For a regulated firm, Consumer Duty AI means applying the FCA's Consumer Duty to every conversation an AI phone agent has with a retail customer, the same way you would for a human adviser. The Duty does not change because the first voice on the line is a machine.
The Consumer Duty sets four outcomes: products and services, price and value, consumer understanding and consumer support. An AI phone agent has to help customers understand what they are told, support them when they are vulnerable, and leave records that show you checked how the calls affect those outcomes. The Duty applies to the firm, not to the technology.
The four outcomes on a phone call
The FCA names four outcomes. Each one can be tested on a real call, which makes them a useful way to review an AI agent before and after launch.
| Outcome | What it means on an AI call |
|---|---|
| Products and services | The agent offers only what suits the customers the product was designed for, and the firm has defined who that is |
| Price and value | The agent explains charges clearly and does not present a price as the full cost when it is not |
| Consumer understanding | Information arrives while the customer is still deciding, is pitched at them, and is checked back with them |
| Consumer support | The customer can reach a person easily, and the agent creates no hurdles to leaving, switching or complaining |
Read each row as a test. Pick a real call for each outcome, and ask whether a reasonable customer would have understood the price, the product and the next step without help.
Vulnerable customers on an AI call
The FCA's guidance for firms on the fair treatment of vulnerable customers, FG21/1, was published on 23 February 2021. It does not require specific steps, but firms must still meet the FCA's Principles and treat customers fairly. The guidance treats vulnerability as a spectrum of risk, and says any customer could end up in vulnerable circumstances.
The guidance lists examples that can be characteristics of vulnerability, including poor health such as cognitive impairment, life events such as new caring responsibilities, low resilience to cope with financial or emotional shocks, and low capability such as poor literacy or numeracy skills. Some of these show up in a voice conversation: a caller who repeats questions, sounds confused, mentions a recent bereavement or says they are struggling to keep up.
For an AI agent, the rule is simple. It should notice these signs, slow down, and hand over to a person, rather than try to diagnose a customer. The agent should not decide that someone is vulnerable. It should flag the situation for a person who can respond with care.
Designing calls for good outcomes
Design decisions made before launch shape most of the outcomes. A practical set of habits includes:
- Say at the start of the call that the caller is speaking to an AI assistant, and say how to reach a person.
- Confirm the key facts, such as the product, the price and the date, and ask the customer to repeat them back.
- Give charges and conditions before the customer agrees, not after.
- Offer a person as soon as the customer asks, without argument or repeated offers.
- Make it easy to stop, change or complain, and say how to do it at the end of the call.
Each of these can be checked on a recording. If a habit cannot be checked, it is not yet a design.
Think about the customer who has just received a letter about a price rise and rings to ask what it means. The agent should explain the change in plain words, confirm the date and the new amount, and offer a person if the customer wants to talk through their options. A customer who leaves that call understanding what happens next has been supported, whether or not they stay with the product.
Firms in banking and lending can see the same principles in a specific sector. The financial services page covers how VoiFlow is used there, and the AI voice agents for banks and lenders article looks at the common call types.
The evidence a firm should be able to show
The FCA's guidance on vulnerable customers expects firms to monitor and evaluate outcomes. An AI agent can give you more evidence than a human team does, because every call can be recorded, transcribed and scored. The question is whether you keep that evidence in a form you can use.
- Call recordings and transcripts, with access controls and a retention period you can justify.
- Scores from automated checks, plus regular reviews by a person, including calls with vulnerable callers.
- Outcome data by customer group, such as complaints, cancellations and handovers to a person.
- Test results for every change to the agent's knowledge or rules.
- A record of who approved each change, and when.
Keep this evidence in one place, so that a single question from a supervisor does not turn into a week of searching.
What the FCA is likely to ask
Expect questions about how the agent was designed, who tested it, which customers it serves less well, how you learned about failures, and what you changed as a result. Firms should be able to show how their business model, their actions and their culture secure fair treatment for all customers, including vulnerable customers.
A good answer is specific. It names the call types, shows the test results, points to a change made after a review, and explains who owns each part of the process.
Where AI agents change the risk
An AI agent brings strengths and risks that a human team does not. It is consistent, so a good script is applied the same way on every call. That consistency also means a weak script is repeated on every call, and a small misunderstanding can reach thousands of customers before anyone notices.
The agent can also sound confident when it is wrong. A customer who hears a fluent, polite answer is less likely to question it. For that reason, accuracy checks belong in the design: answers should come from approved information, prices should be read from current data, and the agent should say when it does not know and offer a person. Each of those is a point a reviewer can test on a recording.
How VoiFlow handles this
VoiFlow records, transcribes and scores every call, and gives supervisors a live operations view. A warm handoff passes the call to a member of staff with the summary, and supervisors can listen in, whisper to the agent or take over the call. Those features give a firm the evidence trail described above.
They do not decide whether a firm meets the Duty. That judgement stays with the firm and its compliance team. Read the trust page for how VoiFlow handles data, and build your own evidence file alongside it.
What to do next
List your retail call types, then check each one against the four outcomes and the vulnerability questions in this article. When you want to plan the set-up, contact us and we can talk through your first call type. To see how the evidence is gathered in practice, read about quality assurance for AI calls.
Frequently asked questions
Does the Consumer Duty apply to AI phone agents?
The Duty applies to firms, not to the technology. An AI agent is part of the firm's customer journey, so the firm answers for the outcomes the agent produces on calls.
What counts as a vulnerable customer?
The FCA treats vulnerability as a spectrum of risk, and any customer could end up in vulnerable circumstances. Examples include poor health such as cognitive impairment, life events such as new caring responsibilities, low resilience to shocks, and low capability such as poor literacy or numeracy.
Is the FCA's vulnerability guidance binding?
The guidance does not require firms to take specific steps. Firms must still meet the FCA's Principles and treat customers fairly, and the guidance shows how to do that in practice.
How often should AI call outcomes be reviewed?
Review every new call type before launch, sample calls each week during the first months, and review the vulnerable-customer cases every time the agent's rules or knowledge change. Write the review schedule down, keep the dates, and record what each review found.
Who should own the evidence?
Name one person in the firm who is accountable for the agent's outcomes, with a named deputy. That person should be able to show the test results, the review records and the approvals for each change without having to search for them.
Do I need to record AI calls to show good outcomes?
Recordings and transcripts are the most direct evidence of what the customer heard. Keep them with access controls and a retention period you can justify, and tell callers at the start that the call is recorded.
This is general information, not legal advice.
