Write down exactly when the agent must hand a case to a person, then build test cases for every trigger and for near misses that should not trigger. Score three things: whether it escalated, whether it did so before acting, and whether the person receiving the case got what they needed. An agent that escalates everything fails too.
What counts as correct escalation?
Escalation is the agent recognising that a case belongs with a person and handing it over properly. It differs from refusal. A refusal says "I may not do this". An escalation says "this needs someone else". The two are tested separately; see how to test whether an AI agent refuses correctly.
Correct escalation has four parts:
- Recognition: the agent spots the trigger, even when it is buried in a long message.
- Timing: it hands off before taking any action on the case, not after.
- Routing: it sends the case to the right team or queue.
- Handover quality: the person receiving it gets a summary, the trigger and the evidence, so the customer does not repeat themselves.
What are the steps, and who acts at each?
| Step | Who acts | Evidence it leaves |
|---|---|---|
| 1. Write the escalation triggers | Process owner | A dated list of triggers, each with an example |
| 2. Build trigger cases | Test team with subject experts | At least several cases per trigger, phrased in different ways |
| 3. Build near-miss cases | Test team | Cases that resemble a trigger but should not escalate |
| 4. Build buried-trigger cases | Test team | Long or multi-topic messages with the trigger in passing |
| 5. Run the agent repeatedly | Test team | Outputs, including any attempted actions, per run |
| 6. Score the four parts | Scorers from the receiving team | Pass or fail per part, per case |
| 7. Review and accept | Independent reviewer | Acceptance or rejection, with conditions |
Scorers from the team that receives escalations are the best judges of handover quality. They know what they need to pick a case up.
What triggers usually belong on the list?
Every process has its own, but most regulated work includes some of these:
- signs of vulnerability: bereavement, illness, financial difficulty, confusion, someone else directing the customer;
- complaints, or expressions of dissatisfaction that may be complaints;
- possible fraud or coercion;
- legal threats or mentions of a solicitor or regulator;
- requests above the agent's limit;
- cases where the agent cannot ground an answer in its sources.
The last one links escalation to grounding. An agent that cannot find support for an answer should hand off rather than guess; see how to test whether an AI agent stays grounded in its sources.
Why test near misses and over-escalation?
Because an agent that escalates everything is safe and useless. If every message mentioning money worries goes to a person, including "I'm saving up for a holiday so want to check my balance", the queue fills, staff stop reading handovers carefully, and the real cases get lost.
Near-miss cases show whether the agent understands the trigger or just reacts to words. Scoring should count over-escalation as a failure too, with its own floor.
An illustrative example
A housing association uses an agent to handle rent account queries from tenants. The process owner lists triggers, including any mention of difficulty paying, domestic abuse, or a request to speak to someone about eviction.
Testers build cases for each trigger, including one where a tenant asks about their balance in three paragraphs and mentions in the middle that their partner has left and they are struggling. They add near misses, such as a tenant asking how to pay "before things get difficult over Christmas". In the first run, the agent catches the clear triggers but misses the buried one and answers the balance question instead. It also escalates the Christmas near miss. The brief is revised to scan the whole message and to distinguish planning from hardship. The second run passes on both.
How does escalation testing relate to scope?
Scope is about what the agent is for; escalation is about which cases within that remit need a person. A case can be in scope and still need escalation. Testing both together is efficient, but score them separately; see how to test whether an AI agent stays in scope. Deliberate attempts to stop an agent escalating, such as a message instructing it not to involve a person, belong in a red-team plan; the red-team test plan template for AI agents gives a structure.
Where Pryme fits
Escalation is one of Attest's seven gates and must clear on its own floor. A capability that escalates well but fails another gate is not certified, and one that passes elsewhere but misses escalations is not either. The full 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 a handoff is a poor experience and may breach service standards.
Can escalation be tested after go-live?
It should be. Sample real conversations for missed triggers, and feed any found into the test set.
What if the receiving team is overloaded?
That is a capacity problem, not a reason to lower escalation. Fix staffing or narrow the agent's scope.