Direct answer
Clause 5.2.2.6 of GB/T 47746—2026 lists five situations in which the service must transfer automatically to a human agent: (a) the interaction has reached the failure threshold and the system still cannot recognise the request; (b) the customer explicitly refuses intelligent service; (c) the conversation involves information security; (d) the situation involves an emergency affecting personal or property safety; (e) the intelligent response times out — the standard's own example being a delay of more than three minutes.
The five scenarios, as written
| Item | Situation |
|---|---|
| a) | The interaction has reached the failure threshold and the system is still unable to recognise the request. |
| b) | The customer explicitly declines intelligent service. |
| c) | The conversation involves information security. |
| d) | The situation involves an emergency affecting personal or property safety. |
| e) | The intelligent response times out — the standard's example being over 3 minutes. |
These are the situations where the transfer is not a matter of preference. They are the operating conditions your AI customer service has to implement.
What each one demands in practice
a) Failure threshold reached. The service must count its own failed attempts and give up at a defined point. This is unimplementable without a threshold value, which means the value has to exist somewhere in your configuration.
b) Explicit refusal. If a customer asks for a person, that ends the AI interaction. Implementing this is harder than it sounds: it requires recognising a request to escalate as distinct from a question about escalation, and it must work in the customer's own phrasing.
c) Information security. A conversation touching security matters moves to a human. This is the scenario most often missing from off-the-shelf configurations, because it depends on recognising subject matter rather than on a keyword.
d) Emergency affecting personal or property safety. Safety-critical content escalates immediately, without waiting for a failure count.
e) Timeout. The response has to arrive within a defined window. The standard gives three minutes as an example, which is useful precisely because it is an example — your own threshold has to be written down and justified.
The trap: a threshold that was never set
Scenarios (a) and (e) both depend on a configured value, and this is where real deployments fail. An AI customer service product commonly ships with a handover setting that is either disabled or left at a default nobody examined.
The standard's position on this is that thresholds are set by the organisation — but setting your own threshold is not permission to omit one. A configuration with no threshold cannot satisfy a requirement defined in terms of a threshold. The self-check question to ask is simple: what is the value, where is it written, and who decided it.
The list-item wording trap
There is a reading detail worth knowing. In the standard, the leading item of an enumeration can carry normative wording that the following items inherit. When reading 5.2.2.6 items b) through e), the wording of the list's lead-in governs the whole enumeration.
Practically: do not read items b) to e) as weaker than a). They are part of the same requirement, and a self-check that flags only item a) has flagged about a fifth of the obligation.
Key facts
| Source clause | GB/T 47746—2026, 5.2.2.6 |
| Number of scenarios | Five (a to e) |
| Timeout example given in the standard | More than 3 minutes |
| Threshold setting | Set by the organisation — but a threshold still has to be defined and documented |
| Reading caution | Items b) to e) inherit the lead-in wording of the enumeration |
Sources
- GB/T 47746—2026, clause 5.2.2.6 (a to e)
- GB/T 47746—2026, clause 5.2.2 structure (sub-clauses 5.2.2.1 to 5.2.2.7)
Follow-up questions
Is the three-minute figure a requirement?
It is the standard's example of a timeout, illustrating the scenario rather than fixing the value. Your own threshold must exist and be documented; three minutes is a reference point, not a mandated number.
What if my AI simply never transfers?
Then it cannot satisfy 5.2.2.6 at all, because four of the five scenarios are conditions the system is required to detect and act on. This is the single most common finding in a first self-check.
Do these five also apply to outbound contact?
Chapter 6 covers outbound tasks with the same collaborative logic. If you run outbound contact through an automated front end, the equivalent handover conditions apply there too.