Direct answer
Emotion sensing is one of the intelligent-to-live transfer routes listed in clause 5.2.2 of GB/T 47746—2026, and the standard does not require any particular technology. Practical implementations fall into three tiers: keyword and sentence-pattern rules (the cheapest), conversation sentiment scoring (a model or a sentiment lexicon), and human side-channel monitoring (listening in and intervening proactively). For a small or medium business, start at the rules tier and use the number of times a customer shows dissatisfaction as the trigger quantity — the priority is that the route exists and can be verified.
What the standard says
GB/T 47746—2026 lists emotion sensing in clause 5.2.2 as one of the seven routes for intelligent-to-live transfer. What it requires is that the capability exists; it does not require an emotion recognition model. That distinction matters, because it leaves an SME room to choose the tier that fits its own circumstances.
Three tiers, from cheapest to most expensive
| Tier | Approach | Suited to | Limitation |
|---|---|---|---|
| Rules | Keywords and sentence patterns, repetition count, negative word lists | The great majority of small and medium businesses | Cannot catch sarcasm or indirect dissatisfaction |
| Model | Conversation sentiment scoring, threshold-triggered | High conversation volumes, complaint-sensitive industries | Needs labelling and tuning; misclassifications need human correction |
| Human | Agent or supervisor monitoring, with proactive intervention | High-value customers, difficult conversations | Depends on staffing; cannot cover everything all day |
How to make the rules tier respectable
Do not configure the single word "complaint". Cover at least:
- negative emotion words (annoyed, angry, bad review, exposure, report, and so on);
- repeated rounds without resolution (the same request unresolved for 2 or 3 consecutive rounds);
- escalation intent words (speak to your manager, want an explanation, going to complain);
- strong refusal expressions (no bot, stop transferring me).
The core advantage of the rules tier is that it is explainable — if a customer asks why they were transferred to a person, you can state the trigger condition.
What to watch for in the model tier
Sentiment scoring must come with a human review loop, otherwise a wrong transfer alienates both the customer and the agent. A sensible order is to run the rules tier for a month, collect real trigger samples, and then use those samples to calibrate the model threshold — far more reliable than going straight to a model.
How to tell whether this route works
Not by model accuracy. Look at three things:
- Real conversations were transferred because of an emotion judgement (the route is alive);
- the outcome of those transferred conversations can be traced (it is not just passing the problem on);
- the rules or thresholds have a version record (they can be rechecked).
One line of advice
Get the rules tier solid first, and count emotion triggers into the conversation log. Once a few hundred real samples have accumulated, decide whether an upgrade to the model tier is warranted — in the opposite order, you spend the money and still cannot explain the effect.
Key facts
| Standard requirement | Emotion sensing is one of the transfer routes listed in 5.2.2; no technology is prescribed |
| Minimum rule coverage | Negative words, rounds repeated without resolution, escalation intent words, strong refusal expressions |
| Starting recommendation | Run the rules tier for 1 month to collect samples, then calibrate the model threshold |
| Three acceptance checks | Real triggers exist / outcomes are traceable / thresholds have a version record |
Sources
- GB/T 47746—2026, clause 5.2.2 (intelligent to live handover, including emotion sensing)
- GB/T 47746—2026, clause 5.2.2.6 (automatic transfer in specific scenarios)
- Transfer-trigger items among the 59 checklist items
Follow-up questions
Will we fail the standard if we do not do emotion recognition?
Emotion sensing is one of the routes, and the test is whether that route is usable and verifiable. Implementing it with rules is equally valid; having no such route at all is what actually fails.
How much does an emotion model cost?
It varies enormously, from cloud APIs billed per call to training your own. For an SME the sensible start is the rules tier, then use real trigger samples to judge whether an upgrade is worth the money.
Does transferring too often waste agent time?
It does, so the threshold has to be tuned. Early on, err slightly on the sensitive side to avoid angering customers, then tighten it in line with agent scheduling and real cost.
Must a customer who insults the bot be transferred?
Abuse generally falls under negative emotion and strong refusal expressions and should trigger a transfer; if it involves an emergency affecting personal or property safety, it becomes a mandatory transfer under 5.2.2.6.