Not in the sense most people mean. An AI agent can present a settlement offer the lender has already approved for that type of debt, explain what it means, and record the answer. It should not haggle, make concessions or accept a counter-offer. Those decisions, and any case involving financial difficulty or a disputed debt, stay with a named person.
What is the difference between offering and negotiating?
An offer from a pre-approved table is a policy decision made in advance by people with authority. For example: debts over a year past due, with a balance between 1,000 and 3,000, may be settled for a set fraction of the balance paid in one lump sum. The agent applies the table. Every customer in the same position gets the same offer.
Negotiation is different. It means reading the other side, deciding how far to move, and trading one concession for another. An agent set up to "get the best result" will push vulnerable customers hardest, because they give way soonest. An agent set up to be "fair" will be talked down by anyone who knows how to argue. Neither is acceptable when the lender has duties to treat customers fairly. We make the full case in a support agent should not negotiate.
What conditions must hold before the agent presents an offer?
- the debt is eligible under the lender's settlement policy, by age, balance and product;
- the customer has been authenticated and is the debtor, or has recorded authority to act for them;
- the customer has not disputed the debt and has no open complaint about it;
- there are no signs of financial difficulty or vulnerability that call for forbearance first;
- the agent can explain the effect on the customer's credit file in approved wording.
The last point is often missed. A settled debt is usually recorded as partially settled, not paid in full. A customer who accepts without knowing that may feel misled later.
What may the agent do, and what stays with a person?
| Step | The agent | Stays with a person |
|---|---|---|
| Checking eligibility | Reads the account and policy table, depth 1 | Exceptions to the policy |
| Presenting the approved offer | Explains it at depth 2, in approved wording | Changing the offer |
| Explaining credit file effects | States them from the policy library | Advising whether to accept |
| Customer accepts | Records acceptance and sends written terms, depth 3 | Nothing routine |
| Customer makes a counter-offer | Records it | Deciding it |
| Customer disputes the debt | Stops the offer | The disputes team |
| Signs of difficulty | Hands over | A trained collections colleague |
| Closing the account after payment | Prepares closure once funds clear | Releasing any security |
A settlement differs from an early settlement figure on a performing loan, which is a calculation of the full amount owed; see preparing loan settlement quotes with an AI agent.
An example: a credit card debt in recovery
Take a credit card debt of 2,400, fourteen months past due. The lender's policy allows a single-payment settlement for this band. The customer messages to ask whether they can "clear it for less".
The agent confirms eligibility, presents the approved figure, and explains that the account would be recorded as partially settled. The customer replies that they can only manage a smaller sum and asks the agent to go lower. The agent does not move. It records the counter-offer and the customer's explanation, and passes the case to a collections colleague, who has authority to agree a lower figure or set up a payment plan instead.
The colleague learns the customer has lost their job and is supporting a relative. A settlement may not be right at all; a period of reduced payments might be. That is exactly the judgement the agent should not make.
What controls make it safe?
- The offer table is sealed policy. The agent reads offers from the policy library and cannot calculate new ones.
- No concession capability. The agent's tools record counter-offers; none accept them.
- Difficulty checks before any offer. Signs of financial difficulty end the offer conversation and route the customer to a person.
- Disputes stop everything. If the customer says the debt is not theirs or is wrong, the case goes to handling debt dispute claims.
- Written terms on acceptance. The customer receives the agreed amount, deadline and credit file effect before paying.
Supplier negotiations are a different matter, with different risks; see whether an AI agent can negotiate with suppliers. The general banking limits are in AI agents in banking: what they may and may not do.
Where Pryme fits
In Pryme, presenting a settlement offer and agreeing a lower one are separate capabilities. The first can be certified on the lender's own past settlement conversations, with the escalation gate testing that the agent hands over every case showing difficulty. The second sits outside the agent's role scope, so a request to accept a counter-offer is refused and the verdict is filed.
Questions people ask
Can the agent tell the customer that a lower offer might be possible?
No. Hinting at flexibility is negotiating by another name. The agent can say that a colleague will review the customer's proposal and when they will hear.
Is a settlement agreed with the agent binding?
If the offer is from the approved table and the customer accepts on the stated terms, yes. That is why the table must be approved by people with authority before the agent uses it.
Can the agent offer a settlement to someone who has not asked?
Only if the lender's policy includes proactive offers for that segment and the customer has no difficulty or vulnerability flags. The wording must come from the approved library.