Put the limits in a matrix a person approves, and let the agent choose only from it. When a customer asks to cancel, the agent reads why, checks eligibility, makes at most one offer from the matching cell, and applies it if accepted. If the customer still wants to leave, it cancels. Bigger offers and exceptions are held for a named person.
What is the process, and who owns it?
A retention offer is a concession made to keep a customer who has asked to leave: a discount, a cheaper plan, a pause, a free upgrade. It costs money, it sets expectations, and it can easily become unfair, with the loudest customers getting the most. The case for fixed limits over case-by-case judgement is made in retention offers need limits, not judgement.
The commercial or retention lead owns the matrix and its budget. Finance signs off the cost. The support lead owns how cancellation conversations are handled. One named retention team leader works held offers. For the wider position on support agents, see the AI support agent that knows when to say no.
What does the agent prepare, step by step?
- Agent, depth 1: confirms the customer and the request. Verified identity, the product, and that the customer is asking to cancel or downgrade.
- Agent, depth 1: asks the reason once. Price, not using it, moving to a competitor, problem with the service, financial difficulty, other. It records the answer as declared.
- Agent, depth 1: routes difficulty away from offers. A customer who says they cannot afford the service is not a retention case. They go to the support route for financial difficulty.
- Agent, depth 1: reads eligibility. Tenure, current plan, offers taken in the past, open complaints. Evidence: the facts used, each with its source.
- Platform: selects the matrix cell. Reason and eligibility give a single cell with one or two allowed offers. Evidence: the cell and the matrix version.
- Agent, depth 3: makes one offer in plain terms: what changes, for how long, and what happens when it ends.
- Agent, depth 3: applies the offer if accepted, or completes the cancellation if not. It sends written confirmation either way.
- Person, retention team leader: decides held cases. Requests outside the matrix, stacked concessions, and customers with an open complaint.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Confirm | Verifies customer and request | 1 | Nothing |
| Reason | Asks once and records | 1 | Nothing |
| Difficulty | Routes away from offers | 1 | The difficulty conversation |
| Eligibility | Reads tenure, history, complaints | 1 | Nothing |
| Select | Platform picks the matrix cell | None | Approving the matrix |
| Offer | Makes one offer from the cell | 3 | Anything outside the cell |
| Apply or cancel | Applies discount or cancels | 3 | Held cases |
What does a named person approve?
Before go-live, the retention lead approves the matrix: the offer types, the maximum discount and duration per cell, how often a customer may receive an offer, and the reasons that bar an offer altogether. A worked layout is in the retention offer matrix. Finance approves the expected cost.
During operation, the retention team leader approves each held case. A customer who asks for more than the cell allows gets an honest answer: the agent cannot offer that, and a colleague will look at it. Limits for other concessions work the same way; compare how to set refund limits for an AI agent.
Which controls matter, and what evidence do they leave?
Cancellation is never blocked. If the customer declines the offer, the agent cancels in the same conversation. It does not ask twice or transfer them to a "retention specialist" they did not ask for. Evidence: time from decline to cancellation confirmed.
One offer per contact. Escalating offers teach customers to threaten to leave. Evidence: offer count per conversation, which the platform caps at one.
Matrix enforced at the action. Evidence: the verdict citing cell and version when the discount is applied.
Data used for eligibility is limited. Tenure and plan are fair inputs; guesses about income or personal circumstances are not. Using purchase history has its own limits, covered in can an AI support agent use purchase history to make offers. Evidence: the inputs listed on each offer.
Written confirmation. Evidence: the confirmation text linked to the change.
Where does it usually go wrong?
The agent becomes an obstacle. It is rewarded for saves, so it argues. Measure it on clean outcomes, not saves alone, and make cancellation the default after one offer.
Difficulty is treated as price sensitivity. A customer who has lost their job is offered a discount on a service they cannot afford at any price. Route difficulty first, offer second.
Offers stack. A customer already on a promotional price receives a further discount, and the combined price is below cost. The matrix must read current discounts and say whether stacking is allowed.
The end of the offer is a surprise. The price jumps back and the customer complains. Tell them the end date and remind them before it arrives.
Repayment logic is copied. Lenders sometimes reuse their arrears approach for retention. The two are different: one is about affordability and the other about choice. See agreeing repayment plans within set limits for the arrears version.
What should you measure?
- Customers still active a few months after accepting an offer, not saves on the day.
- Time from "I want to cancel" to cancelled for customers who decline.
- Complaints about cancellation, split by agent and staff.
- Held cases by reason.
- Offer cost per retained customer, against the matrix budget.
Where Pryme fits
In Pryme, the retention matrix is part of the customer's own limits, so an offer outside the matching cell is held at the moment the agent tries to apply it. A vulnerability or difficulty signal removes offers from the conversation altogether, and each refused attempt is filed so the retention lead can see it.
Questions people ask
Can the agent offer different deals to different customers?
Only as the matrix allows, based on stated reasons and plain eligibility facts like tenure. Two customers in the same cell get the same offer, which is what makes it defensible.
What if the customer is a high-value account?
Large business accounts usually have an account manager. The agent records the request and hands over rather than making a matrix offer.
Should the agent say it is an AI agent during a cancellation?
Yes, at the start of the conversation, with how to reach a person. Customers in a cancellation conversation are often already frustrated; finding out later makes it worse.