No, not on every action. Every agent needs a named person accountable for it, but approving each action one by one only makes sense where the action is consequential, hard to undo, or not yet proven. For routine work an agent has earned on real cases, a person on the loop, watching samples and alerts, is the better control.
What is the difference between in the loop and on the loop?
A person in the loop approves each action before it happens. The agent drafts; the person sends, pays or decides. A person on the loop does not approve each action but supervises the flow: reviewing samples, watching for alerts, able to stop the agent at once.
Both are human oversight. The question is which fits each capability, not which is morally better. Putting a person in the loop on thousands of trivial actions does not add control. It produces fatigue and approvals nobody reads.
Who decides which applies?
The process owner proposes, per capability, and risk approves. The decision follows from the action's consequences and the evidence the agent has earned, not from a general view about AI.
| Depth | Typical human role | Why |
|---|---|---|
| 1. Read | Owner reviews outputs periodically | Nothing changes in your systems |
| 2. Draft | In the loop: a person sends or applies each draft | The agent has not yet earned execution |
| 3. Execute routine work | On the loop: sampling, alerts, stop control | Proven on real cases; each action is small and correctable |
| 4. Move money or hard-to-undo actions | In the loop above set limits; on the loop within tight ones | Consequences are larger and slower to reverse |
| 5. Coordinate other agents | On the loop over the orchestrator, with each agent's own controls intact | Coordination adds no new authority |
Two legal points shape this. Under GDPR Article 22, people have rights around decisions based solely on automated processing that have legal or similarly significant effects, so those decisions need meaningful human involvement or another lawful route. Under the EU AI Act, Article 14 requires high-risk systems to be designed for effective human oversight. The difference between the two ideas is set out in human-in-the-loop versus human oversight under the EU AI Act.
Where is it enforced, and what is recorded?
At the moment of action. The depth set for each capability decides whether an action executes or becomes a held draft for a named person. A rule in a procedure saying "a person checks these" is not enforcement; a held verdict that cannot proceed without a named approval is.
The record shows which: an allowed action filed with the agent's identity, or a held action filed with the approver's name, the time, and what they decided. Sampling on the loop is recorded too, with the reviewer and findings.
What does this look like in practice?
Take a card issuer's disputes team. The agent reads incoming disputes, gathers transaction data and drafts the case. For one category, duplicate charges where the ledger shows two identical transactions minutes apart, the agent has been certified at depth 3 and raises the chargeback itself within a set amount. A team lead samples a set number of these each week and can stop the capability instantly.
For disputes alleging fraud, the agent stays at depth 2. It drafts the case and a person decides, because the outcome affects the cardholder's liability and depends on judgement about what they say happened. The same agent has a person in the loop on one capability and on the loop on another, which is the point.
What makes a human in the loop a real control?
The person must be able to say no, and must have what they need to say it. A reviewer who sees only a summary and an approve button, under pressure to clear a queue, is not a control; the argument is made in human in the loop is not a control unless the person can say no.
So held drafts should carry their sources, the bound that caused the hold, and anything unusual. Approval queues should be sized so careful review is possible. And rejection should be as easy as approval.
Where Pryme fits
In Pryme, depth is certified per agent and earned through Attest on your own cases, so the in-the-loop or on-the-loop decision follows from evidence. Held verdicts wait for a named person and are filed with their decision. Some customers warrant more human involvement than others; see whether an agent's limits should be the same for every customer. More in our guide to agent governance.
Questions people ask
Is a person on the loop enough for regulated decisions?
It depends on the decision. Where the law requires meaningful human involvement in an individual decision, a person must be in the loop for that decision.
Can an agent move from in the loop to on the loop?
Yes, per capability, when evidence from real cases supports it and risk agrees. If performance drops, it moves back the same day.
Who is accountable when nobody approved an action?
The named owner of the agent and the people who set its depth. Removing per-action approval moves accountability up a level; it does not remove it.