention position and what your privacy information tells clients about how their data is used.

An operating model that works

Treat AI in the CRM as a capable assistant with no accountability. It can remove the blank page, gather what is known, spot an omission and draft the routine. It cannot hold responsibility for advice, for suitability or for what a client understood.

In practice that means three rules. Anything client-facing is approved by a named person before it leaves. Anything that forms part of the advice record is checked against the source rather than accepted because it reads well. And anything the system was unsure about is surfaced rather than smoothed over.

Deciding whether to pay for it

Do not buy a CRM for its AI. Buy it for the things that will still matter in three years — the record structure, the workflow, the oversight, the integrations — and treat the AI as an upside that may or may not mature.

Then, if you do adopt the features, measure them. Track how often output is corrected and how long the corrections take. A feature that saves eight minutes and costs five minutes of checking is worth having; one that costs six is worth switching off. You will only know which is which if somebody is counting.