The moment you record a client meeting, you've created a record you have to keep, produce on request and stand behind. That's the part of voice AI nobody demos. Speech to text has been good enough for years, so the useful questions are what the transcript becomes, who knew it was being made, and whether the numbers inside it were heard correctly. Voice earns its place in an advice firm when it removes typing at the moment of the conversation and drives the work that follows. It doesn't earn its place by producing a longer document nobody reads.
Key takeaways
- Transcription is solved. What the transcript turns into is the product.
- A recording is a record: retention, access requests and disclosure all apply from the moment it exists.
- Voice is weakest exactly where advice is most exposed: numbers, names and crosstalk.
- The strongest use isn't note taking, it's capture straight into structured fields and hands-free execution.
- Test it on your accents, your product names and your worst meeting room, not on a demo recording.
What is voice AI in financial services?
3 different things get sold under 1 label, and they're worth separating.
1. Transcription. Turning speech into text. Mature, commoditised, and by itself it produces a long document that still needs reading.
2. Structured capture. Turning speech into fields: this figure is the client's income, this sentence is an objective, this is a stated attitude to risk. Harder, and where the value sits, because a populated fact find is worth more than 40 pages of transcript.
3. Voice as a control surface. Asking the system to do something and having it done, including moving around the platform on your behalf. This is the newest of the 3 and the one that changes how a day feels, because it takes the navigation out of the work.
Most tools sold as voice AI stop at 1, describe themselves in terms of 2, and don't attempt 3.
Where voice genuinely helps
In the meeting. The best moment to capture a client detail is when they say it. Typing during a conversation damages the conversation, and writing it up afterwards means writing it from memory, which is where detail is lost and where files diverge from what was said.
Between tasks. Dictating an instruction is faster than finding the screen that performs it. For a principal moving between a client call, a provider chase and a review, the saved navigation adds up more than the saved typing.
On the move. Advisers spend real time driving between clients. Capturing an action while it's fresh beats a note that says "call Sarah" and no longer means anything by Thursday.
What voice is bad at
Be specific about the failure modes, because they're predictable.
- Numbers. "Fifty thousand" and "fifteen thousand" are 1 vowel apart. In an advice context, a misheard figure is not a typo, it's a wrong recommendation. Numbers captured by voice need confirming on screen, always.
- Names and product names. Provider names, fund names and client surnames are exactly the vocabulary a general model has least of. A system that lets you load your own terms will do materially better than one that doesn't.
- Crosstalk. 2 people talking over each other, a joint meeting with a couple, a speakerphone in a hard room. Attribution breaks first, and a transcript that assigns a statement to the wrong person is worse than no transcript.
- Accents and speech differences. Recognition quality varies across accents and speech patterns. Test with the people who'll use it and the clients you serve, not with the vendor's sample.
- Inference from silence. A model asked to fill a structured field will often produce something rather than nothing. What it should do is leave it blank and tell you.
The part nobody demos: a recording is a record
The moment a meeting is recorded or transcribed, your firm holds personal data about that client, and probably special category data if health or family circumstances came up.
That brings obligations you have to answer before the first meeting, not after:
- A lawful basis, and telling people. Clients need to know the recording exists and what happens to it. Put it in your privacy notice and say it at the start of the meeting.
- Retention. How long you keep recordings and transcripts, decided deliberately. Keeping everything forever because storage is cheap isn't a policy.
- Access requests. A client can ask for what you hold. That now includes the transcript of the meeting where they said something they'd rather they hadn't.
- Disclosure. In a complaint, the recording is evidence. Sometimes that protects you. Sometimes it doesn't. Either way it exists.
- Where the audio goes. Which provider processes it, in which country, and whether it's used to train shared models. Get it in writing, because you need it for your own records.
Separately, if your firm has telephone taping obligations, a meeting recorder isn't the same thing and doesn't discharge them. Check what applies to your permissions rather than assuming a new tool covers it.
None of this is an argument against recording. Recorded meetings, handled properly, are usually better evidence than a note written from memory 2 days later. It's an argument for deciding the policy before the tool arrives, because the tool won't ask.
From transcript to structured data
This is the difference between a novelty and a system.
A transcript is unstructured, which means someone still reads it and types the important parts somewhere useful. That's the same work as before, with an extra document to store. The version that changes a firm's week takes what was said and populates the client record: the income figure into the income field, the stated objective into objectives, the concern about a parent's care costs into the file where it will surface at the next review.
Judge any voice tool on what exists when the meeting ends. If the answer is "a transcript", you've bought a recorder. If the answer is "a populated fact find you check and approve", you've bought capacity.
What to test before you buy
- Your worst room. The one with the hard walls, or the client who prefers speakerphone.
- Your accents. Everyone who will use it, not a sample.
- Your vocabulary. 10 provider and fund names you use every week. Count what comes back wrong.
- Numbers under pressure. Say 6 figures at conversational speed and check every one.
- A joint meeting. 2 people, some overlap. Look at attribution, not only the words.
- The gap case. Ask about something the client never mentioned and see whether the field comes back empty or invented.
What voice should never do
Execute something irreversible without a person confirming it. Voice is a fast, low-friction interface, and low friction is the point right up until the action sends a letter to a provider or a report to a client. A spoken instruction should assemble the work and hold it for sign-off. The confirmation step isn't a limitation of the technology, it's the design.
Frequently asked questions
Can financial advisers record client meetings with AI?
Yes, with the right handling. You need a lawful basis under UK GDPR, the client needs to be told the recording is happening and what you'll do with it, and you need a retention policy the firm follows. Check where the audio is processed and whether the provider uses it to train shared models, because you'll be asked and you should have the answer written down.
Is voice AI accurate enough for financial advice?
For general speech, yes. For the parts that matter most in advice, treat it as a first pass. Numbers, provider names and fund names are the weak spots, and anything captured by voice into a structured field should be confirmed on screen before it reaches a document. Accuracy also varies by accent and by room, which is why you test with your own people.
What's the difference between AI meeting notes and voice AI?
Meeting notes are an output: a summary you read and then act on. Voice AI in the fuller sense is an input method: what was said becomes structured data in the client record and, at the far end, instructions the system carries out. The first saves you writing up. The second removes the re-keying that follows the writing up.
Do AI meeting recordings count as client records?
Treat them as records. If a recording or transcript informed the advice, it's part of the story of how that advice was reached, and it's disclosable in a complaint. Decide deliberately what you keep, for how long, and where, rather than letting a tool's default retention make that decision for your firm.
Should clients be told a meeting is being recorded?
Yes, and it's better practice to say it out loud at the start as well as covering it in your privacy notice. It takes 1 sentence, most clients don't mind, and a client discovering later that they were recorded is a trust problem you never needed to have.