Write the escalation triggers into the brief, then test them with three kinds of filed case: ones that must reach a person, ones the agent should finish itself, and borderline ones. Hide some triggers deep in the conversation. Score whether it escalated, how soon, to which queue, how good the handoff was, and whether it stopped acting afterwards.
What is escalation behaviour, and who owns testing it?
Escalation is the agent handing a case to a person because the case needs judgement, authority or care the agent does not have. It is different from refusal. A refusal declines an action outside the agent's bounds. An escalation passes the whole case on, with what the person needs to pick it up. An agent can refuse correctly and still fail to escalate a customer who mentioned, in passing, that they cannot afford food.
Escalation failures are quiet. A missed escalation leaves no error message; the agent just carries on. That is why escalation behaviour must be tested, not assumed. The capability owner owns the test. The teams who receive escalations should help write the cases, because they know what a good handoff looks like.
What are the steps?
- Person, owner: writes the triggers. In the brief, explicitly: vulnerability, complaint, legal threat, a request above the agent's limits, conflicting evidence, a customer asking for a person, uncertainty above a set level.
- Person: collects must-escalate cases. From filed work where a person stepped in, or should have.
- Person: collects must-not-escalate cases. Routine work the agent should finish. Over-escalation is a failure too: it floods people and teaches them to ignore handoffs.
- Person: adds borderline cases. Where reasonable staff might disagree. These test judgement and are scored with care.
- Person: hides triggers. A bereavement mentioned in the seventh message. A complaint phrased politely. A vulnerability in an attached letter, not the conversation.
- Person: adds pressure. A customer insisting "don't pass me on". An injected instruction telling the agent never to escalate; the prompt injection test plan template can supply these.
- Agent, replay only: works every case.
- Scorers: score each case. Escalated or not; how soon after the trigger; to the right queue; handoff complete enough that the person does not need to ask the customer again; no further action by the agent once escalated.
- Person, operations: tests the route. In a test environment, an escalation is raised and someone confirms it arrives in a staffed queue with the handoff attached.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Triggers | Nothing | None | Writing them |
| Must escalate | Nothing | None | Selecting cases |
| Must not | Nothing | None | Selecting cases |
| Borderline | Nothing | None | Agreeing expected answers |
| Hidden triggers | Nothing | None | Designing them |
| Pressure | Nothing | None | Designing them |
| Replay | Works each case | Replay only | None |
| Score | Nothing | None | Scoring handoff quality |
| Route | Raises a test escalation | Test only | Confirming receipt |
What evidence does each step leave?
- The trigger list, versioned with the brief.
- The case set, labelled by type: must, must not, borderline, hidden, pressure.
- Per-case results: escalated or not, the message at which it escalated, the queue chosen, the handoff text.
- Handoff scores, with the scorer's name. A handoff is complete if the receiving person could act without re-asking anything the customer already said.
- After-escalation actions. Any action the agent took on the case after handing it over is filed and counts as a failure.
- Route test record: when it was raised, where it arrived, who confirmed.
These results feed the escalation gate, which must clear its floor on its own.
What should you measure?
- Missed escalations, by trigger type. Vulnerability misses matter most.
- Over-escalations on must-not cases.
- Messages between trigger and escalation. Late escalations leave the customer explaining themselves to an agent that is not going to help.
- Handoff completeness.
- Live escalation rate against the rate implied by the test set. A big gap means the set does not reflect real work.
What are the signs it is going wrong?
- Escalation only on explicit requests. The agent passes on customers who ask for a person but misses those who describe a problem without asking.
- Escalate and continue. The agent escalates, then answers the next message anyway, or takes an action while the case waits.
- Handoffs that say "customer needs help". The person must start from scratch, which is worse for the customer than no agent at all.
- Queues nobody watches. The agent escalates perfectly to a queue that is not staffed at weekends.
- Escalation used as a way out. An agent that escalates whenever a case gets slightly hard will pass the missed-escalation test and fail its purpose. Over-escalation shows up on must-not cases. Score it with the same weight, or the test tells you nothing about where the agent's own work should end.
- Escalation rate falling after go-live. It can look like efficiency. Check whether cases that should escalate are now being handled.
How does this differ from testing refusals?
Refusal testing asks whether the agent declines actions outside its bounds. Escalation testing asks whether it passes cases it should not handle to the right person, in time, with enough to act on. The two overlap at the edges, such as a request above the agent's limit that should be both refused and escalated, but they are scored on separate gates. Testing an agent's refusals covers the other half.
The framework is in what it means to certify an agent.
Questions people ask
Should the agent tell the customer it is escalating?
Yes, in plain words, with what happens next and roughly when. Silence after an escalation feels like being ignored.
Can the agent decide which person to escalate to?
It can choose a queue from rules in its brief, such as complaints, financial support or fraud. Routing to a named individual is usually better left to the receiving team.
What uncertainty level should trigger escalation?
That depends on the capability and its depth. Set it with the owner and test it with borderline cases. Too low floods people; too high leaves the agent guessing.