How to work through an AI agent's held queue each day

Work the queue in two fixed passes, morning and afternoon, starting with items closest to a deadline and those carrying the most value. For each item, open the evidence rather than trusting the summary, check where each fact came from, then approve, amend or reject with a reason. End the day with nothing older than the target.

What is the held queue?

When an agent proposes an action it may not take alone, the action is held: drafted, recorded and waiting for a named person. The held queue is the list of those items. It is a work queue like any other, but with one difference: each item is a decision the organisation has deliberately kept for a person. If the queue is worked badly, that decision is not really being made.

Named approvers work the queue. Their team leader owns its ageing and quality. The agent owner reads it for what it says about the agent: as the held queue tells you more than any dashboard argues, it is where the agent's limits show most clearly.

What are the steps each day?

1. Approver, morning: opens the queue sorted by deadline first, then value or risk, then age.
2. Approver: reads the proposed action and why it was held: which bound decided, and what the agent was unsure of.
3. Approver: opens the evidence. The source documents and records behind the proposal, not just the agent's summary.
4. Approver: checks provenance. Facts from systems of record are sealed; facts from a named party are signed; facts from the customer are declared; anything from email or the web is external. A payment that rests on an external fact needs independent confirmation before release.
5. Approver: decides: approve as proposed, amend and approve, or reject. Every amendment and rejection gets a reason from a short list, plus a note if needed.
6. Agent, depth 3 after approval: carries out the approved action and records the result. A rejected action is closed and, where needed, routed to a person to handle.
7. Approver, afternoon: second pass, focusing on items that arrived during the day and anything carried from the morning.
8. Approver, end of day: carries forward any item still open, with a reason and a named owner for tomorrow.
9. Team leader: checks that nothing is older than the target and that cover is in place for tomorrow.
10. Agent, depth 1: compiles the day's approvals, amendments, rejections and reasons for the weekly review.

A one-page version of this routine is in the held queue daily review checklist.

Which controls make it safe?

Named approvers only. Items can only be released by people on the roster for that action type and value. Evidence: the approver on each decision.

Evidence opened, not just summaries read. Evidence: whether the approver opened the sources, recorded by the tool.

Reasons for amendments and rejections. Evidence: the reason code and note.

Age limits. Each held action type has a maximum age. Evidence: ageing report.

Cover for absence. A deputy for every approver, set before the absence. The arrangements are covered in covering approver absence for an AI agent's held work. Evidence: the cover rota and items routed to deputies.

Separation. No approver releases an item they prepared, or that concerns them. Evidence: the platform's enforcement record.

What are the signs it is going wrong?

Approval is fast and almost universal. Hundreds of items approved unchanged, each in seconds. Either the agent is ready for more depth, or no one is reading. Check with a sample.

The queue ages. Items wait days. Customers wait with them, and the pressure to widen depth "to clear the backlog" grows. That pressure is a staffing problem, not evidence of the agent's readiness.

Bulk approval. An approver selects forty items and approves them together. For low-risk types that might be acceptable by policy; for anything that moves money, it is not a review.

The same rejection reason repeats. If one reason accounts for most rejections, the agent has a fixable problem. Send it to the agent owner.

Items are rejected without a route. A rejected refund is closed, but no one handles the customer's request. Every rejection needs a next step.

The held queue and the exceptions queue blur. Approvers start fixing process exceptions in the held queue, or vice versa. Keep them separate; the exceptions routine is in working an operations exceptions queue.

How much time should it take?

It depends on volume and item type. Measure it: time per item from the tool, by type. If the queue takes more time than doing the work by hand would, either the agent's drafts are not good enough or the capability belongs at a different depth. That is a decision for the agent owner, using the evidence, not a reason to rush reviews.

Where Pryme fits

In Pryme, every held item shows the bound that held it and the provenance stamp on each fact, so approvers can see at a glance what came from the system of record. Only named approvers for that action type and band can release it, and every decision is filed. The sequence follows train, shadow, certify, widen.