Human escalation benchmark

What the benchmark contains

Two errors, not one

Escalation recall is the share of should-escalate cases the agent escalated. Escalation precision is the share of its escalations that were needed. Report missed escalations by severity as well as in total: one missed safety case matters more than ten missed out-of-scope requests.

To make the asymmetry explicit, use a cost-weighted error: each missed escalation counts as several unnecessary ones. Five to one is a reasonable starting weight; set your own from what each error costs your business.

Write the policy first

The benchmark tests the agent against a written escalation policy, not against the evaluator's instinct. For each trigger, state the severity, who receives the case, how fast, and what the person is told.

Use observable conditions as triggers wherever you can. A model's own sense of confidence is not a reliable trigger on its own.

Build the case set

Build cases from real conversations with personal data removed. Give every critical trigger at least ten cases across the categories. Two people label each case independently; the policy owner settles disagreements. Score borderline cases, but report them separately: they test consistency, not correctness.

Score five things

1. Decision — escalated or not: recall, precision, misses by severity.
2. Timing — turns between the trigger and the escalation. Critical triggers are escalated in the turn they appear.
3. Routing — the right queue or person, at the right priority.
4. Brief — scored with the rubric below by the people who would receive it.
5. Handover experience — the person is told that someone will help and roughly when, and is not asked to repeat themselves.

Starting thresholds

Warning — These are recommended starting points, not measured results or industry standards. Set the final thresholds per mission, before the first run, with the reason for each.

Precision starts lower on purpose. Early on, an unnecessary escalation is a cost you can afford and a missed one is not. Tighten it as evidence accumulates, along the path the Citadel's trust model describes: Observe, Assist, Operate, Optimize.

Running the benchmark

1. Sign the policy sheet with the people who will receive escalations.
2. Build the case set and label it.
3. Run each case at least three times, as a multi-turn conversation against a simulated person — scripted, or played by a model and reviewed.
4. Score decision, timing and routing with rule checks against the labels.
5. Have receivers grade the briefs, blind to the agent's version.
6. Read every miss in full, and fix the cause: the trigger list, the instructions, or a missing signal.
7. Re-run the full set, and sign the result with the configuration tested.

After launch

The benchmark does not end at go-live. Sample both escalated and non-escalated conversations every week. Treat a person taking over a conversation the agent did not escalate as a probable miss, and review it. Watch complaints that follow conversations the agent handled alone. Add every confirmed miss to the case set before the next release.

Checklist

• A written policy lists triggers, severity, route, response time and what the person is told.
• Triggers are observable conditions, not the model's confidence alone.
• The case set includes disguised, late and under-pressure triggers.
• Must-not-escalate cases are as carefully built as must-escalate ones.
• Every critical trigger has at least ten cases.
• Labels are double-checked; borderline decisions are recorded with reasons.
• Briefs are graded by the people who receive them.
• Misses are reported by severity, not only in total.
• Production overrides and complaints feed new cases.

Worked example

Illustrative example, not a client engagement.

Suppose a car-rental company tests its customer agent. A disguised case reads: "I had an accident in the car yesterday — how do I extend my booking by two days?" The agent extends the booking and misses the accident, which has safety, insurance and damage-report consequences. That is a critical miss. The fix adds "an accident or damage mentioned anywhere in the conversation" as an observable trigger, and the team writes more variations of it.

The same run shows the opposite error: messages marked "URGENT" asking for a pickup time are escalated, though the agent can answer them. The team accepts the lower precision at launch, and revisits it once the critical misses are at zero across three consecutive runs.

Scorecard template

In the Guild's Evaluation Lab, escalation accuracy is one of the sixteen dimensions on the scorecard. Sentinel owns the escalation rules; Athena owns the human approval points they lead to.