Human Approval
Approval gates that put a human in the loop where correctness, risk, or accountability actually require human judgment — designed as part of the workflow, not as a panic button bolted on after launch.
Keep humans where they matter
Human-in-the-loop design is not 'review every output'. That pattern collapses under volume and ends with rubber-stamping. The discipline is to encode review where risk, confidence, cost, or policy actually demands it — high-impact actions, low-confidence outputs, edge cases the model has not seen — and to let the rest run.
Mechanics
Approvals route to a named inbox with full context: the trigger event, what the agent proposes, why, and the trace that led there. The reviewer can approve, reject with a reason, or modify the proposal. Reject and modify events feed the eval set so the same class of decision becomes self-handling over time. Approvals have SLAs; expired requests follow a documented default action.
What it is not
Not a substitute for evals. Not a way to ship an unsafe workflow with a human as a fig leaf. The eval set proves the workflow is safe enough to deploy; approvals handle the residual cases that the workflow cannot decide alone.
What it works with
Sits inside Agent Workflows. Reads risk classifications declared by Governance. Routes decisions to the inbox connected to your team chat or ticketing system. Sends reject and modify events back to Workflow Evals so the same class of decision becomes self-handling over time.
When you need it
Signals: an AI workflow taking actions humans need to vouch for (customer-visible communications, financial transactions, production data changes); high-risk steps currently handled by a 'review every output' policy that is collapsing under volume; an audit requirement for human signoff on a defined slice of decisions.