The FCA's Consumer Duty does not change because an AI agent answers the customer. Firms must still deliver good outcomes for retail customers: support that helps rather than obstructs, communication people understand, fair treatment of the vulnerable, and evidence of all three. A support agent must be designed and monitored against that, not just response times.
What does the Consumer Duty expect of customer support?
In substance, the Duty requires firms to act to deliver good outcomes for retail customers. For support, that means customers can get help when they need it, can act on their own choices without unreasonable barriers, receive information they can understand and rely on, and are not treated worse because of vulnerability. Firms must also monitor outcomes and act when they fall short. Our explainer on what the FCA Consumer Duty is sets out the framework, and our answer on whether it applies to AI agents confirms that the obligations attach to the firm, whatever tool it uses.
The Duty is outcome-focused. The FCA does not prescribe how a support channel must work; it expects the firm to show that the channel produces good outcomes. That puts the burden of proof on the firm's own evidence.
Where can an AI support agent fall short?
Support agents create some risks that a well-run human team would rarely produce, and amplify others.
Friction by design. An agent configured to retain customers may make cancelling, switching or complaining harder than buying. It may loop, ask unnecessary questions or steer people away from what they asked for. Each conversation may look reasonable; the pattern is a barrier.
Answers that sound right but are not. A fluent agent can state product terms, fees or eligibility rules incorrectly, with complete confidence. A customer who relies on that answer has been misled.
Missed vulnerability. People rarely announce vulnerability. Signs appear in what they say and how: confusion, distress, mention of bereavement, illness or financial difficulty. An agent that does not recognise these signs will treat the customer as it treats everyone.
Inconsistency. Two customers with the same question can receive different answers on different days, which makes fair treatment hard to show.
No way out. Customers who want a person cannot reach one, or must work hard to do so.
What controls answer each risk?
The controls are about bounds, sources, routing and records, not about how the model works.
| Risk | Control | Evidence it leaves |
|---|---|---|
| Friction on cancelling, switching or complaining | The agent may not obstruct these requests; it must complete or route them at once | Time from request to completion or handover, by request type |
| Wrong answers on terms or fees | Answers must be grounded in the current policy library; ungrounded answers are refused or routed | Grounding results in certification; sampled answer reviews |
| Missed vulnerability | Signs of vulnerability route the conversation to a trained person | Count and handling of routed conversations; samples reviewed |
| Inconsistency | Answers drawn from the same sealed sources; certification on real cases | Consistency checks across similar questions |
| No way out | A person is always reachable on request, without obstruction | Time to reach a person; abandonment after a request |
Our guide on Consumer Duty and AI agents in customer service works through each in more detail.
How should the agent's authority be set?
Narrowly at first, and widened on evidence. A support agent might answer factual questions about products at depth 3, within its grounded sources, while drafting complaint responses at depth 2 so a person sends them. It should hold no authority to refuse a cancellation, deny a hardship request or make a judgement about a customer's circumstances. Those stay with people.
Some actions should never be available to it. Discouraging a customer from complaining, or presenting a retention offer in a way that delays a cancellation the customer has clearly requested, are examples firms may choose to forbid outright. Writing these down as refused actions, checked when the agent acts, is stronger than writing them in its instructions.
The same thinking applies beyond service. Where support agents touch lending, such as arrears conversations or forbearance requests, the stakes are higher still; our page on Consumer Duty and AI agents in lending covers that ground.
How do you show outcomes are good?
By monitoring them, not assuming them. The Duty expects firms to monitor customer outcomes and to look for differences between groups, including customers with characteristics of vulnerability. For an AI support agent, that means sampling conversations against a standard, measuring how quickly cancellations and complaints are completed, tracking how often customers ask for a person and how long they wait, and comparing outcomes across groups.
Agents can help with this monitoring itself, for example by reviewing large volumes of conversations for signs of poor outcomes. Our piece on whether an AI agent can monitor customer outcomes for Consumer Duty discusses where that works. The monitoring agent should be separate from the support agent, with different owners, so the firm is not marking its own work.
Accountability for the result sits with named people. Under the Senior Managers and Certification Regime, senior individuals answer for areas of the business, and support delivered by an agent is still within someone's responsibility.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What is the trade-off?
Designing for good outcomes limits what a support agent may do. It cannot be tuned purely for containment or retention. It must route vulnerable customers and customers who want a person, even when it could handle the conversation. Grounding every answer and monitoring outcomes by group costs effort.
An agent tuned for containment will look efficient and may deliver outcomes the firm cannot defend. For each support agent, list the actions it must never obstruct, the signs that should route to a person, and the outcomes you will monitor, by group. Then check each is enforced and measured. The wider mapping of rules to controls is in our guide to AI agent regulation, mapped to controls.
Questions people ask
Must customers be told they are talking to an AI agent?
Clear communication is part of the Duty's expectations, and telling customers they are dealing with an AI system supports understanding and trust. Where EU rules apply, Article 50 of the EU AI Act separately requires people to be told when they interact with an AI system.
Is a high containment rate a problem under the Duty?
Not in itself. It becomes one if containment is achieved by making it hard to reach a person, complain or cancel. Measure outcomes for contained conversations, not just the containment figure.
Does the Duty apply to business customers?
The Consumer Duty is aimed at retail customers. Firms serving business customers have other obligations, but the controls above remain sound practice for any support agent.