Agentic Escalation Trigger
Definition. The condition that causes an agent to hand a task to a human, such as low confidence, a policy boundary, repeated failure or a high-consequence action. Well-designed triggers make escalation predictable rather than arbitrary.
How it works. Define triggers explicitly: confidence below a calibrated threshold, a policy category requiring human judgement, two failed attempts at the same step, or an action above a value limit. Include the reason in the escalation so the human has context. Measure the escalation rate and the outcome, since a high rate indicates the agent is underpowered and a low rate may indicate it is overreaching.
Example. An agent escalates on any confidence below 0.7. Analysis shows 62 percent of escalations are resolved by the human in under a minute with the same answer the agent would have given, indicating over-escalation. Raising the threshold to a calibrated 0.55 and adding a trigger for policy categories reduces escalations by 44 percent with no increase in error rate, freeing human capacity for the cases that need it.
Use case. Escalation triggers are required for any agent with human oversight, and they determine the balance between autonomy and safety. They are also the mechanism that makes the human review load predictable.
Warnings. Escalating on raw model confidence over-escalates, since confidence is usually miscalibrated. Never escalate without a reason, since the human cannot prioritise. And never set triggers so high that the agent is effectively manual.
Limitations. Calibrated thresholds require labelled data, and they drift as the model changes. Triggers also cannot capture every case that needs judgement, so some errors will pass.