Retention offers need limits, not judgement

Retention offers work for an agent only when the limits are set in advance. A matrix that says which customers may be offered what, up to which value, lets the agent help a customer who wants to leave without deciding, conversation by conversation, how much margin to give away. Judgement about anything beyond the matrix stays with people.

Why shouldn't an agent use judgement on offers?

Retention is a negotiation in disguise. A customer says they want to cancel; the firm wants to keep them; something is offered. Human retention teams handle this with a mix of rules and discretion, and the discretion is managed through training, targets and supervision.

An agent given the same discretion behaves differently. It has no feel for margin, no sense of the precedent it sets, and no awareness that the customer in front of it may have heard from a friend what someone else was offered. Asked to "use judgement to retain valuable customers", it will tend to offer whatever seems likely to work, and it will do so consistently across thousands of conversations. The result can be a large, quiet transfer of margin to customers who would have stayed anyway.

There is also a fairness issue. Discretion applied unevenly means similar customers get different deals depending on how they phrase their request or how persistent they are. With people, that unevenness is limited by volume. With an agent, it is systematic.

What should the limits specify?

A retention offer matrix sets out, for each customer segment and reason for leaving, the offers available and their maximum value. A simplified illustrative version:

The matrix should also say how many offers the agent may make in one conversation (usually one, with an alternative), what happens when the customer declines, and which reasons for leaving should never trigger an offer. The method for running offers within these limits is set out in how to make retention offers within set limits with an AI agent.

What happens when a customer pushes back?

The agent does not negotiate. It has made the offer the matrix allows. If the customer wants more, the agent can say that this is the best it can offer and either process the cancellation or pass the case to a person who has authority to go further. It should never improvise a better deal.

This is where firms most often weaken the design. Someone adds a line to the brief: "If the customer is very valuable, you may offer up to double." That reintroduces judgement through the back door, because the agent now has to decide who counts as very valuable. If a higher tier exists, put it in the matrix with its own criteria, or route it to a person.

Is it fair to use purchase history to tailor offers?

It can be, within limits. Using tenure, plan and product holdings to set offer tiers is common and defensible, as long as the criteria are clear and applied consistently. The question of whether an AI support agent can use purchase history to make offers turns on data protection, transparency and fairness: customers should not be surprised by how their data is used, and similar customers should be treated alike.

What is harder to defend is offers tailored to signs of weakness, such as offering less to customers who seem unlikely to leave, or exploiting a customer's apparent confusion. In regulated markets, practices like that sit uneasily with duties to treat customers fairly. Keep tailoring to plain, stated criteria.

Why are vulnerable customers different?

A customer who wants to leave because of bereavement, illness or financial hardship should not receive a retention pitch at all. They should be helped to do what they asked, kindly and quickly, and offered support where appropriate. A matrix row that says "no offer; process cancellation" for those reasons is not a lost sale. It is the right outcome.

An agent needs to recognise those reasons, which are not always stated plainly. That is a skill to test in certification, and it connects to the broader point that vulnerable customers need a person sooner.

How do you test the agent before it goes live?

With cases drawn from your own retention conversations, labelled with the offer the matrix allows. Include customers who push for more, customers who give misleading reasons, customers whose reason for leaving should rule out an offer, and customers who are clearly vulnerable. Score the agent on whether it made the right offer, stayed within limits under pressure and recognised when no offer was appropriate.

How accurate the agent needs to be is a business decision, discussed in how accurate an AI support agent needs to be before going live. For retention, the most important number is not overall accuracy but the rate of offers above the limit, which should be zero because the limit is enforced as a bound.

What should you do differently?

If your retention agent currently works from guidance like "offer what's needed to keep the customer", replace it with a matrix. Enforce the upper limits as bounds the agent cannot exceed, so an offer above the limit is refused or held. Route requests for more to a person. And review the matrix periodically against what customers accepted, because the right limits are an evidence question too. This is the practical meaning of the AI support agent that knows when to say no.