How to review sales calls for conduct risk with an AI agent

Sampling a handful of calls per adviser each month misses most of what happens. The agent can review every sales call transcript against your conduct standards: disclosures given, needs explored, pressure avoided, vulnerability recognised, the customer's understanding checked. It drafts findings with quotes and timestamps. Quality assurance confirms them, and managers decide coaching, remediation or customer contact.

What is the process, and who owns it?

Sales calls are where conduct risk becomes customer harm: a product sold that does not meet the customer's needs, a risk not explained, a cooling-off right not mentioned, a vulnerable customer pushed into a decision. What conduct risk is sets out the concept. Regulators expect firms to monitor sales for these outcomes, and in the UK the FCA Consumer Duty puts good customer outcomes at the centre of that monitoring.

Sales quality assurance, often in the first line, reviews calls; compliance monitoring tests whether QA works; sales managers act on findings. The traditional model reviews a small sample by hand. The agent changes the coverage, not the ownership. Whether an AI agent can review customer calls for conduct risk sets out the boundary.

What does the agent prepare, and what does a person approve?

1. Person, QA lead: sets the standard. A call assessment framework per product: required disclosures, needs questions, suitability checks, prohibited statements, vulnerability indicators and signs of pressure.
2. Agent, depth 1: reads every transcript from the call recording system, with the product sold and the outcome.
3. Agent, depth 1: assesses each call against the framework. Each item met, not met or unclear, with the quote and timestamp that support it.
4. Agent, depth 1: flags vulnerability and pressure. Statements about health, bereavement, money worries or confusion; repeated pushing after hesitation; urgency the product does not justify.
5. Agent, depth 2: drafts findings for calls that fail any critical item, ranked by potential customer harm.
6. Person, QA reviewer: listens to the flagged calls and confirms or rejects each finding. Transcripts miss tone; the reviewer must hear it.
7. Person, sales manager: decides adviser outcomes, such as coaching or supervision.
8. Person, compliance or product owner: decides customer remediation where a sale may have caused harm.
9. Agent, depth 1: reports patterns. By adviser, product, script, time of month and incentive period.

Which controls apply, and what evidence does each leave?

Full coverage, human confirmation. Every call assessed; every adverse finding confirmed by a person who listened. Evidence: assessment for every call, reviewer name on every confirmed finding.

Quotes and timestamps. Findings point to the moment in the call. Evidence: reviewers and advisers can check exactly what was said.

Fair process for advisers. An adviser's outcome is decided by a manager on confirmed findings, not on agent output. Evidence: the confirmation and the decision, kept together.

Remediation decided by accountable people. Evidence: decisions on each potentially affected customer.

Framework under change control. Evidence: framework version on each assessment.

What should you measure?

• Findings confirmed versus rejected by reviewers, by framework item.
• Calls with critical failures, by product and adviser.
• Vulnerable customer calls handled well and badly.
• Patterns linked to incentives or targets, such as failures rising at month end.
• Customer outcomes after flagged sales: cancellations, complaints, arrears.

Where does it usually go wrong?

Ticking boxes. An adviser reads the disclosure at speed and the item is "met". The framework should test whether the customer was given a fair chance to understand, not just whether the words were said.

Transcription errors treated as facts. Speech-to-text struggles with accents, names and figures. A finding that rests on one misheard word can be unfair. Reviewers listen before confirming.

Surveillance culture. If advisers experience the agent as a trap, they game the script. Share the framework openly and use findings for coaching as much as discipline.

Only sales calls reviewed. Conduct problems often show up in the calls after the sale: a customer ringing in confused, trying to cancel, or asking what they bought. Linking post-sale calls to the original sale shows whether the customer understood, which a review of the sale call alone cannot.

Missing the product question. Some harm comes from the product or the script, not the adviser. Patterns across all advisers point to the product, which may need its own review; see whether an agent can review a new product for compliance risk.

Agents selling too. Where voice agents handle sales, their calls need the same review, by a separate agent or by people. The argument is in conduct risk moves when agents do the selling, and the voice agent setup in answering customer calls with a voice agent.

The overall stance on compliance agents is in AI compliance agents.

Where Pryme fits

In Pryme, the reviewing agent reads transcripts and sale outcomes at depth 1 and drafts findings at depth 2. Deciding adviser outcomes and customer remediation is outside its role scope. Where a voice agent sells, it is reviewed by a separate agent, so no agent marks its own work, and every finding is filed in the append-only record.