Human escalation for AI agents: send the decision, not the mess

Human escalation is the handoff an AI agent makes when it reaches a decision or action outside its authority or available evidence. A useful escalation gives the person a concrete choice, the relevant facts and the prepared work. It should reduce the person's effort rather than transfer an unstructured problem.

Escalate the actual uncertainty

Imagine a customer requesting a service change not covered by the current agreement. The assistant can retrieve the agreement, summarize the request and prepare a response. The open decision is whether to approve the change, not whether the assistant should keep thinking.

Avoid broad questions when a specific missing fact would resolve the task. Also avoid repeatedly asking for approval that already clearly covers the same action.

Prepare the decision packet

• Situation: what changed or cannot be established.
• Evidence: source links and the relevant verified facts.
• Proposed action: the exact draft or change to review.
• Consequence: what approval would cause.
• Owner: the person authorized to decide and the fallback if unavailable.

Do not invent urgency. Use a deadline only when it comes from the task or an agreed operating rule.

Preserve work while waiting

Store the draft and task state in a shared location. Mark the task as awaiting a specific decision, not generically blocked. When an answer arrives, confirm it applies to the current version of the proposed action.

If the situation changes while waiting, show the change before executing. Approval of yesterday's recipient or amount may not cover today's version.

Learn from repeated escalations

Review patterns with the process owner. Some indicate missing documentation; others identify decisions that should remain human. Do not optimize only for fewer escalations. The useful goal is the right level of independent work with timely, understandable handoffs when judgment or authority is missing.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.