It can apply a price, not set one. An AI agent can take an assessed application, read its risk grade, term and amount, apply the lender's approved rate card and quote the result. Setting the rate card, discounting, matching another lender's offer, and repricing existing borrowers are decisions for named people in credit and pricing.
What is the difference between applying and setting a price?
Pricing a loan happens in two stages. First, the lender decides its pricing policy: a grid of rates by risk grade, term and amount, approved by people who answer for the lender's margins, its fairness to customers and its compliance with lending rules. Second, each application is placed on that grid. The first stage is judgement. The second is arithmetic once the inputs are known.
An agent can do the second stage well, as long as the inputs come from the right places: the risk grade from the credit decision, the amount and term from the application, the grid from the policy library. It cannot do the first, and it should not drift into it by "adjusting" the grid case by case.
Why do discounts belong with a person?
Discretionary pricing is where unfairness creeps in. If the agent may knock something off for customers who push back, the customers who negotiate get better terms than those who do not, and the ones who do not are often the most vulnerable. In markets with fair lending rules, such as ECOA and Regulation B in the United States, uneven discretionary pricing can also look like discrimination even when nobody intended it.
There is a second reason. Credit scoring of individuals is listed as high-risk under the EU AI Act's Annex III, and risk-based pricing rests on that score. The more the agent's own judgement shapes the price, the more the agent itself becomes part of a high-risk system with documentation and oversight obligations of its own.
What may the agent do, and what stays with a person?
| Task | The agent | Stays with a person |
|---|---|---|
| Reading the risk grade, amount and term | Reads them from the origination system, depth 1 | Disputes about the grade |
| Applying the rate card | Calculates the rate and repayments | Changing the grid |
| Quoting to the customer | Issues the quote in approved wording, depth 3 | Nothing routine |
| Explaining how the price was reached | Explains the factors the policy lets it share | Arguing the price is fair |
| Customer asks for a lower rate | Records the request | Any discount |
| Customer shows a competitor's offer | Records it | Deciding whether to match |
| Repricing an existing loan | Prepares the analysis | Approving it under the contract |
Quoting a price in other settings follows the same split; see whether an AI agent can quote a price to a customer and whether an AI agent can approve a price match.
An example: a car loan quote
Take a lender whose credit decision engine approves a car loan application of 15,000 over four years, with a risk grade of B. The agent reads the grade and the terms from the origination system and applies the rate card. It quotes the rate, monthly repayment and total cost in the lender's standard wording.
The customer replies that a dealer offered a lower rate and asks the agent to match it. The agent records the competitor's offer as the customer stated it, explains that a colleague reviews such requests, and passes it to the pricing desk. A person checks the offer, decides the lender will not match, and explains why. The record shows the grid version used, the quote issued and the person's decision.
What controls make it safe?
- The rate card arrives sealed. The agent uses the approved version from the policy library and records which version it applied.
- Inputs from the system of record. The risk grade comes from the credit decision, never from what the customer says it should be.
- No discount capability. The agent cannot quote below the grid; requests are recorded and held.
- Quotes checked against rules. The quote is checked for required disclosures before it goes out.
- Every quote filed. Grid version, inputs and output are recorded together.
Pricing sits inside loan origination, and the system connections are covered in connecting an AI agent to your loan origination system. The general limits are in AI agents in banking: what they may and may not do.
Where Pryme fits
In Pryme, applying the rate card and changing a price are different capabilities. Attest replays the agent on the lender's past applications and checks every quote against the grid in force at the time; the grounded gate fails any quote the grid does not support. Discounting is outside the role scope, so the verdict on a discount request is refused and filed.
Questions people ask
Can the agent explain why one customer pays more than another?
It can explain the factors the lender's policy discloses, such as risk grade, term and amount. It should not speculate about model features the lender does not publish.
Can the agent change the rate on a variable-rate loan?
Rate changes under a variable-rate contract follow the lender's published process, decided by people. The agent can tell borrowers about a change once it is approved, as part of loan servicing.
Can the agent offer a promotional rate?
Yes, if the promotion is in the rate card for that product and the customer meets the conditions. It cannot create a promotion for one customer.