A human gate is a point in an agent's workflow where an action stops and waits for a named person to approve, edit or reject it before it takes effect. It is how a held verdict is put into practice. It is placed by rule, on a capability or above a limit, rather than on everything the agent does.
Why does a human gate matter?
Some decisions should be made by a person even when an agent could make them well. They carry legal effect, they are hard to undo, or the business wants a person accountable for each one. A human gate is how that preference is enforced. Without it, "a person reviews this" is a policy statement that may or may not happen.
The value of a gate depends on where it sits. Too few, and the agent takes decisions nobody intended to delegate. Too many, and reviewers face a queue so long they stop reading. Placing gates by capability and by the customer's limits, rather than on every step, is part of the approach described in agent governance is a bound, not a model.
How does a human gate work in practice?
A useful gate has five parts:
- A trigger: the rule that sends an action to the gate, such as depth 2 for this capability or an amount above the customer's limit.
- A named approver: a specific person or rota, not a shared mailbox.
- The evidence: the agent's draft, its sources and its reasoning, in one view.
- A real choice: approve, edit or reject, with rejection as easy as approval.
- A record: the decision, the person and the time, filed against the draft.
Setting one up is covered in how to add a human gate to an agent workflow. Whether every agent needs one at all is discussed in should every AI agent have a human in the loop?
What does that look like in an example?
A pensions administrator's agent handles transfer requests. It may check documents and calculate values at depth 3, but the transfer itself moves a member's savings out of the scheme and is hard to undo. The firm puts a human gate on every transfer out. The agent prepares the pack: the member's request, the receiving scheme's details checked against a register, the transfer value and any warning signs. A named administrator reviews and releases it.
How is a human gate different from a refusal gate or a stable gate?
The names are similar, but they work at different times. A human gate operates at the moment of action, every time, in live work. A refusal gate and a stable gate are certification tests: they run when the agent is being certified, checking whether it declines what it should and whether it gives the same answer to the same question across runs. Certification gates decide what depth the agent earns. Human gates apply to actions above that depth or above the customer's limits.
A human gate is also more specific than "human oversight" in general. Oversight can mean monitoring, sampling or the ability to stop a system. A human gate means one action waits for one person. Supervisors and frameworks expect more than gates alone, as covered in NIST AI RMF and human oversight of AI agents.
Which related terms should you know?
- Held: the verdict for an action waiting at a human gate.
- Named approver: the identified person who decides at the gate.
- Human-in-the-loop: a design where a person takes part in each decision, which a human gate implements for a defined set of actions.
Questions people ask
Should the approver at a human gate see the agent's reasoning?
Yes, but presented as evidence rather than a recommendation to accept. Show the sources and the facts first, then the agent's proposal. Approvers who see only a confident conclusion tend to agree with it.
How do you know a human gate is working?
Look at what approvers do. A gate where every item is approved in seconds and nothing is ever edited or rejected is probably not being used as a decision. Healthy gates show some edits, some rejections, and review times that vary with how complex the item is.