Name the people, match their authority to the actions, and cap their load. Every held action waits for a specific person who already holds the mandate for that action. Each approver is trained on what the agent can get wrong, has a limit on items per hour, and has a named deputy for when they are away.
What is in the held queue, and why does staffing it matter?
A held action is one the agent has prepared and is permitted to propose, but not to complete: a refund above the customer's limit, a payment to a new beneficiary, a letter that changes a customer's terms. The agent has done the work. A person decides.
It is not the same as an exception queue, which holds cases the agent could not process at all, or a general work queue. Held items are mostly ready to go, which is what makes them dangerous. A tired reviewer facing a queue of plausible, well-formatted proposals starts approving on sight. The control then exists on paper only.
Who should approve held actions?
| Kind of held action | Who approves | Why |
|---|---|---|
| Payments and refunds above limits | Staff with an existing payment mandate at that amount | The agent does not change who may authorise money |
| Changes to customer terms or plans | The team that owns that decision today | Accountability stays where it was |
| Customer communications outside approved wording | A senior handler or team lead | Tone and conduct risk |
| Actions in a regulated decision, such as a credit decline | The named decision-maker under your policy | Regulatory accountability is personal |
| Anything unusual or outside normal patterns | A specialist, with an escalation path | Judgement, not throughput |
The rule underneath the table: an agent never creates a new kind of approver. If nobody in your organisation could approve that action before the agent arrived, the agent should not be proposing it.
How do you set up and run the staffing?
- Process owner: maps each held action type to an approver role using existing mandates. Evidence: a role-to-action map.
- Team lead: names individuals and deputies for each role and each shift. Evidence: a rota.
- Training lead: trains approvers on the agent's scope, typical errors and how to read its sources. This is part of the AI literacy duty discussed in Article 4 for staff who approve agent actions. Evidence: completion records.
- Team lead: sets a load limit in items per hour per person, based on how long a proper review takes. Evidence: the stated limit and queue reports against it.
- Agent: routes each held item to the named role, with the facts, their sources and the bound that caused the hold. Evidence: a filed verdict per item.
- Approver: approves, rejects or returns with a reason. Evidence: the decision, the person and the time, filed.
What does a well-staffed queue look like?
Say a payments operations team at a business bank has an agent preparing outbound payments from approved invoices. Payments under the customer's own limit go at depth 4 where certified; anything above it, or to a beneficiary added in the last few days, is held.
On a normal day 150 items are held. Reviews take about three minutes when done properly, so the team staffs eight reviewer-hours across two people with separate payment mandates, plus a deputy on call. They cap each person at 15 items an hour. When the queue passes the cap, the lead calls in the deputy instead of asking people to go faster. Once a week, a second person resamples 20 approved items to check that approvals involved a real look at the sources. The daily review checklist for the held queue is the routine they follow.
Which controls keep approvals meaningful?
- Authority follows existing mandates, never a new role invented for the agent.
- Each held item shows its sources and their provenance, so approvers can see if anything rests on an unverified document.
- Load limits enforced by the rota, not by goodwill.
- Sampling of approved items to catch rubber-stamping.
- Rejections reviewed weekly, because they show where the agent's brief or bounds need work.
Exceptions need their own handling, covered in working an operations exceptions queue with an AI agent. The full sequence is in our guide to deploying an AI agent.
Where Pryme fits
In Pryme, a held verdict means the action is drafted and waiting for a named person, not for whoever opens the queue first. The verdict shows the facts, where each came from, and which bound caused the hold, whether the agent's certificate, its role scope or the customer's own limit. The approver's decision is filed in the same append-only record.
Questions people ask
Can the same person approve and also own the agent?
They can approve items, but someone else should sample their approvals. An owner has an interest in the agent looking good.
What happens to held items at the end of the day?
They stay held until a named person decides, and they show their age in the queue. Items should never be released automatically because time has passed.
Should approvers see the agent's reasoning?
They should see the facts, their sources and the bound that caused the hold. Reasoning helps, but approvers should check the facts themselves rather than the agent's account of them.