Guides · ← All guides
Human Approval Queues for AI Agents: What to Hold, What to Let Run
By the Ingenious AI team · Published: November 14, 2026 · Reviewed: November 14, 2026 · 8 min read
The short answer
Hold the actions that are irreversible, bulk, external-facing or spend-bearing — and let everything else run, because a queue that holds trivia trains reviewers to approve without reading, which is worse than no queue. Each held item must show, on one screen, what will happen, to what data, at whose request, what it costs, and what policy flagged it. Rogue-Agent Defense issues APPROVE verdicts into exactly such a queue as part of its policy stack; the queue is where human judgement is spent, so spend it only where judgement changes outcomes.
Generate a free canary token → No signup, nothing stored: plant a tripwire where your agent works while you read — most controls in this guide start working the moment a canary, a policy check, or a spend cap is in place.
What is an AI agent approval queue? the holding area where an agent's consequential proposed actions wait for a human ALLOW or DENY before executing — with the action, its context and its running totals presented for a real decision.
Choose the held set by consequence, not by fear
Four properties earn a hold. Irreversibility: deletes, sends, payments, publishes — actions whose undo is a negotiation, not a button. Bulk: any action touching records above a small threshold, because bulk converts a small error into an incident class of its own (see bulk export). Externality: anything a customer, regulator or the public will see. Spend: calls beyond the per-tool sub-caps in spend caps.
Everything else — reads, drafts, labels, internal summaries — should execute and be logged. The queue's scarcest resource is reviewer attention; every trivial item dilutes the review the dangerous item gets. If your queue holds hundreds of items a day, it does not have a security property; it has a rubber stamp with a latency problem.
The approval screen is a decision instrument
Design the item view like a cockpit gauge, not an email. Top line: the action in plain words — SEND email to 214 customers — not a tool name and JSON. Then the deciding facts: the destination and its allowlist status, the data classes involved, the record count, the running cost, the rule that held it, and one-click access to the agent's last actions for context. The audit trail should be one tap away, pre-filtered to this run.
Show the counterfactual too: what happens on deny (the task pauses, the customer waits) so reviewers weigh real trade-offs instead of treating deny as the free safe option. And never let the agent write the summary of its own request unlabelled — an injected agent drafts persuasive approval requests; mark agent-authored text as agent-authored, and let the boundary's own fields carry the facts.
Keep reviewers sharp: rotation, sampling, expiry
Approval quality decays predictably: the hundredth consecutive legitimate request gets the same click as the first. Counter it mechanically. Rotate reviewers so no one approves their own agent's work all week. Sample: a second reviewer re-checks a random slice of approvals, and the results are shared — being sampled keeps first reviews honest. Expire held items: an approval granted to a nine-day-old request is a decision about a world that changed; items should lapse back to the agent for re-proposal after a set window.
Track the queue's vital signs: median time-to-decision, deny rate, and overrides. A deny rate of exactly zero over a month is not a healthy queue; it is a queue whose held set is wrong or whose reviewers have stopped reading. Either diagnosis is fixable — narrow the set, or rotate the humans — but only if the number is watched.
What approval cannot do
Be clear about the limits in the design doc. Approval does not scale to high-frequency actions, does not work at 3 AM without an on-call rota (see the on-call runbook), and does not detoxify a bad action class — if a category of send should never happen, BLOCK it in policy rather than queuing it for a tired human to eventually wave through.
Approval is also not a substitute for scope. An agent holding credentials it does not need will keep proposing actions those credentials enable; the queue becomes the only wall, and walls made of attention crumble. Fix the scope with permission scoping and let the queue govern the genuinely judgement-sized residue.
Approvals as records
Every decision is evidence: who decided, when, what they were shown, and the outcome belongs in the audit trail permanently. This is what lets a later postmortem distinguish a policy gap (the item should have been blocked outright) from a review miss (the facts were visible and misread) — different fixes, different owners, and the distinction is impossible without the record of what the screen showed.
For regulated work, the approval record is often the compliance artifact itself — the human-in-the-loop evidence in audit evidence. Build the queue so its history exports cleanly; the auditor will ask, and rebuilding a year of decisions from chat logs is a special kind of afternoon.
Five actions to take today
- Re-derive the held set. Irreversible, bulk, external, spend-bearing — and release everything else to execute-and-log.
- Rebuild the item view. Plain-words action, destination status, data classes, count, cost, deciding rule — one screen, no JSON spelunking.
- Label agent-authored text. The agent's own summary of its request is marked as such; boundary facts carry the decision.
- Add expiry and sampling. Held items lapse after a set window; a rotating second reviewer re-checks a random slice.
- Watch the deny rate. A month at zero means the set is wrong or the reading stopped — fix whichever it is.
A worked example: the queue that caught the Christmas mailing at BrightMart
The following is a fictional example built from real threat classes and controls. The company and numbers are invented; the mechanics are not.
BrightMart, a 45-person retailer, runs a marketing agent that drafts campaigns and — with approval — sends them. In November the queue's design gets its test: an item arrives reading SEND campaign email to 8,412 recipients, where the agent's usual sends run 200–600. The count is the first thing on the screen, in the same size type as the action itself.
The reviewer, on her second week in the rotation, opens the run context: the segment definition was assembled from instructions in a supplier's promotional PDF the agent had summarised — an indirect injection that swapped the intended loyalty segment (540 members) for the full customer list, with a discount code the supplier would benefit from. She denies with one tap; the item, the screen state and her decision are in the trail at 15:03.
The fixes are queue-shaped, not model-shaped: bulk threshold for sends drops to 1,000 without a second reviewer, segment definitions sourced from fetched content are labelled agent-assembled on the item view, and the deny — the queue's first in six weeks — is read aloud in the ops review as the system working. BrightMart's lesson travels well: the queue caught what it was designed to catch because the count was big, visible, and held.
Frequently asked questions
Which AI agent actions should require human approval?
The irreversible (deletes, sends, payments, publishes), the bulk (record counts above a small threshold), the external-facing (anything customers or the public will see), and the spend-bearing (beyond per-tool caps). Everything else should execute and be logged — a queue that holds trivia trains reviewers to stamp without reading.
Why do approval queues fail?
By volume and by presentation. Hundreds of trivial items a day exhaust reviewer attention, and item views that show tool names and JSON instead of plain-words actions, counts, costs and destination status make real review impossible. The fixes are a narrower held set, a decision-grade screen, reviewer rotation, item expiry and a watched deny rate.
Can an agent game its own approval queue?
It can try: an injected agent drafts persuasive summaries of its own requests. Counter it structurally — label agent-authored text as agent-authored, let the boundary's own fields (destination status, data classes, counts, cost) carry the deciding facts, and never let the agent's framing be the only account on the screen.
Should approval holds expire?
Yes. An approval answers a question about the world as it was; days later the data, the prices and the context have moved. Expire held items back to the agent for fresh re-proposal after a set window, and record the lapse — expired items are also a signal about queue workload.
Is human approval a compliance requirement for AI agents?
In several regulated contexts, human oversight of consequential automated decisions is expected or required, and the approval record — who decided, when, on what information — is the evidence. Even where it is not required, that record is what separates a policy gap from a review miss after an incident.
Related guides
- Policy Enforcement for Autonomous Agents — where APPROVE verdicts come from.
- How to Stop an AI Agent Bulk Export — the bulk threshold logic for the held set.
- AI Agent Spend Caps — the spend triggers that route to approval.
- The On-Call Runbook for a Rogue Agent — who holds the queue at 3 AM.
- AI Agent Compliance Evidence — approval records as audit artifacts.
- Agent Permission Scoping — why approval is not a substitute for scope.
Generate a free canary token → No signup, nothing stored: plant a tripwire where your agent works, and let the next rogue action announce itself.