AI transcription for mortgage advisers

Speech to text is the least glamorous AI capability and the one most likely to survive in a brokerage, because it does something narrow and verifiable. It turns audio into words. Everything else you might build on top of it depends on that layer being accurate.

Which makes it worth understanding the layer itself, rather than buying whatever is bundled into a product you were already looking at.

Four different jobs, often confused

Firms tend to talk about transcription as one thing. It is at least four, with different accuracy requirements.

Dictation. You speak, it types. One voice, controlled conditions, and the output is corrected immediately by the person who spoke. The easiest case by a distance, and a genuinely good fit for advisers who think faster than they type.

Meeting capture. Two or more people, variable audio, output reviewed later. Harder, and dependent on speaker separation working properly.

Call transcription. Telephone audio is narrowband and compressed, which degrades recognition compared with a laptop microphone. Expect worse results than the vendor's demo, which will have been recorded in ideal conditions.

Bulk processing of existing recordings. Useful for compliance sampling and complaint investigation, and the one place where imperfect accuracy is least damaging, because you are searching rather than quoting.

Buying a tool that is excellent at one of these and using it for another is the most common disappointment.

Testing accuracy before you commit

Do not accept a published accuracy figure. Test on your own audio.

Take six real recordings that represent your actual range: a quiet office conversation, a call over a mobile in a car, a client with a strong regional accent, a client whose first language is not English, a three-way conversation with a partner or a family member present, and a call with background noise. Run each through the tool. Read the output against the audio.

Count the errors that would matter. A misspelled word is nothing. A wrong number, a wrong lender name, a wrong date, or a "not" that has disappeared are everything.

Pay particular attention to whether accuracy varies across the six. A tool that performs well for some speakers and poorly for others is a fairness problem as well as a quality one, and if it shapes how cases get handled you should be able to explain why that is acceptable.

Also test speaker separation. Attributing the client's statement to the adviser, or the reverse, is a serious defect in any record you might later rely on, and it happens more often than vendors advertise.

Where the audio and text go

Every recording of a client conversation is personal data, and one covering health or financial hardship will often be special category data.

Establish, in writing, before you sign anything: which company processes the audio, whether processing happens in the UK or overseas, whether recordings or transcripts are retained by the vendor after processing, and whether your data is used to train or improve their models. Get a processor agreement that reflects the answers. A vendor that resists this conversation is telling you something.

Then decide your own retention. A common and defensible pattern is to keep the transcript with the case record for as long as you keep the file, and delete the underlying audio much sooner, because the transcript serves the business purpose and the audio carries more risk. Whatever you choose, write the period down and actually enforce it. Recordings that accumulate indefinitely because deletion was nobody's job are a liability that grows quietly.

Your privacy notice needs to describe this, and if the recording captures a third party who is not your client, such as a spouse sitting in, they have rights too.

Getting the transcript into the right place

A transcript sitting in a vendor's web app is nearly useless. It needs to reach the case record, because that is where anyone will look for it in eighteen months when a query lands.

Ask specifically how the export works, whether it can be automated, and what the file looks like when it arrives. Ask what happens if you leave: can you take every transcript with you, in a readable format, without paying for the privilege. Firms discover the answer to that question at the worst possible moment.

Mistakes that recur

Treating the transcript as the note. It is raw material. A record that requires someone to read fifty minutes of speech to find the point is not a usable case file, and a file checker will say so.

Quoting from it without listening. If something in a transcript is going to be relied on in a complaint response or a dispute with a lender, verify it against the audio first.

Recording without a clear announcement. Tell people at the start of the call, in words, not in a line buried in a terms document.

Letting advisers use personal accounts. Consumer transcription apps installed on a personal phone, holding client conversations, outside any firm agreement, is a data breach waiting to be discovered. This happens more than principals expect, and usually with good intentions.

Assuming the transcript proves what you did. It proves what was said. It does not evidence that advice was suitable, that a risk was explained adequately, or that a client understood. Those still need the adviser's own record.

Where it pays

The honest case for transcription is not that it produces polished output. It is that it removes the fear of missing something while listening, which changes how advisers conduct conversations. Advisers who are not writing while a client talks ask better follow-up questions, and clients notice being looked at rather than typed at.

That benefit is real and it does not show up on a time saving calculation. It is probably the strongest reason to bother.