Setting the limits an AI agent negotiates within

Can an AI agent really negotiate a payment on a live call?

Yes. A conversational AI agent, meaning software that understands free replies and answers them, can find out what the customer can manage, propose an arrangement, answer a counter-offer and close on terms both sides accept. What the lender will accept is decided before the call, by people, and handed to the agent as its mandate: the list of offers it may make and the limits on each. AI payment plan negotiation is the part of payment recovery that used to need a collector on every call.

Inside that mandate the agent works the way a good collector does. It listens for what the customer's month looks like, picks the approved offer that fits, and says it plainly. When the customer asks for something the mandate allows, it agrees there and then. When they ask for something it doesn't, it says what it can do instead, and if that won't work, it passes the conversation to a person who can.

The real question for a lender is whether the mandate is clear enough for anyone, human or software, to follow without guessing.

What goes into a negotiation mandate?

Five or six levers are usually enough for a consumer book; the table shows one set. Each needs a limit and a named owner, usually someone in credit risk or collections strategy. Limits normally vary by product and by how far behind the account is, so a card balance at 20 days past due gets different room from a personal loan at 80.

Write the mandate as a table, not as prose. A table can be checked line by line against what the agent actually agreed; a paragraph of guidance can't.

In what order should the offers come?

In the order that recovers the most of the balance soonest, adjusted to what the customer has already told you. Start with payment in full if they can, then a near date if the money arrives within days, then a short plan, then a longer one. Any settlement discount, where your mandate allows one at all, comes last, because a discount offered early can set an expectation for the next call.

That order is a default, not a script. If the customer opens by saying they've lost their job, pushing payment in full first is tone-deaf, and the conversation belongs with a person who can look properly at what the customer can afford, an affordability assessment. If they say they get paid on the 3rd, the agent should go straight to a date after the 3rd. Asking about their month before naming a figure is what makes the order work, and it's the same discipline a kept promise to pay, a date and amount the customer commits to, depends on.

One rule is worth writing into every mandate: the agent never presents an offer as final when a better one is available to it. It can lead with three months, but if three won't work and six are allowed, it has to say so.

What does a negotiation at the edge of the mandate look like?

Take the mandate in the table above and a personal loan customer who is 45 days behind on a balance of 1,800 euros. The agent confirms it is speaking to the customer themselves, a right party contact, and asks what's changed. The customer explains their hours were cut, they get paid on the 3rd, and they can put 200 euros a month towards this.

Payment in full is out, and so is a date within 14 days. An instalment plan is next. Six instalments is the longest the mandate allows, and 1,800 divided by six is 300 euros a month, which is 100 more than the customer said they can find. At 200 euros a month the balance takes nine months, three past the limit. No combination of the approved levers fits, and there's no discount at 45 days.

So the agent doesn't stretch the rules or push the customer to promise 300. It says plainly that it can offer up to six months at 300 a month, that it understands 200 is what works, and that a colleague can look at a longer plan. If the customer wants to pay something today, it takes 200 as a part payment and passes the call or arranges a callback with the facts already written down: the balance, the reason, the 200 a month, the payday on the 3rd. The person who picks it up decides, without asking the customer to start again.

What should happen to requests outside the mandate?

They should be recorded, every one, with what the customer asked for and why. A customer who wants nine months instead of six, or a date 20 days out instead of 14, is telling you something about your limits, and individually those calls look like exceptions. Counted across a month, they show where a limit sits in the wrong place.

Review them monthly with whoever owns the mandate. If a large share of handovers are customers asking for seven or eight months on mid-sized balances, and the people who take those calls approve most of them, the six-month limit is generating work without protecting anything. Move it and watch what happens to kept payments. If the requests are scattered and the approvals rare, the limit is doing its job.

How do you check the agent never goes past the limits?

Before go-live, test calls that try to push the agent past its mandate should each end with an approved offer or a person, never with a new term. The check that matters more runs after launch.

After go-live, compare every arrangement the agent records with the mandate, automatically and every day. Any arrangement outside it should show up on a list someone reads that morning, rather than waiting to be found through a complaint.