It can be safe, provided the tool is a business-grade service with a data-processing agreement, doesn't train on your calls, keeps recordings only as long as your record-keeping rules require, clients are told before recording starts, and a broker checks every summary before it goes on file. Free consumer note-takers and unchecked summaries are where it becomes unsafe.
Mortgage calls raise the stakes because of what's said on them. A single fact-find call can cover income, debts, missed payments, a divorce, health conditions for protection, and enough identity detail to open an account in someone's name. That changes which tools qualify, how long anything is kept and who can see it. It also means an AI summary that gets a detail wrong isn't a minor inconvenience; it can end up in a lender application or a protection form.
What's said on a mortgage advice call
It helps to be concrete about the data before judging any tool. On a typical purchase or remortgage call you'll hear most of this:
| Type of data | Example | Why it matters for AI |
|---|---|---|
| Identity | Full name, date of birth, address history | Useful to fraudsters if a transcript leaks |
| Financial | Salary, bonuses, debts, bank details, deposit source | Figures are easy for AI to mishear, and errors flow into applications |
| Credit problems | Missed payments, defaults, past arrangements | Sensitive and embarrassing; access should be tight |
| Health | Conditions, medication, smoking, for protection | Special-category data under data-protection law such as the GDPR; extra care required |
| Family and relationships | Separation, dependants, who's on the deeds | Personal; may involve third parties who never agreed to anything |
| Vulnerability | Bereavement, illness, difficulty understanding | Needs to reach the file accurately so the firm can respond properly |
Any AI tool that listens to these calls is processing all of it. The question is whether you can show it's handled at least as carefully as your existing call recordings and file notes.
Six risks, and the control for each
| Risk | How it shows up | Control |
|---|---|---|
| Vendor uses your calls to improve its models | Buried in the privacy policy, often as "de-identified" data | Business terms that exclude training, in writing |
| Recordings kept longer than needed, or not long enough | Default retention is the vendor's, not yours | Set retention to match your record-keeping policy; test deletion |
| Wrong figures or omissions in summaries | "Fifteen" becomes "fifty"; a health condition is left out | Broker checks every summary against the call before it's filed |
| Client unaware AI is listening | A note-taker bot joins a video call unannounced | Notice in advance and at the start; option to proceed without |
| Too many people can access transcripts | Shared workspace, links forwarded by email | Access limited to the case team; no public sharing links |
| Data leaves approved systems | An adviser uses a personal note-taker app on their phone | One approved tool, a written policy, and personal accounts banned for client calls |
The third row causes the most day-to-day trouble, and the first causes the most unpleasant surprises. Both are worth looking at in detail.
Reading a vendor's small print: a real example
Popular note-takers differ more in their terms than their features. Otter.ai is a useful example because its privacy policy is explicit: it says Otter trains its proprietary AI technology on de-identified audio recordings and on transcriptions, which may contain personal information. That may be acceptable to some businesses. For a mortgage firm, "transcriptions which may contain personal information" being used for training is a line your compliance support will want to discuss before any client call goes near it. Otter has also faced a class action, filed in August 2025, alleging its meeting assistant recorded participants without proper consent; those are allegations, but they show why the consent question isn't theoretical.
The meeting platforms you may already pay for have their own options. Zoom's built-in meeting summaries (the features formerly branded AI Companion, a name Zoom retired in June 2026) come with paid Zoom Workplace plans, and its separate agent product, ZoomMate, is a paid add-on from $20 a user a month. Microsoft Teams' intelligent recap requires a Teams Premium or Microsoft 365 Copilot licence, and keeps the data within your Microsoft 365 environment. Advice-specific tools built for regulated firms often offer meeting capture with audit trails and fact-find population. Whichever you consider, get written answers to these:
- Is any of our audio or transcript data used to train your models or anyone else's, including in "de-identified" form?
- Where is the data stored and processed, and which sub-processors are involved?
- Can we set our own retention period, and can we prove deletion?
- Who at the vendor can access recordings, and under what circumstances?
- Can we restrict access to named users and disable public sharing links?
- Can we export recordings and transcripts if we leave?
- Does a note-taker join visibly, and can it be stopped from joining calls automatically?
- Will you sign a data-processing agreement, and what security certifications do you hold?
What to check in an AI vendor's data processing agreement goes through the clauses behind questions 1 to 4 and 8, and whether AI meeting note-takers are safe for client calls covers the general case for any client-facing business.
Telling the client: wording that works
Clients should hear about AI before the call and again at the start. Filled-in wording an illustrative brokerage might use:
In the booking confirmation:
"We record advice calls so we have an accurate record of what
we discussed. We also use a transcription tool to help write
up our notes. Recordings and notes are kept securely for [X
years] in line with our record-keeping duties, and are never
used to train AI models. If you'd rather we didn't use the
transcription tool, just tell us and we'll take notes by hand."
At the start of the call:
"Before we start, this call is recorded, and I'm using a
transcription tool that helps me write up notes afterwards.
I check everything it produces. Is that okay with you?"
If the client says no:
"No problem. I'll switch the transcription off and take notes
myself. The call will still be recorded for our records."Adjust the last line to match your actual practice. If your firm must record advice calls, the client's choice is about the AI transcription, not the recording, and the wording should make that clear. Your compliance support should sign off the final version.
Two situations the script doesn't cover come up often enough to plan for. The first is a third person on the call. A parent gifting part of the deposit joins for ten minutes to confirm the amount and that it isn't a loan. They never received the booking confirmation, so the notice didn't reach them. The broker should repeat the start-of-call line when they join ("before you go on, this call is recorded and a transcription tool is running; is that all right?") and the checked note should record that they were told and agreed. Their name and the gift details are now their personal data on your file too.
The second is a client who wants to stop part-way: "Can you turn that off while I explain what happened with my ex?" Pause the transcription (and the recording, if your rules allow), say out loud that you've done so, and write that part of the note by hand. Then add a line such as "Transcription paused 22:10-27:40 at client's request; notes of that section taken manually" so the gap in the transcript isn't mistaken later for a missing record. Check in your mock calls that the tool can actually pause without ending the session.
Testing accuracy before you trust it
Every tool claims high accuracy. Test it on the calls you actually make. Run five mock advice calls with staff playing clients, using realistic figures and some deliberately awkward details, then compare the AI summary with the script. Illustrative results from such a test:
| Mock call | Errors in figures | Omissions | Example |
|---|---|---|---|
| 1. First-time buyers, two incomes | 1 | 0 | "Deposit of fifteen thousand" became "fifty thousand" |
| 2. Remortgage with debt consolidation | 1 | 1 | Card balance of 4,300 recorded as 3,400; payday loan not mentioned |
| 3. Self-employed, two years' accounts | 0 | 1 | Summary said "profits stable"; client said the second year was lower |
| 4. Protection needs, health questions | 0 | 1 | Medication mentioned in passing was left out |
| 5. Joint applicants talking over each other | 2 | 0 | Incomes attributed to the wrong applicant |
Five calls, eight problems, none of them unusual for speech-to-text on fast conversation. The test isn't to find a tool with no errors. It's to learn what kind of errors yours makes, so the broker's check looks in the right places: figures, attribution between joint applicants, and anything said quickly or in passing.
You can also shape the summary so those errors are easier to find. Many note-takers let you set a summary template or custom instructions; where yours does, ask for figures and sensitive points to be pulled out with timestamps rather than folded into prose:
After each call, produce:
1. FIGURES: every amount, date and term mentioned, one per line,
with the timestamp and which person said it.
2. DISCLOSURES: anything about health, credit problems, debts or
vulnerability, quoted word for word, with timestamps.
3. UNCLEAR: any word or number you're not confident about, marked
[unclear] with the timestamp.
4. ACTIONS: documents requested and next steps.
Do not assess suitability, recommend products or summarise the
client's circumstances in your own words.An illustrative extract of what came back for mock call 2:
FIGURES
05:41 Applicant 1: outstanding card balance 3,400 [unclear]
07:02 Applicant 1: car finance 212 a month, 19 months left
DISCLOSURES
09:15 "we had a payday loan last year but that's cleared"
ACTIONS
Payslips x3, bank statements x3, card statement
Suitability: client appears well suited to a five-year fixed rate.Progress: the payday loan the prose summary missed now appears, quoted, and the card balance is flagged as unclear, which sends the broker to 05:41 where the client clearly says 4,300. The last line has to go. The instructions said no suitability statements and the tool added one anyway, and a line like that in a file note reads as advice the broker never gave. If it keeps appearing, remove it every time and raise it with the vendor; don't let it become normal to see on file.
The broker's same-day check, built from the mock-call results, then fits on a card beside the screen:
- Every figure in the FIGURES list: listen to its timestamp. Deposits, balances and incomes first.
- Every [unclear] mark: listen and correct; never leave one on file.
- Joint applicants: check each income and debt is attached to the right person.
- Health and credit: scan the transcript for "well", "except", "apart from" and "tablets"; that's where passing disclosures hide.
- Self-employed income: check year-by-year figures, not the summary's description of the trend.
- Anything the tool added that nobody said (suitability, product names, "client happy with"): delete.
What a checked summary looks like
One omission on a protection call shows why the check is non-negotiable. On a protection call, the client says: "No, no health issues really. Well, I'm on tablets for my blood pressure, but that's it." The AI summary reads: "Client confirmed no health conditions." Filed unchecked, that note could feed a protection application that misstates the client's health, a non-disclosure problem that may surface only when a claim is made, years later.
The before and after, as it would appear on the file:
AI draft: "Client confirmed no health conditions. Non-smoker. Wants cover to match the mortgage term."
Broker-checked note: "Client initially said no health issues, then disclosed medication for high blood pressure (name and dosage to be confirmed before application). Non-smoker for over five years. Wants life cover to match the 30-year mortgage term; critical illness discussed, client to consider. Checked against recording, 14:32-16:05."
The second note is longer because it's accurate. The time reference to the recording also makes any later review quicker. Summaries like this should be checked the same day, while the call is fresh; a queue of unchecked summaries at the end of the week is where omissions slip through.
It's fair to ask whether the check eats the time saved. A quick sum for an illustrative broker running twelve advice calls a week: writing a file note by hand from memory and scribbles takes about 25 minutes a call, or five hours a week. Checking a structured AI summary against timestamps takes perhaps 10 minutes, or two hours. That's three hours back, and a note that's more accurate than the handwritten one, because the check is against the recording rather than recollection. If your checks are running close to 25 minutes, the summary format is wrong or the tool is making too many errors; either is worth knowing before you roll it out to every adviser.
For more on meeting records in advice work, AI note takers for financial advisers covers the compliance side of file notes in depth.
Where recordings and transcripts live afterwards
Most of the long-term risk sits in storage, not the call itself. Decide these in advance:
- One home for records. The checked note goes into your CRM or case file. If the tool keeps its own copies, know how long, and either align that with your policy or delete them after export.
- Retention by rule, not by default. Your record-keeping obligations set how long advice records must be kept; data-protection law expects you not to keep personal data longer than needed. Set the tool's retention to match and check that deletion actually happens.
- Access by case. Only the people working on a case should be able to open its recordings and transcripts.
- Requests from clients. If a client asks for a copy of their data, you'll need to find their recordings and transcripts quickly. Test that you can.
A go/no-go check before the first real call
Put this in front of whoever approves the tool. Every line should be a yes:
- Business plan with a signed data-processing agreement, and no training on our data in any form.
- Retention set to our policy; deletion tested.
- Access limited to named users; public links disabled.
- Client notice in booking confirmations and a start-of-call script approved by compliance.
- Five mock calls tested; the broker checklist covers the error types we found.
- Every summary is checked by the broker the same day before it's filed.
- Personal note-taker apps banned for client calls, in writing.
If a line is a no, fix that first. The rest of the case, from sourcing to updates, has its own automation questions; what mortgage brokers can automate and what stays advice maps them out, and what an AI note-taker costs per seat helps with the budget line.
Spot-checking the file notes three months in
A process that works in week one can slip by month three, usually because checks get quicker as brokers start trusting the tool. Once a month, have someone other than the adviser pick three checked notes per broker at random, listen to the recording around each figure and disclosure, and count differences. It takes about an hour for a four-adviser firm.
An illustrative first audit: eleven of twelve notes matched their recordings. The twelfth, from a busy Friday, still carried an [unclear] mark on a monthly childcare cost and a "client happy with the proposed approach" line the tool had added. Neither made it into an application, but both showed the Friday check had been a skim. The response was procedural rather than a telling-off: notes can't be filed while an [unclear] mark remains, and calls booked after 4pm on Fridays are checked first thing Monday, before anything else. Record each audit's result, because a run of clean months is also the evidence your compliance support will want if anyone asks how the firm knows the notes are reliable.
AI on mortgage calls: questions brokers ask
Do we need the client's consent, or is telling them enough?
It depends on your market's rules and on the legal basis your firm relies on for processing, so ask your data-protection adviser or compliance support. In practice, most firms tell clients in advance, repeat it at the start of the call, and offer a way to proceed without AI transcription. Whatever the basis, the client should never discover afterwards that an AI tool was listening.
Can we use the AI summary as our file note, or do we still need our own?
The checked summary can become the file note, provided the broker has read it against the call, corrected it and signed it off. The unchecked AI draft should not be the record. Some firms keep the original transcript as well, so the file shows both what was said and the broker's summary of it; decide which you keep and for how long as part of your retention policy.
Is a transcript safer than a recording?
Not automatically. A transcript is easier to search and share, which makes it more useful and easier to leak. It also contains the same sensitive details as the audio. Treat both the same way: stored in approved systems, with access limited to the people working on the case, and deleted on your retention schedule.
Further reads
- AI Meeting Note-Takers Compared for Small Teams — How the main note-takers compare on features and data.
- Can Lawyers Use AI Note Takers in Client Meetings? — How another confidential profession handles the same question.
- Can AI Keep Mortgage Clients Updated During an Application? — The next stage of the case: keeping clients informed.
- How to Anonymise Client Data Before You Paste It Into AI — Test tools on dummy calls rather than real client data.
- SOC 2 and ISO 27001 Explained: Checking an AI Vendor's Security — What security certifications actually tell you about a vendor.
- GDPR and AI Tools: What a Small Business Must Do — The data-protection basics behind any AI tool decision.
- Choosing a Mortgage CRM With AI Built In: What Actually Matters — A demo-ready checklist for mortgage CRMs with AI: document extraction, chasing, client-bank alerts, compliance logs, integrations and exit terms.
- From Enquiry to Fact Find: An AI Workflow for Mortgage Brokers — Five hand-offs between a new mortgage enquiry and the fact-find meeting, with prompts, sample outputs and the checks that stop AI misreading payslips.
- Best AI Tools for Mortgage Brokers at Each Stage of a Case — Eight stages of a mortgage case, the AI tool worth using at each, what it costs at list price, a real-looking example of it working, and when to skip it.
- Using AI to Spot Remortgage Opportunities in Your Client Bank — Turn an untidy client bank into a watchlist of deal-end dates, early-repayment-charge dates and loan-to-value moves, with AI filling the gaps.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Otter.ai privacy policy (training on de-identified audio and transcriptions) and pricing; Zoom newsroom and product pages on ZoomMate and the retired AI Companion name (June 2026); Microsoft Learn, intelligent recap licensing. All checked September 2026. Not legal advice.