Set response-time targets for held actions by two things: the risk of the action and the time of day it arrives. A drafted reply can wait hours; a held payment or a fraud hold cannot. This template gives each class of held action a target in core hours, extended hours and overnight, an escalation point, and a cover role.
What is the template for?
A held action is drafted and waiting for a named person. Nothing happens until that person decides, and that is the point of holding it. But a hold with no clock turns into a backlog, and a backlog turns into pressure to approve in bulk. Targets keep the queue honest without changing the rule that matters most: a missed target never releases an action. It escalates to another person.
Use the template before an agent starts holding work, and revisit it when the agent gains a capability, when its depth changes, or when the record shows targets missed for a whole class. The term itself is defined in what is a response-time target?, and the method for choosing numbers is in how to set response-time targets for an AI agent review queue.
How do you fill it in?
- Held action class: group held actions by consequence, not by agent. A held refund from two different agents belongs in one class.
- Depth and risk: depth 4 actions (money or hard to undo) usually carry the tightest targets. Customer harm, regulatory deadlines and fraud exposure raise the class.
- Core, extended and overnight targets: define your hours first. Many teams use core hours, an extended evening window, and overnight with weekends.
- Escalate after: the point at which an unopened item goes to the next named person.
- Cover role: who holds the item if the first approver is absent. They must hold the same certificate.
- If the target is missed: what the business does, which is never "auto-approve".
- Record check: what the weekly review looks for in the record for this class.
The template
| Held action class | Typical depth | Risk | Core hours target | Extended hours target | Overnight and weekend target | Escalate after | Cover role | If missed | Record check |
|---|---|---|---|---|---|---|---|---|---|
Add one line under the table defining your core, extended and overnight hours, and one naming the owner.
A worked example
A hypothetical retail bank's operations team handles four classes of held action. Targets are sample values.
| Held action class | Depth | Risk | Core hours | Extended hours | Overnight and weekend | Escalate after | Cover role | If missed | Record check |
|---|---|---|---|---|---|---|---|---|---|
| Payment held for suspected fraud | 4 | High | 15 minutes | 30 minutes | 1 hour, on-call analyst | Half the target | On-call fraud lead | Payment stays held; customer told | Every miss reviewed |
| Release of a large outbound payment | 4 | High | 1 hour | 2 hours | Next core hours unless customer flags urgency | Half the target | Second certified approver | Stays held | Misses by approver |
| Customer complaint final response | 2 | Medium | 1 working day | Next core hours | Next core hours | 4 hours before deadline | Team lead | Escalate to complaints manager | Deadline breaches |
| Drafted reply to a routine query | 2 | Low | 4 hours | Next core hours | Next core hours | End of day | Any certified reviewer | Agent's draft waits | Ageing report weekly |
The fraud row shows why time of day matters. Fraud does not keep office hours, so the overnight target is short and a named on-call analyst carries it. The large payment row shows the opposite choice: outside core hours, the bank prefers a delay to a tired approver, unless the customer has flagged urgency, in which case the on-call approver picks it up.
Training follows from the targets. An overnight fraud analyst decides alone, so they need to read a held payment fast and recognise pressure, including a caller insisting the payment is urgent. Two sector views of what that looks like are how internal auditors handle an action an AI agent has held for them and how AML investigators handle an action an AI agent has held. When a missed target is part of something larger, hand over to your AI incident response process. The skills each role needs are in AI agent skills by role.
Where the Academy fits
The Operator track at Pryme Intelligence Academy includes modules on human review, reading the workspace and running the rhythm, which is where response targets live. The practicum is two weeks of real review decisions, with at least 95 in every 100 correct and no missed red line, so a certified operator has already worked held items under real time pressure.
Questions people ask
Should a held action ever approve itself when the target is missed?
No. A timeout that releases the action defeats the hold. Missed targets escalate to another person, and the item stays held until someone decides.
Who sets the targets?
The operations manager, with the agent owner and risk. Targets for customer-facing classes should also reflect any deadline the business owes customers.
How do we know the targets are realistic?
Read the record. If one class misses its target every week, either the target is wrong or the rota is thin. Fix the cause, not the number.