Edge Cases and Escalation

The maturity of a Voice AI Agent shows in how it meets the non standard moment, not in how well it runs the standard one. Three categories need particular care:

1. Comprehension failureIf the customer is misunderstood two or three times in a row, the AI Agent should change its approach rather than repeat the same strategy, and hand over to a person past a defined threshold. A concrete example: the customer says "I have a problem with last month's invoice", the AI Agent asks for the invoice number. The customer says "I don't know it, I mean last month's invoice". A badly designed AI Agent asks the same question again, in the same words. A well designed AI Agent changes strategy here: it searches by national ID or phone number instead of invoice number, or hands over to a person on the second failure. Persistence is not a solution.
2. Rising emotional intensityWhen a customer becomes angry, several signals can detect it: speech speeding up, sentences getting shorter, volume rising, more negative word choices. Once detected the AI Agent should adjust its tone, becoming calmer and using less filler, and route to a person where needed, rather than stalling with empty empathy lines like "I understand how you feel".
3. Out of scope subjectsLegal advice, medical diagnosis, financial commitments beyond its authority, sensitive personal circumstances. That last category demands particular care: if a customer mentions a bereavement or a serious illness during a call, the AI Agent must not process it as an ordinary objection or a reason for delay. It should route to the right process, whether that is a person, the relevant department or a special procedure. On these subjects an AI Agent should never speculate. A content safety layer makes sure conversations like these are routed to the right flow from the start.

The escalation rule should not be uniform. The three categories above are about what happens inside a conversation. But escalation also has a rule design dimension: the same threshold and the same handover point should not apply everywhere.

1. By persona and roleEach Digital Employee should have its own threshold and its own handover point, set by its risk profile. A persona working in a sensitive, regulated area such as TAHSİLDAR should hand over to a person even at a relatively high confidence level, while a simple appointment reminder AI Agent can run with far more autonomy. The handover point changes with the persona too: TAHSİLDAR hands over to the collections team, GÜVEN to a claims assessor, ADİL to the crisis coordination team. Instead of a single "hand over to a human" rule, each persona should have its own handover map.
2. By transaction value and riskSome escalations come from the size of the task rather than the course of the conversation. A small balance enquiry can run fully automatically, while a high value payment plan or an account closure request should move to a person automatically even when the call is going well. The trigger here is not confusion on the AI Agent's part. It is the size of the risk the task carries.
3. By system or integration failureIf a back end system such as core banking or CRM does not respond or returns an error, it is wrong for the AI Agent to stall with "one moment, let me try again". When a technical fault is detected the escalation should be immediate and clear. The AI Agent should not behave as though it is solving something it cannot solve.
4. By repeat contactIf the same customer calls about the same subject several times in a short period, for instance for the third time this week, that is a signal no matter how well the current call is going. It means automation has not solved the problem. Here the right move is to route to a person straight away, without waiting for a threshold.

A common mistake: measuring escalation quality by how warm the handover sentence sounds. Teams often spend a great deal of time perfecting it ("I'm transferring you to an experienced colleague right away, please don't worry..."), when what actually matters is not how well that sentence reads but whether the system correctly detects when it should hand over. A perfect handover sentence does not make up for a late handover. The right question is not "how good is the handover line" but "how early and how accurately is the threshold detected".

CBOT treats escalation as a planned part of the design rather than an error state: any interaction that falls below a defined confidence threshold is routed automatically to a person or to LiveChat before the customer feels misunderstood. A well built escalation pattern protects trust. An AI Agent does not have to know everything. It has to know when it does not know.