AI calling for mortgage brokers

Synthetic voice on the phone has improved enough that a caller may not immediately notice. That is the interesting part and also the uncomfortable part, and for a regulated intermediary the uncomfortable part deserves more attention than the demo reel.

This piece sets out what these systems genuinely do well, where they break, and what a brokerage has to settle before letting one speak to anybody.

What the technology is doing

A voice agent is three things stitched together: speech recognition turning what the caller says into text, a language model deciding what to say back, and speech synthesis reading it out. Each layer adds delay, and each layer adds a chance of error.

That architecture explains almost every strength and weakness you will encounter. It is excellent at gathering a handful of known fields in a predictable order. It is poor at anything that depends on hearing something unusual and reacting to it, because the model is generating a plausible continuation rather than understanding a situation.

Where it holds up

Short, structured, low-stakes exchanges. Confirming who is calling and what they want. Collecting a name, a callback number, a rough loan size and a property type. Booking an appointment into a diary with real availability. Telling someone your opening hours or that their adviser is with a client.

Outbound, the equivalent is narrow too: confirming an appointment is still wanted, checking whether a document has been sent, asking whether a caller who left a message still needs a call back. Errands with one question in them.

The reason these work is that success is easy to define and easy to verify afterwards. If the diary entry exists and the number is correct, it worked.

Where it does not hold up

Accents and speech patterns are the first practical problem. Recognition accuracy varies with regional accents, with speakers whose first language is not English, with speech differences, and with background noise. A system that performs beautifully in a quiet office demo will perform noticeably worse with someone calling from a building site. In a market where you have obligations to all customers equally, a tool that works less well for some groups is not merely a technical annoyance.

The second problem is emotional register. Property transactions produce distressed calls. Bereavement, divorce, redundancy, a chain collapsing three days before exchange. A voice agent has no reliable way of recognising that the person on the line is in difficulty, and Consumer Duty expects your firm to identify and respond to characteristics of vulnerability. Anything that stands between a distressed caller and a human being is a genuine risk, not a theoretical one.

The third is the tendency to answer questions it should not answer. Ask a general-purpose model whether you can get a mortgage on a maisonette above a takeaway, and it will say something. What it says may be wrong, and if it came from your phone number it is your firm speaking. Any agent you deploy needs hard boundaries and a fast route to "I will get an adviser to call you about that", not clever answers.

The compliance questions to settle first

Disclosure. Tell people they are talking to an automated system, promptly and in plain words. Beyond the fairness principle in UK GDPR and the FCA's expectation that communications are clear and not misleading, there is a simpler reason: people who find out later feel deceived, and that is the sort of thing that turns into a complaint.

Recording and retention. If calls are recorded and transcribed, that is personal data. You need a lawful basis, a retention period, and somewhere in your privacy notice that says so. Note also that a transcript of a conversation about a client's finances belongs on the case record, not solely in a vendor's dashboard.

Where the audio goes. Ask the supplier directly which sub-processors handle the audio, in which country the processing happens, and whether recordings are used to improve their models. Get it in the contract rather than in an email from a salesperson.

Escalation. Define what triggers a handover to a person: an explicit request, a complaint word, any mention of distress, any question about advice, and repeated recognition failures. Then test that the handover actually works, including when your office is closed.

Complaints. If a caller expresses dissatisfaction, DISP obligations start running whether the ear that heard it was human or not. Your agent needs to capture and flag that, not soothe it away.

A realistic trial

Do not start with inbound calls from strangers. Start somewhere you already know the outcome.

Outbound appointment confirmations to existing clients is a sensible first test, because the recipients know your firm, the conversation is short, and you can measure success precisely: did the appointment get confirmed, and did anyone complain.

Listen to the recordings yourself. Not a sample summary from the vendor's reporting screen, the actual audio, including the calls the system marked as successful. Firms are routinely surprised by how often a call marked complete involved the caller repeating themselves four times.

Track the calls that ended without a resolution and work out why. That set is the real measure of the system, and it is the one that vendor dashboards tend to present least prominently.

The honest summary

Voice AI is a decent way to stop calls going unanswered and to take the tedium out of confirming things. It is not a receptionist, it is not an adviser, and it should never be the only thing standing between a worried client and a person who can help.

If you deploy it, deploy it narrowly, tell people what it is, keep a human within reach at every moment, and review what it actually said rather than what it was supposed to say.