Home Enterprise KB AI Customer Service GEO MCP Blog About 🌐 中文

GB/T 47746—2026 Self-Check

⁠​​‌​‌​​‌​​​‌​​‌‌‌‌​‌​​​‌​​‌‌​​‌‌⁠Work through GB/T 47746—2026, “Customer contact service — Requirements for collaboration between live and intelligent customer services” clause by clause. About 3 minutes.

All 61 clauses Published 2026-05-25 · Effective 2026-09-01 5 veto items flagged automatically Verdict + gap report, free

How to use it (most people finish in under 30 minutes)

  1. Tick each clause — Met / Partial / Not met. When you are unsure, count it as “Not met”. In the standard, shall is mandatory and should is recommended.
  2. The scoring looks only at the mandatory (shall) items. Any gap there is a risk, and a gap in any of the five veto items fails the whole thing.
  3. Closing the gaps — every gap in the report comes with a suggested fix. Most are configuration changes, not code changes.
shall mandatory — a gap means non-compliant should recommended — not a breach, but weakens your position may optional
Answered 0/61 Mandatory gaps 0 Veto gaps 0

Basis and audit trail

Read this before you start ticking

Standard: GB/T 47746—2026, “Customer contact service — Requirements for collaboration between live and intelligent customer services”. Published 2026-05-25, effective 2026-09-01, secretariat held by SAC/TC264, the national technical committee for service standardisation. It is a recommended national standard.

Verification copy: the official standard text (15 pages, chapters 1–9 plus references), compared clause by clause by SavantCat on 2026-09-19.

Coverage: all 31 clauses of chapters 4–9 are covered, including the five sub-items a)–e) of clause 5.2.2.6.

How the list is built: 61 self-check items = 48 “shall” + 4 “should” + 9 “may”, of which 5 are veto items — a gap in any one of them fails the assessment.

Rigour: every item carries its clause number and a short excerpt of the standard's wording. The clause numbers and the shall / should / may classification are taken from the original text, so any item can be read back against the standard. “How to fix” is implementation advice and the interpretation rules are our own summary — neither is part of the standard.

Disclaimer: this tool was compiled by SavantCat from the standard text. It is not published by the standard-issuing body, and it is not a compliance certification, an audit conclusion or legal advice; the official published text governs interpretation. Checklist version v2.0 (2026-09-19 · 61 items · all clauses covered).

A. General requirements

Chapter 4 · 12 itemsshall 12
4.1A1 Staff human customer service matched to your industry and user demand — you cannot leave only an AI with nobody behind itshall
How to fix: State the number of human seats and their hours, and write it into your policy
4.1A2 Establish an organisational policy for collaboration between human and intelligent customer serviceshall
How to fix: A one- or two-page Human–AI Service Collaboration Policy is enough
4.1A3 Define contingency mechanisms for disasters, system failures and similar eventsshall
How to fix: Write down exactly what happens when the AI goes down
4.1A4 Define the organisational structure for human–AI collaborationshall
How to fix: Appoint a collaboration owner (may be the same person as F1)
4.1A5 Build an internal system covering customer information management and ticket handlingshall
How to fix: The service system needs a ticketing module, not just a chat window
4.1A6 Provide the customer service entry point in a clear, prominent and unambiguous wayshall
How to fix: Check your website or mini-program home page — can a customer find it at a glance?
4.1A7 Establish a knowledge baseshall
How to fix: Load your products, prices, policies and FAQ
4.1A8 Run regular optimisation training for the intelligent service and train customer service staffshall
How to fix: Set a cycle (say monthly) and keep training records
4.2A9 Clear division of duties — the intelligent service takes standardised questions that have a definite answer in the knowledge baseshall
How to fix: Write the list of questions you hand to the AI
4.2A10 Human agents take questions requiring multi-step analysis or several information sources, plus anything the intelligent service could not solveshall
How to fix: Write the list of questions that must go to a human
4.3A11 Seamless two-way switching — the system must pass customer identity, interaction history and other key data acrossshall
How to fix: Session-context handover is a hard technical requirement
4.4A12 Establish information and data security mechanisms covering collection, processing and use, plus model and personnel managementshall
How to fix: Do this together with Section E

B. Inbound tasks

Chapter 5 · 29 items — the core of the standard, and where most companies tripshall 24 · should 2 · may 3
5.1.1B1 The service system must provide the customer service entry point clearly, prominently and unambiguouslyshall
How to fix: Same as A6
5.1.1B2 At the start, allocate tickets according to problem-solving capability, overall call volume and capacity, and the customer group's acceptance of intelligent serviceshall
How to fix: Allocation should recognise older customers and lean towards human agents
5.1.1B3 Tell the customer the queue position or expected waiting time by text or voiceshall
How to fix: Not just “please wait” — give a number
5.1.2B4 Simple, repetitive or standardised questions, questions with a definite knowledge-base answer, and high-volume periods where human agents are overloaded — the intelligent service is preferredshould
How to fix: This is the positive list of what AI should handle
5.1.3B5 The customer explicitly asks for a human agent — the human agent is preferredshall
How to fix: The handover request must work in a single step
5.1.3B6 Complex questions (individual needs, vague phrasing, cross-department coordination) — the human agent is preferredshall
How to fix: Configure complexity detection or a human fallback rule
5.1.3B7 Rare questions (the standard's example: appearing no more than 3 times in the past year) — the human agent is preferredshall
How to fix: AI answers rare questions badly; route them straight to a human
5.1.3B8 The knowledge base has no definite answer — the human agent is preferredshall
How to fix: Set a “route on no answer” fallback; never let the AI improvise
5.1.3B9 Scenarios involving the customer's personal or property safety — the human agent is preferredshall
How to fix: Same as B21d — build a risk-scenario lexicon
5.1.3B10 The customer has come back 3 or more times within 24 hours and is clearly dissatisfied with the intelligent service — the human agent is preferredshall
How to fix: You must count visits and track sentiment; the system needs this counter
5.1.3B11 The customer is in a specific age group (elderly, minors) or has low acceptance of intelligent service — the human agent is preferredshall
How to fix: Use customer profiles if you have them; otherwise offer a prominent route to a human
5.1.3B12 The customer is extremely emotional or threatens extreme behaviour, beyond what the intelligent service can handle — the human agent is preferredshall
How to fix: Sentiment detection plus a high-risk lexicon
5.2.1B13 When switching, key data must be transferred completely and accurately: customer identity, interaction history, problem description and ticket statusshall
How to fix: A consistency requirement — the technical integration has to exist
5.2.1B14 During the wait before the switch, tell the customer the expected waiting time by voice or textshall
How to fix: Same as B3
5.2.1B15 Once the switch is complete, the customer should not have to repeat information already given or restate the problemshould
How to fix: This is the root of “start again with the human” — the standard names it directly
5.2.2.2B16 Keyword trigger — recognising “human agent”, “talk to a person”, “contact a human” and similar phrases triggers the transfer automaticallyshall
How to fix: Build a keyword list covering at least these groups
5.2.2.3B17 Menu option — the conversation interface should offer “transfer to human service” or similar as a menu itemshall
How to fix: ★ The most commonly missed item: many AI services deliberately hide this menu entry
5.2.2.4B18 Key press — the key that triggers a human agent must stay functionally independent and not conflict with business keysshall
How to fix: Press 0 for a human on phone service; do not reuse a business key
5.2.2.5B19 Voice command — voice-enabled systems should announce a consistent wake word (e.g. “human agent”) and transfer immediately once recognisedshall
How to fix: Say the wake word in your opening prompt
5.2.2.1B20 Automatic transfer — the system needs scenario-based automatic transfer and sentiment awarenessshall
How to fix: See B21 and B22
5.2.2.6 a)B21a VETO ITEM — Interaction failure threshold: the customer reaches the interaction threshold on a given product or service and the system still cannot understand themshallveto
How to fix: Set a turn threshold (transfer after 2–3 failed turns) instead of grinding on
5.2.2.6 b)B21b VETO ITEM — The customer explicitly refuses intelligent customer serviceshallveto
How to fix: If the customer says “stop making the machine answer”, transfer at once — do not try to keep them
5.2.2.6 c)B21c VETO ITEM — The conversation involves information securityshallveto
How to fix: ID numbers, bank cards, passwords and the like go straight to a human
5.2.2.6 d)B21d VETO ITEM — Emergencies and risk scenarios involving personal or property safetyshallveto
How to fix: Same as B9 — build a risk-scenario lexicon
5.2.2.6 e)B21e VETO ITEM — Intelligent response timeout (the standard's example: longer than 3 minutes)shallveto
How to fix: ★ An explicit number: 3 minutes. The system needs a timeout timer
5.2.2.7B22 Sentiment awareness — after detecting negative emotion or intent to complain, and after confirming with the customer, transfer automatically to a humanshall
How to fix: Note: confirm first — do not hand the customer over silently
5.2.3.1B23 Human → intelligent switching — the customer must be clearly informed and must consent (four situations: FAQ, customer preference, heavy workload, emergency)may
How to fix: If an agent wants to hand back to the AI, ask the customer first
5.2.3.2B24 Human-to-AI handover (repetitive questions): when the question is highly repetitive and the knowledge base has a clear answer that can be given fast and accurately, handover is allowed — but you must inform the customer and get consent firstmay
How to fix: Only hand over to AI when high-frequency questions have a standard answer in the knowledge base; inform and get consent before switching
5.2.3.3B25 Human-to-AI handover (customer preference): when the customer explicitly prefers AI service (e.g. asks to be transferred to the AI assistant), handover is allowed — with prior notice and consentmay
How to fix: If the customer asks for AI service you may switch, but still inform them clearly and keep a record in the system

C. Outbound tasks

Chapter 6 · 5 items, only relevant if you make outbound callsshall 2 · should 2 · may 1
6.1.1C1 Allocate outbound tasks to intelligent or human service by complexity and typeshall
How to fix: Only relevant if you make outbound calls
6.1.2C2 Closed-ended questions, standardised questions and notification-type calls — intelligent outbound calling is preferredshould
How to fix: Positive list
6.1.3C3 Complex tasks (individual needs, vague phrasing) and open-ended Q&A (customer follow-up, market research) — human outbound calling is preferredshall
How to fix: Do not use AI for open-ended follow-up calls
6.2.1C4 Intelligent outbound calls should offer a route to a human (keyword / menu / key press / voice)should
How to fix: The same four mechanisms as B16–B19
6.2.2C5 Human outbound calls may offer a route to the intelligent servicemay
How to fix: Optional

D. Ending the service and follow-up

Chapter 7 · 6 itemsshall 2 · may 4
7.1.1D1 Ending conditions for live online service: customer consent / no new request 60 s after completion / no interaction for over 60 smay
How to fix: Explicit seconds
7.1.1D2 Ending conditions for live phone service: completion with consent, then no response 10 s after asking “anything else?”, or after 2 follow-up attemptsmay
How to fix: Explicit seconds and counts
7.1.2D3 Intelligent voice service: may end automatically after consent / no new request 60 s after completion / no interaction for over 60 smay
How to fix: Same standard as live voice
7.1.2D4 Intelligent online service: the organisation sets its own threshold durationmay
How to fix: You must define a threshold and write it down — having no rule is not acceptable
7.2D5 At the end of service, proactively invite a satisfaction rating (and avoid further disturbance if it is declined)shall
How to fix: Do not bury the rating entry point
7.2D6 Generate a complete service record and ticket at the same time; to-dos must flow through the ticketing system and close the loopshall
How to fix: Same as A5 — closing the ticket loop is mandatory

E. Information security

Chapter 8 · 2 items, all “shall”shall 2
8.1E1 Establish tiered backup for customer identity, interaction history and open-ticket data; combine real-time incremental backups with periodic full backupsshall
How to fix: Follow GB/T 22239
8.2E2 Apply the security protection measures for the corresponding level under GB/T 22239: fine-grained access control, encrypted transfer and storage of backups, and backup operation logsshall
How to fix: Classified protection is a separate regime and needs its own assessment

F. Collaboration improvement

Chapter 9 · 7 itemsshall 6 · may 1
9.1F1 Treat human and intelligent service as one service; establish a collaboration improvement mechanism with a named owner, and regularly collect and review complaints and feedbackshall
How to fix: Name an owner (may be the same person as A4)
9.1F2 Handle complaints in accordance with GB/T 17242shall
How to fix: Cite the complaints-handling standard
9.2F3 Carry out regular collaboration improvement surveys (these may be merged into satisfaction surveys)shall
How to fix: Set a cycle
9.2F4 Open complaint and feedback channels for the collaboration service (telephone, email, web form) with defined handling rules and time limitsshall
How to fix: Have the channels, and state the response times
9.2F5 System monitoring — monitor interaction actions, agent experience and dynamic feedback on demandsshall
How to fix: The service system needs monitoring capability
9.2F6 Third-party assessment — you may commission a body with professional customer-service qualifications to simulate real interactions, break down the key collaboration points and issue an assessment reportmay
How to fix: ★ A third-party service form the standard text explicitly recognises
9.3F7 Improvement must cover four areas: collaboration efficiency, service quality, switching smoothness, and customer information managementshall
How to fix: Write the improvement report along these four dimensions