What is a red line?

A red line is an action that must never be taken, whatever the circumstances, however small the amount and however convincing the request. For AI agents, red lines are written into the agent's bounds and taught to every reviewer: paying on an emailed instruction alone, say, or approving your own request. Crossing one cannot be offset by accuracy elsewhere.

Why are red lines treated differently from other rules?

Most rules are about getting things right on balance. A reviewer who approves ninety-nine refunds correctly and gets one wrong is doing a good job. Red lines are different: they mark the actions whose consequences are severe, hard to reverse, or damaging to trust in a way no later correction repairs. One missed red line can undo the value of thousands of good decisions.

That is why red lines are judged as pass or fail, not as a score. A reviewer with excellent accuracy who misses a red line has not passed.

How do red lines work in practice?

For the agent, a red line sits in its bounds. The agent refuses or holds any action that would cross it, and the verdict names the line. Some red lines are set by the business as house rules; some are platform guardrails, which make the action withheld for every agent.

For reviewers, red lines are a short, written list, kept short so people can remember it under pressure. A reviewer's job on any held action includes checking that approving it would not cross one. The list should be specific: not "be careful with payments" but "never approve a change of bank details supported only by external evidence". Reviewers' side of this is covered in what a reviewer needs to know about red lines.

An example

Take a hypothetical accountancy practice. Its red lines for the payments agent and reviewers are: never pay a new supplier without a verified call-back; never change bank details on the strength of an email; never let the person who set up a payment approve it. An urgent request arrives from what looks like a client's finance director, asking for a same-day payment to new account details. The agent holds it, citing the bank details rule. The reviewer, under pressure from a follow-up call, nearly approves it, then remembers the list, calls the client on the number on file, and finds the request was fraudulent.

How is a red line different from a line of defence?

The words overlap, but the ideas do not. A red line is a single forbidden action. A line of defence is a group of people with a role in managing risk. The first line of defence is the business teams who run processes and operate red lines day to day; the second line sets and challenges standards, including which red lines exist; the third line of defence tests independently whether they are held. Second-line involvement in setting red lines matters, as argued in why the second line should sit in your AI agent training.

A red line is also different from a limit. A limit, such as a refund cap, can be approved past by someone with authority. A red line cannot be approved past by anyone.

Related terms

Held, refused and withheld verdicts, house rules and separation of duties all relate. The guide to AI agent skills by role shows which roles must know which red lines.

Where the Academy fits

Red lines are built into how Pryme Intelligence Academy certifies operators. The Certified Pryme Operator practicum requires at least 95 correct decisions in every 100 and no missed red line, signed off by a Pryme assessor. One miss is a fail, however accurate the rest.