In most industries a call recording is useful. In regulated financial services it is a record — potentially evidence in a complaint, an audit or a regulatory review.
That changes what a QA programme is for. It is still about service quality, and it is also about being able to demonstrate what was said.
Mandatory disclosures are binary, and therefore easy
Financial services calls frequently require specific statements — risk warnings, fee disclosures, recording notifications, cooling-off rights. Whether a required statement was made is one of the few genuinely unambiguous QA parameters.
Which makes it a poor use of scarce human review time and an excellent use of automated checking across every call. The value is not the average score; it is being able to say what proportion of calls contained the required disclosure, and identify the ones that did not.
| Check | Type | Why coverage matters |
|---|---|---|
| Recording notification given | Binary | A single omission can void the record |
| Risk warning stated | Binary | Regulatory exposure per call, not on average |
| Fees disclosed before commitment | Binary | Central to mis-selling complaints |
| Identity verified before account detail | Binary | Fraud and privacy exposure |
| Pressure or urgency applied | Judgement | The mis-selling signal that is hardest to catch |
Mis-selling shows up in tone, not words
The compliance failure that causes real damage is rarely a missing sentence. It is an agent creating urgency that was not warranted, minimising a risk, or pressing past hesitation.
None of that is detectable in a transcript. The words can be entirely compliant while the delivery does the selling — which is why transcript-scoring systems under-report exactly the behaviour a regulator cares most about, and why analysing audio rather than text is not a technical preference here.
Fatal incidents carry more weight here
A dismissive agent in retail support costs you a customer. A dismissive agent handling someone in financial difficulty is a vulnerable-customer issue with a different consequence entirely.
So fatal-incident detection warrants a broader definition in financial services — including impatience with a distressed caller, and continuing a sale after the customer has expressed doubt. Reported as a count with human review of each, never averaged away.
Evidence, and how long you keep it
Two competing pressures, and they resolve differently for audio and for analysis.
Dispute and regulatory timelines can be long, which argues for retention. Privacy and breach exposure argue for deletion. The resolution is usually to keep the audio for the regulatory minimum and the derived scores and analysis for longer — the scores carry the reporting value at a fraction of the sensitivity.
Whatever you choose, confirm that expiry deletes from storage rather than hiding from a screen. That distinction will not survive a security questionnaire.
Practical starting set
- Every mandatory disclosure as its own binary parameter, checked on every call.
- Pressure and urgency as a scored judgement parameter, weighted heavily on sales queues.
- Identity verification as fatal-level rather than scored.
- Vulnerable-customer handling in the fatal-incident definition.
- Retention set to the regulatory minimum for audio, longer for derived analysis.
More on how banks apply this on the banking use case page.
Financial services QA has an advantage most industries lack: a large share of what matters is binary and checkable. Disclosures either happened or they did not, and that is worth verifying on every call rather than sampling.
Spend your human review time on the judgement parameter instead — pressure and urgency — because that is where mis-selling actually lives and where automated detection is least certain.