An AI agent in a cancellation conversation should never pitch again after the customer has said no, never miss an opt-out however it is worded, and never sell to someone who is complaining or vulnerable. It should also never argue, never handle a legal threat on its own, and never close in a way that pretends the conversation went well.
We build and run retention agents for UK broadband providers. These are the rules we hold them to, and most of them come from testing exactly where retention conversations go wrong.
Why does a cancellation conversation need its own rules?
Because the customer does not have to be there.
Since One Touch Switch went live for most UK customers on 12 September 2024, the new provider organises the switch. The customer does not need to contact their current provider at all. Some will have picked it precisely so they would not have to talk to you.
So the agent is talking to someone who owes it nothing. The job is one clear offer, an easy way out and a clean record. An agent tuned for persistence damages the brand with exactly the customers you want back.
Never pitch again once the customer has decided
This one rarely comes from the prompt. It comes from the plumbing.
A common pattern: a customer declines, then writes again later. The first conversation has ended, so the system opens a new one, finds a recent retention record and starts the flow again. The customer hears the same offer after already saying no.
The fix is a state rule, not a wording change. Once the last outcome was a decision (declined, opted out, moving somewhere you cannot serve), the next message goes to the normal support agent, not back into retention. Test it by sending a second message straight after a decline.
The same applies inside a conversation. Ask the save question once. If the answer is not one the agent recognises, it should move on, not ask the same question reworded.
Never miss an opt-out because of how it was worded
Customers rarely type STOP. They write "leave me alone", "block", "end the conversation" or "never message me again".
If only the keyword counts, those customers keep getting messages, and that is how a polite decline turns into a complaint. Every phrasing that means stop contacting me should trigger the opt-out, and the opt-out should outrank every other route. If a message is both a complaint and an opt-out, the opt-out wins.
Then check it reaches every system that sends. A suppression that stops outbound messages but is ignored on the reply path is an easy gap to leave.
Never sell to someone who is complaining or vulnerable
When a customer says they are leaving because the service is broken, a discount reads as buying them off. A tempting design is to offer a choice between getting the fault looked at and carrying on with the offer. That second option keeps selling a cheaper price on a service the customer has just said is letting them down. Offer a person looking at the problem, or no thanks.
Answer the question first, too. A customer asking why they are charged an exit fee is asking a question, not complaining. Treat it as a grievance and it never gets answered.
Vulnerability is stricter still. Ofcom's General Conditions (C5) have required providers to have policies and procedures for treating vulnerable customers fairly since October 2018, and its guidance names bereavement, job loss, health problems and changes in household income among the circumstances that can make anyone vulnerable. Ofcom expects staff to be trained to recognise the signs. An agent is frontline. If a customer mentions any of those, the offer stops and the route to help or a person starts.
More on that boundary in can conversational AI handle complaints.
Never handle a legal threat or a formal complaint alone
A customer who mentions court, a solicitor or the ombudsman should reach a person, with a ticket raised and the whole conversation attached. So should a repeated complaint that has not been resolved.
The trap is an outcome that closes by design. If an outcome such as "not interested" ends the conversation with no ticket, then a legal threat in that conversation reaches nobody. Something has to look for it specifically.
There is a clock on this. Since 8 April 2026, Ofcom's rules let customers take an unresolved telecoms complaint to an ADR scheme after six weeks, down from eight. A complaint the agent quietly closed is still a complaint.
Never argue, and never talk about the competitor
The agent should not dispute the customer's reasons or guess at a competitor's motives. Left alone, an agent will sometimes guess at why another provider said something to the customer. Write that out. It should restate what staying gets them and say nothing about anyone else.
Never close as if it went well
One scripted close for every decline sounds efficient. Then it thanks a customer warmly straight after they have complained.
Use one neutral close that is true in every case: understood, thanks for letting us know, we will not message you about this again. Then make sure the system keeps that promise.
What to do before go-live
Write these as test scenarios before the build and run them against the live agent, not a copy: a decline then a second message, an opt-out in five wordings, a complaint then an offer, a mention of court. The wider retention picture is in AI for ISP and altnet customer retention, and the handover side in what happens when an AI agent does not know.
If you are planning a retention agent, we are happy to talk through our test list. You can also try our agent and see how it takes no for an answer.
