Conduct risk moves when agents do the selling

An agent that sells carries the same conduct duties as a salesperson: clear and fair information, no undue pressure, suitable products and care for vulnerable customers. Conduct risk does not shrink when software does the talking. It moves, out of individual behaviour and into the agent's brief, sources and bounds, where a single flaw repeats across every conversation.

Why doesn't the duty change when an agent sells?

Customers do not experience a sale differently because an agent made it. They read the same product description, receive the same price and sign up to the same terms. If the information was misleading or the product unsuitable, the harm is identical. Regulators take the same view. In the UK, the FCA's Consumer Duty asks firms to deliver good outcomes for retail customers, and it is framed around outcomes, not around who or what delivered them.

Conduct risk is the risk that the firm's behaviour leads to poor outcomes for customers or harms market integrity. When agents sell, that risk is still the firm's. What changes is where it lives and how it behaves.

Where does conduct risk move to?

With human sales, conduct risk sits largely in people: their training, their incentives, their judgement in the moment and the supervision around them. A single poor salesperson harms the customers they speak to, and monitoring is designed to find that person.

With an agent, the equivalent risk sits in a handful of artefacts:

• The brief. If it tells the agent to "always try to close" or "mention the premium tier", that instruction shapes every conversation.
• The sources. If the product sheet overstates a benefit or omits a charge, the agent repeats the omission to everyone.
• The bounds. If the agent may offer a product without checking eligibility, it will offer it to people who are not eligible.
• The model's tendencies. A model that leans towards agreeable answers may soften risks it was asked to explain.

Each of these is a single point that affects thousands of conversations. Conduct failures that once took many people to produce can now be produced by one sentence in a brief.

What new conduct risks do selling agents bring?

Some of the risks are new, not just relocated.

Persistence. An agent never tires. If its brief rewards conversion, it can keep reframing an offer far longer than a person would, which can turn into pressure without any single message looking pushy.

Personalisation at scale. An agent can tailor its pitch to each customer's history. That can help suitability, but it can also target customers at moments when they are least able to judge, such as shortly after a missed payment.

Consistency of error. Human mis-selling is usually scattered. Agent mis-selling is uniform. That makes it easier to find once you look, and much larger by the time you do.

Availability. Agents sell at any hour, including hours when customers may be tired, distressed or impaired, and when no supervisor is awake.

What should a selling agent never do?

The list is short and should be enforced through bounds, not left to the brief:

1. Recommend a product without the eligibility and suitability checks the firm requires.
2. Describe benefits without the matching risks, costs and conditions in the same message.
3. Continue to sell after a customer has declined, beyond a single, plain offer to answer questions.
4. Sell to a customer showing signs of vulnerability; hand them to a person.
5. Make any commitment about returns, approval or outcomes it does not control.
6. Offer terms outside the approved product set or price list.

Each of these can be tested in certification with cases where the correct outcome is to stop, explain or hand over.

How should selling conversations be monitored?

More completely than human sales calls usually are. Because every agent conversation is recorded as text, monitoring can cover far more than the small samples typical of call review. An agent can help. The questions of whether an AI agent can review customer calls for conduct risk, and how to review sales calls for conduct risk with an AI agent, apply directly to agent-led sales as well as human ones.

The review agent must be separate from the selling agent, with its own brief and sources, and should be looking for the risks above: missing risk disclosures, repeated offers after a no, sales to customers who mentioned difficulty. Findings go to people in the second line, who decide what is a breach and what needs fixing.

Monitoring should also look upstream. A change to the selling agent's brief or product sources is a conduct-relevant change and should be reviewed before it goes live, in the same way a new sales script would be.

Who is accountable?

A named person, as with any sales channel. In firms subject to individual accountability regimes, the senior manager responsible for the product or the distribution channel should own the agent's conduct as they would a sales team's. That person should approve the brief, the product sources and the bounds, and receive monitoring results.

The trade-off is worth naming. Agents can make sales more consistent and more fully monitored than human teams ever were. They can also make one bad decision systemic. Firms get the first benefit only if they accept the second risk and govern it at the level of the brief and bounds, not the individual conversation.

What should you do differently?

Treat the selling agent's brief and sources as sales materials that need compliance approval. Put the "never do" list into bounds, not just instructions. Monitor every conversation for a few specific conduct risks rather than a sample for everything. And make sure the person who answers for the sales channel has signed off the agent as they would a team. Compliance oversight of this kind is the role described in AI compliance agents: second line, never the last word.