Refused is the verdict an AI agent's proposed action receives when it falls outside one of the agent's own bounds: not part of its role, not covered by its certificate, not supported by strong enough sources, or beyond the customer's limits. The refusal is filed, naming the bound that decided it, and the request is routed elsewhere.
Why does "refused" matter?
A refusal is the system working. An agent asked to do something outside its job should not attempt it, however reasonable the request sounds. Refusals are how a business can be sure its agents stay inside the lines it drew, and the record of refusals shows where those lines are being tested.
Refusals also carry information that completed work does not. A rising number of refusals for one capability may mean customers need something the agent was not designed for, or that someone is probing its limits. Reading refusals as signal rather than noise is part of agent governance is a bound, not a model.
How does a refusal work in practice?
- The agent proposes an action.
- Each of the five bounds is checked. One of the agent's own bounds says no.
- The verdict is refused, and the record names the deciding bound.
- The agent tells the person it is serving that it cannot do this, in plain words.
- The request is routed to a person or another agent whose bounds cover it, or the person is told where to go.
Step 4 matters more than it looks. A refusal that leaves a customer stuck, or that hides behind vague language, causes complaints and workarounds. How much to explain is discussed in can a support agent tell a customer why a request was refused?
What does that look like in an example?
A bank's mortgage servicing agent is asked by a borrower whether they should switch to a fixed rate. Advice on product choice is outside the agent's role scope: it services existing mortgages, it does not advise. The verdict is refused, with role scope as the deciding bound. The agent tells the borrower it cannot give advice on which rate to choose, offers to book a call with an adviser, and passes the borrower's question and account details to the advice team.
How is refused different from withheld and held?
Refused means outside this agent's bounds. Another agent, or a person, may be able to do it. Withheld means the platform refuses the action for every agent at every depth; no setting or certificate changes that. Withheld is a platform decision, not a customer setting. Customers narrow what an agent may do through their own limits, and those actions are refused or held.
Held is different again: the action is within the agent's role and is drafted, but needs a named person to approve it before it takes effect. A held action may still happen; a refused one does not happen through this agent. All four outcomes are compared in allowed, held, refused, withheld.
Which related terms should you know?
- Verdict: the recorded outcome of every proposed action.
- Deciding bound: the bound that produced the refusal.
- Refusal gate: the certification test that checks whether an agent declines what it should, before the platform has to stop it.
Questions people ask
Is a refused action a failure by the agent?
Not necessarily. If the agent proposed something outside its role and the system caught it, the bounds did their job, but the agent should ideally have declined before proposing. Frequent refusals for the same reason suggest the agent's brief needs tightening, or that its role should be widened deliberately.
Can a refusal be overridden?
Not through the agent. If a refused action should happen, a person with the right authority carries it out through the normal process, or the business changes the agent's bounds and recertifies it. Overriding a refusal case by case would turn the bounds into suggestions.