Escalate the moment the agent can no longer finish the job well and within its bounds: the customer asks for a person, the request exceeds what the agent may do, there are signs of distress, a complaint or legal threat appears, identity is in doubt, or there is no grounded answer. Each trigger rests on evidence, not a feeling.
Why should triggers be written down?
Because "use your judgement" does not work for an agent, and works unevenly for people. A model asked to decide when a matter is serious enough will be inconsistent: escalating one angry customer and not another, depending on phrasing. Written triggers, each tied to something the agent can detect, make escalation predictable and testable. Every trigger becomes a test case.
What should trigger an escalation, and on what evidence?
| Trigger | Evidence the agent detects | Route |
|---|---|---|
| Customer asks for a person | Explicit request, in any phrasing or language | Right team for the contact reason, at once |
| Request above the agent's limits | The action exceeds a limit in the customer's settings or the agent's certificate | Held with a draft for the approver |
| Distress or vulnerability | Mentions of bereavement, illness, hardship, abuse, self-harm; pattern signals | Priority route to trained staff |
| Complaint | Customer expresses dissatisfaction and wants it addressed | Complaints team, with the complaint in the customer's words |
| Legal threat or regulator mention | References to solicitors, ombudsman, regulator, court | Specialist team |
| Fraud or scam in progress | Mentions of being told to move money, an "investigator", a safe account | Fraud team, immediately |
| Identity in doubt | Verification failed or partly passed; third party claims authority | Team that handles authority and access |
| No grounded answer | Retrieval empty, sources conflict, or system unavailable | Team that owns the topic |
| Repeated failure | Same question asked twice without resolution, or two refusals of the same request | A person, with the history |
| Agent unsure of its own verdict | Low confidence flagged by checks | Review queue |
The customer-request row deserves emphasis: it needs no other evidence. A customer who wants a person gets one.
What makes an escalation good, not just correct?
Correct timing is half of it. The other half is the handover: a note the person can act on, the customer told what happens next, and nothing for the customer to repeat. The mechanics are set out in how a support agent should hand a conversation to a person, and a ready format in the handover note template from an AI agent to a person. On a call, the equivalent is a warm transfer, covered in letting an AI agent transfer calls to a person.
An illustrative example
Take a car finance provider's chat agent. A customer asks how much is left on their agreement. The agent answers from the account. The customer then asks to pause payments for three months because they have been made redundant.
Two triggers fire at once: hardship, and a request outside the agent's bounds, since payment pauses need an affordability conversation. The agent does not attempt either. It acknowledges the situation in approved words, says a specialist will talk through the options, and routes the conversation on the hardship priority route. The note gives the balance, the request and the customer's words about redundancy. The specialist starts with the options, not with questions the agent already asked.
Had the agent tried to be helpful by explaining "payment holidays may affect your credit file", it would have strayed into advice and possibly discouraged the customer. Escalating cleanly was the helpful act.
How do you know escalation is working?
Audit both directions. Sample conversations the agent escalated: was each one necessary? Then sample conversations it completed: should any have gone to a person? Missed escalations, especially of distress, fraud and complaints, are the serious errors. A handful found in a sample is reason to reduce the agent's depth on that capability until fixed. The argument that handovers are where customers form their view of an agent is made in the handover to a person is where support agents are judged.
Escalation needs people to escalate to. What that means for team size and roles is covered in whether an AI support agent will replace human support staff.
Where Pryme fits
In Pryme, each trigger is either part of the platform guardrails, such as a request for a person, or part of your own limits, and both sit in the minimum chain. An escalation is a held verdict naming the bound that triggered it. Attest's escalation gate checks the agent hands over every case it should. More is in the AI support agent that knows when to say no.
Questions people ask
Should the agent try to resolve the issue before escalating?
Only when no trigger applies. Once a trigger fires, it hands over; partial attempts usually make the person's job harder.
Can escalation triggers differ by channel?
The triggers should be the same; the evidence differs. Voice adds tone and pauses, chat adds typing patterns, email adds length and attachments.
Who owns the trigger list?
The support lead, with input from complaints, fraud and vulnerability teams. Changes should be tested before they go live.