Escalation

Escalation is the AI's act of stopping autonomous execution and routing a case to a human (or higher-tier agent) because confidence is low or stakes are high.

Escalation is how AI systems stay safe at the edges. Triggers: low model confidence, repeated tool failures, dollar amounts above a threshold, customer language indicating distress, novel categories the agent wasn't trained on. A good escalation includes the full case, the agent's working hypothesis, and the specific reason for stopping.

Example

AI Support Rep sees the word 'lawsuit' in a ticket, immediately stops auto-reply, flags red, and routes to a human with full transcript and explanation.

How OpenLabor uses it

OpenLabor lets you define escalation triggers per employee — keywords, dollar caps, low-confidence thresholds.

How do I tune escalation thresholds?

Start aggressive (escalate often) and relax as the human reviewers report most escalations were needless.

Related: hand-off, approval-gate, guardrails.

AI Labor Glossary