← Insights
Engineering3 min
What an escalation contract actually is
How should an AI agent decide when to hand a task to a human?
Against a defined confidence floor, set before deployment and enforced in code. Below the floor the agent escalates with full context attached rather than answering. An agent that never escalates is not confident — it is unmonitored.
Escalation is a design decision, not a fallback
Most deployments treat handover as what happens when something breaks. It should be a defined path with a threshold, a destination, a payload and a response expectation — specified in the same document as everything the agent is supposed to do.
What the contract specifies
- The confidence floor, and how confidence is computed.
- What travels with the escalation — the full context, not a bare ticket.
- Who receives it, and what the response window is.
- What the agent tells the customer while it waits.
- Which categories always escalate regardless of confidence.
The metric nobody publishes
Escalation rate belongs on the case study next to the success figures. A system reporting a two percent escalation rate on genuinely varied work is either extraordinary or not measuring. Publishing the number invites that question, which is precisely why it builds trust.