A held action is something an AI agent has prepared but is not entitled to complete on its own, so it waits for a named person to decide. The agent has drafted the payment, the letter or the account change. The person approves, corrects or declines it, and their decision is filed with the agent's original proposal.
Why do agents hold actions?
Every action an agent proposes meets five bounds at once: its role scope, its certificate, the strength of its sources, the customer's own limits and the platform guardrails. The lowest bound wins. If one of them says the agent may draft but not execute, the verdict is held.
That happens for ordinary reasons. The amount is above the limit the customer set. The capability is certified only to depth 2. A key fact arrived from an external source, such as an emailed invoice, which cannot by itself support a payment. Holding is not a failure. It is the agent doing what it was designed to do when it reaches the edge of its authority.
What does the person see?
A held item shows the agent's proposed action, the facts it relied on with the provenance of each, the bound that caused the hold, and who it is waiting for. The reviewer's job is to check whether the facts are true and sufficient, and whether the proposed action is right. A KYC analyst, for example, focuses on whether identity evidence is sealed, signed or merely declared, as covered in how KYC analysts handle an action an AI agent has held for them.
An example
Take a hypothetical wealth manager. A client emails asking to move a large sum to a new external account. The agent drafts the transfer but holds it: the amount exceeds the firm's limit, and the new account details came only from the client's email. The adviser sees both reasons. She calls the client on the number already on file, confirms the request and the destination, and approves the transfer with a note of the call. Had the client not recognised the request, she would have declined it and escalated to fraud. The adviser's side of this is covered in how wealth advisers handle an action an AI agent has held for them.
How is a held action different from a refused or withheld one?
| Verdict | What it means | What a person can do |
|---|---|---|
| Held | Drafted, waiting for a named person | Approve, correct or decline |
| Refused | Outside the agent's bounds | Act themselves, through their own authority |
| Withheld | Refused on purpose at the platform floor, for every agent | Act outside the agent if the business permits |
The key difference is that a held action is ready to go and needs only a decision. A refused action never gets drafted for execution. A withheld action, such as reversing a settled payment on Pryme's own rail, is one no agent will take at any depth.
A held action also differs from an irreversible action, though the two often coincide: actions that are hard to undo are among the most commonly held, precisely because a mistake cannot be corrected later.
How fast should held actions be decided?
Fast enough that holding does not become a reason to loosen bounds. Teams set targets by type: a held payment might need a decision within hours, a held letter within a day. The response-time target template for held actions gives a structure. For the shorter definition of the verdict itself, see held: what it means for AI agents, and for the roles that decide held work, the guide to AI agent skills by role.
Where the Academy fits
Deciding held actions is the centre of Pryme Intelligence Academy's Operator track. Learners study how a governed agent decides and how human review works, then complete a two-week practicum of real review decisions. A business can choose to tie the right to approve held actions at each depth to a current Operator certificate.
Questions people ask
Can a held action be approved by the person who asked for it?
It should not be. Separation of duties means the person who requested or prepared an action is not the one who approves it.
What happens if nobody decides a held action?
It waits, and nothing happens on its own. That is why held queues need rotas, response targets and cover arrangements.