Financial services call QA: quality vs compliance

Short answer

A bank or insurance call raises three different questions, and they must not be merged into one score: how good was the service, were the mandatory disclosures made in full, and is there any sign of a potential violation? The first serves the agent's development, the second the fulfilment of regulatory and internal rules, and the third is a risk signal that needs human review. When "average score 84" combines all three, an interest rate that was never stated hides behind good manners.

An important boundary: an AI evaluation is not a legal verdict. A "violation flag" only points to a call a person needs to look at; whether a violation happened is decided by the company's compliance function.

Why three separate results

1. 1Service qualityClarifying the need, a clear answer, a next step. Used for agent coaching and meaningful as an average.
2. 2Mandatory disclosuresTexts approved by the company: interest rate, fees, insurance exclusions, the right to withdraw. An average means nothing here — it was either said or not.
3. 3Potential-violation flagA prohibited phrase, a guarantee, an action without the customer's consent. Not a score, but a signal for human review.

AWS's evaluation-form documentation describes marking an answer option as an "automatic fail", which sets the score of a section or the whole form to zero. That is one platform's approach. Whichever you choose, disclosures and flags should appear in the report separately from the service score.

How to write mandatory disclosures

• The list of disclosures for each product is approved by compliance and kept with a date
• If the text must be read in full, that is stated; if conveying the meaning is enough, the key elements are listed — for example, annual rate, fee, term
• For each disclosure, the call types it applies to are written down — a loan disclosure does not apply to a balance enquiry
• An approved version exists in each language
• The point in the call where the disclosure must be made is noted — for example, before consent

Illustrative example: one call, three results

The example below is illustrative, not a real bank call. The customer calls about a consumer loan.

Turn these three results into one number and the call looks "good". Show them separately and it is clear: the agent is good at service, but one mandatory disclosure was missed and an unconfirmed promise was made. The first is a coaching topic; the second and third are cases for compliance.

Illustrative disclosure list: a consumer loan

The list below is an illustrative example; the real list is approved by your bank's compliance function. Each item becomes a separate step in the call standard.

1. 1Annual interest rateThe figure, and whether it is fixed or variable.
2. 2Fees and extra chargesCharges for issuing the loan, maintaining the account and any insurance.
3. 3Term and monthly paymentAn estimated monthly payment for the chosen amount and term — with the word "estimated".
4. 4Early repayment termsWhether early repayment carries a penalty or fee.
5. 5Condition of approvalThat the decision follows the bank's checks; that no guarantee is given.

When a disclosure is missed: agent or script?

If several agents miss disclosures at the same point, the cause is often not the agents. The script is too long, the disclosure sits at the end of the call after the customer has already said "fine, thanks", or the text is not in spoken language and agents shorten it. In that case the fix goes to the script: move the disclosure to a natural moment in the call and simplify the wording — with compliance's approval. Lowering agents' scores does not solve the problem.

From flag to decision: human review

1. 1FlagThe system or a QA specialist points to a potential violation: which criterion, which transcript line.
2. 2Listening to the audioA compliance officer listens to the audio itself, not the transcript — ruling out recognition and speaker errors.
3. 3ContextWhat happened before and after the call: was written information sent, what does the contract say?
4. 4Decision and recordViolation confirmed, not confirmed, or unclear. The decision and its reasoning are recorded.
5. 5The agent's sayThe agent can explain and dispute; the decision does not rest on the AI result alone.

Insurance-specific criteria

• Was it stated what the policy covers and what it does not — the exclusions?
• Were the customer's questions about past illness or damage answered correctly; did the agent make an unconfirmed promise such as "that will be covered"?
• Were the documents and deadlines needed for a claim stated correctly?
• Was the right to withdraw and its deadline stated, where the rules require it?

Bank-specific criteria

• Was the customer's identity verified per the bank's rules — without asking for more than needed?
• The agent did not ask for a full card number, PIN or SMS code over the phone
• Were interest, fees and penalties stated correctly and in full?
• No guarantee of loan approval was given
• No transaction was carried out without the customer's explicit consent

Three columns in the report

• Service score — averaged by agent and team, split by call type
• Disclosures — separately for each disclosure: on how many applicable calls it was made, and on how many not
• Flags — how many were raised, how many reviewed, how many confirmed, how many rejected
• The three columns do not mix: a disclosure result does not lower or raise the service score

This structure lets the same data serve both agent development and compliance reporting, while asking each its own question.

Typical mistakes

• Averaging disclosures into the same score as service criteria
• Reporting a flag as a confirmed violation
• Keeping the disclosure list without dates — an old call gets judged by a new rule
• Counting a disclosure that does not apply — on a balance enquiry, say — as "no"
• Making a compliance decision without explaining it to the agent

Limits

• An AI result is not a legal verdict and does not automatically ensure compliance with any regulation
• A transcript does not confirm that written documents were sent to the customer or that a transaction was carried out
• Disclosure requirements vary by country, product and regulator, and must be checked with compliance
• Keeping financial data in call recordings brings its own risks and requirements

What Vexvon Audio Analyzer offers

• You write service criteria and disclosures into your call standard as separate steps
• Each step is judged as met, partial, missed or not applicable and shown with transcript lines — so the results of the disclosure steps can be counted separately
• Vexvon computes one 0–100 score per standard; there are no criterion weights or automatic zeroing — the three separate results are built by your reporting rule
• Risky phrases appear as transcript lines, so a compliance officer finds them directly

More: Vexvon Audio Analyzer. How critical errors affect the score is covered in critical errors and agent scores, the agent appeal process in call center QA appeals, and protecting personal data in call recording data privacy. Rules for bank campaign calls are in bank call compliance.

First step

Pick one product — a consumer loan, say — and write its mandatory disclosures with compliance as a dated list. Then compute the three results separately on 10 calls. This article belongs to the agent QA by industry section. To try it on your own recordings, get in touch.