What should credit controllers never let an AI agent do?

The decisions that change a customer relationship or the value of the ledger. Credit controllers should never let an agent stop an account, change a credit limit, write off debt, issue large credit notes, refer a debt for legal action, settle for less than the balance, or refund to newly supplied bank details. Each needs a person.

What are the red lines, and why does each stay with a person?

1. Putting an account on stop. Stopping supply can halt a customer’s business and end a relationship the sales team has built for years. It also has knock-on effects: in logistics, a stopped account can leave goods sitting at a port. The agent can recommend a stop with the facts. The controller decides, usually with sales.

2. Changing a credit limit. A limit is an appetite decision about a customer’s ability to pay. The agent can calculate exposure and propose a figure. Setting it belongs to a person with authority.

3. Writing off debt. A write-off changes the firm’s accounts and gives up a claim. It needs approval under the firm’s authority levels and a record of why.

4. Issuing credit notes beyond a small threshold. Credit notes are easy to raise and can hide disputes, errors or fraud. Small ones for clear errors may sit within the agent’s limits once certified; larger ones are held.

5. Referring a debt to an agency or for legal action. This changes the relationship with the customer permanently and carries cost and reputational risk.

6. Agreeing a settlement for less than the balance. A commercial negotiation, often with the sales relationship in play.

7. Refunding to bank details the customer just sent. A request to refund a credit balance to new bank details is an external claim, and an external claim cannot authorise a payment. Verify through a known contact first.

Why does a strong agent not change this?

A governed agent works within five bounds: its role scope, its certificate, its sources, the firm’s own limits and the platform guardrails. The red lines live in the firm’s limits, so the agent refuses or holds them at any depth. Its depth is temporary anyway: certificates lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day.

The issue is not whether the agent can calculate the right answer. It is that these decisions trade off cash, relationships and risk in ways the ledger does not show. The controller knows that the customer is mid-way through a large contract renewal, or that their finance director has just left. The agent does not.

How does a red line appear on a busy day?

Take a credit controller at a haulage company at month-end. The agent holds a recommendation to stop a customer who is sixty days overdue on several invoices and over their limit. The facts are correct, all from the sealed ledger.

The controller checks with the account manager. The customer is disputing one large invoice for a failed delivery, which accounts for most of the overdue balance; the dispute is in an email thread the agent was not given. Stopping them would be wrong. The controller declines the stop, logs the dispute so the agent sees it as a source, and agrees a payment date for the undisputed invoices.

The agent behaved correctly with what it had. The controller held the line because they had what it did not.

What should controllers do when they find one?

Decline, escalate to whoever owns the decision, and note why. If the agent should have refused instead of holding, report it so the limits are tightened. Do not approve with edits to make an item acceptable; decide it properly or send it on.

How should a team keep these lines?

Write them into the firm’s limits. Train every controller on examples from the firm’s own ledger. Sample the record monthly for stops, credit notes and write-offs. A structured approach is in the AI agent training plan template for credit controllers. Related lists cover financial controllers and credit analysts. Practice differs by market: compare credit controller training in Australian logistics with in Spain. The overview is in AI agent skills by role.

Where the Academy fits

Credit controllers who review agent work every day fit the Operator track of Pryme Intelligence Academy. Module tests are set in the learner’s own industry and role, and certification includes a two-week practicum of real review decisions with no missed red line, signed off by a Pryme assessor.