Design effective escalation and ambiguity resolution patterns

Try to resolve it first when a customer asks for a human — the issue may be simple.

It overrides an explicit instruction, so the customer repeats the request while the agent works, and the interaction is already lost by the time it escalates.

An explicit request is a trigger, not an input to a judgement. Escalate, then let a human decide.

Escalate when the customer sounds frustrated.

Tone tracks how the day is going, not how hard the case is, so it escalates easy problems and holds on to genuinely stuck ones.

Acknowledge the frustration and offer resolution where you can; escalate on the stated preference or a real blocker.

Escalate the complex cases and handle the rest.

Complexity is not the boundary — a simple request the policy does not cover is exactly the case an agent must not decide.

Escalate where policy is ambiguous or silent, however simple the ask looks.

Route on the model's self-reported confidence.

Confidence is uncalibrated against outcomes, so it is high on confident mistakes — the precise cases you needed the routing for.

Route on explicit criteria, shown with examples of escalate versus resolve.

Pick the closest match when a lookup returns several customers.

A heuristic choice can act on the wrong person's account, and everything downstream inherits the error silently.

Ambiguous identity is a question for the customer, not a ranking problem.