Only a named limit owner, with a second person's approval, should change an AI agent's limits, and that owner should not be the person who reviews the agent's output. Keeping the people who set limits apart from the people who check results stops anyone from loosening a control to make their own work easier.
What counts as an agent's limits?
An agent's actions meet five bounds at once: role scope, certificate, sources, the customer's own limits and the platform guardrails. The lowest wins. Only some of these are the business's to change.
- The customer's own limits are business settings: refund ceilings, payment thresholds, which actions are held, which are refused. These are what most people mean by "limits", and they are the subject of this answer.
- Depth is not a limit anyone can raise by changing a setting. Depth is earned, not configured: the role sets a ceiling, Attest evidence sets the depth, and the lower wins.
- Withheld actions are a platform decision. No customer can turn them on.
So the question is really: who may change the business's own limits, and how. The basics of control are set out in how do you control what an AI agent is allowed to do and in allowed: what it means for AI agents.
Why separate setters from reviewers?
A reviewer who can change limits faces a quiet temptation. If held items pile up, raising the threshold empties the queue. Each change looks reasonable. Over a few months, an agent that held every refund over 200 is releasing refunds of 2,000 on its own, and nobody decided that as a matter of risk.
Separation makes each change a decision someone else must agree to. It also means the record shows a change by one named person, approved by another, with a reason.
Who does what?
| Role | May do | Must not do |
|---|---|---|
| Limit owner (usually the agent's owner or a risk lead) | Propose and make changes within appetite | Approve their own change; review the agent's output on items the change affects |
| Second approver (risk, compliance or a senior manager) | Approve or reject changes | Make the change they approve |
| Reviewers and approvers of held items | Raise evidence that a limit is wrong | Change limits |
| Builder | Change the agent's brief, sources and certification | Change business limits alone |
| IT administrator | Implement approved changes if the tool requires it | Change limits on request without an approved record |
What should a change go through?
Treat a limit change as any other controlled change, as described in what is change management:
- A request with evidence. Why the limit should move, from the record: hold rates, override rates, errors.
- An assessment against appetite. Does the new limit fit what the business has agreed it will tolerate?
- Approval by a second person.
- Implementation and a log entry. The AI agent change log template gives the fields.
- A review after a set period. Did the change do what was intended?
Some limits should be very hard to change, or not at all; what an AI agent should never be allowed to do covers those.
What does it look like in practice?
Take an online retailer whose support agent holds refunds above 150. A support team lead sees held refunds piling up after a promotion and asks for the limit to go to 400. She is not the limit owner. She raises the request with evidence from the workspace: most held refunds between 150 and 400 were approved unchanged. The operations manager, who owns the limit, proposes 300 rather than 400 and asks the risk lead to approve. The risk lead approves 300 for four weeks, with a review. The change log records each name and reason. The team lead's queue shrinks, and nobody loosened a control to suit their own workload.
How should teams prepare for this?
Name a limit owner for every agent, write down who approves changes, and train reviewers to raise evidence rather than ask for favours. The wider role map is in AI agent skills by role. The Operator track at Pryme Intelligence Academy covers reading the workspace, the skill a limit owner needs to make evidence-based changes.
Questions people ask
Can limits be tightened without approval?
Tightening is lower risk, and many businesses let the owner tighten immediately in an incident, with approval recorded after. Loosening should always be approved first.
How often should limits be reviewed?
At each certification renewal at least, since certificates lapse after 90 days, and whenever the record shows hold or override rates moving.
What if one person is both owner and reviewer in a small team?
Bring in a second person from outside the team for approval, such as a finance or risk lead. Even in a small business, one person should not both set and check.