Held is the verdict an AI agent's action receives when it is drafted in full but cannot take effect until a named person approves it. The agent has done the work; one of its bounds, usually depth or the customer's limits, requires a person to decide. The decision is filed with the approver's name.
Why does "held" matter?
Held is where agents and people meet. It lets a business use an agent for work it is not yet ready to let the agent complete: the agent gathers the facts, prepares the action and explains it, and a person makes the call. Done well, this gives most of the speed of automation with the judgement still in human hands.
Held also produces evidence. Every held item that a person approves unchanged is a data point that the agent's draft was right. Every edit or rejection is a data point that it was not. Over time, that record informs whether the agent's depth for that capability should rise. The verdict is one of the four set out in agent governance is a bound, not a model.
How does it work in practice?
- The agent drafts the action fully: the refund, the letter, the payment, the change.
- The verdict is held, and the record names the bound that required a person.
- The item is routed to a named approver, not a shared inbox, with the evidence and the agent's reasoning.
- The person approves, edits or rejects, and the decision is filed against the original draft.
- Approved actions execute; rejected ones are closed with a reason.
Who should sit at the end of that route, and how much authority they need, is the subject of who approves an AI agent's held actions?
What does that look like in an example?
A telecoms provider's billing agent is certified at depth 4 for goodwill credits, and the customer's limit for unreviewed credits is 20. A customer with a long outage asks for compensation. The agent calculates 45 under the provider's outage policy, drafts the credit with the outage log attached, and the verdict is held. A team leader opens the item, checks the outage dates, and approves it in under a minute. The record shows the draft, the approver, the time and that no edit was made.
How is held different from refused?
Held means the action is inside the agent's role and capability but needs a person's decision before it takes effect. Refused means the action is outside the agent's bounds altogether: not its job, not certified, or not supported by sources strong enough to act on. A held action can still happen. A refused action does not happen through this agent, though the request may be routed to someone who can handle it.
There is a third neighbour, withheld, which applies when the platform never lets any agent take the action. The four verdicts are compared in allowed, held, refused, withheld.
What is the main risk with held items?
Rubber-stamping. When an approver sees a long queue of well-prepared drafts that are almost always right, approval becomes a click rather than a decision. The control looks present and is not. Signs include approvals in a few seconds, approval rates that never vary, and approvers clearing items outside their expertise. The problem and its remedies are discussed in rubber-stamping is the hidden risk in held items.
Questions people ask
What happens if nobody approves a held item?
It should not execute by default. Held items need a service level and an escalation route, so an unapproved item moves to a second named approver after a set time. Silent expiry is acceptable for low-stakes items if the customer is told; automatic approval on timeout defeats the point of holding.
Can an approver change a held action before approving it?
Yes, and the change should be filed alongside the original draft. Edits are valuable evidence: frequent edits on one type of action suggest the agent's brief or sources need attention before its depth for that capability is reconsidered.