Skip to main content

China's New AI Customer Service Standard: How Bytedesk Delivers a Visible Human Handoff and Accountable AI Replies

· 8 min read
Jack Ning
Jack Ning
Maintainer of Bytedesk

On September 1, 2026, China's first national standard for human–AI customer service collaboration came into force: GB/T 47746—2026, "Customer Contact Services — Requirements for the Collaboration of Human and Intelligent Customer Service." Media coverage distilled the standard into four red lines:

  1. A clear, prominent human entry point: customer service menus must offer a visible "talk to a human" option — no burying it behind layers of navigation;
  2. Scientific division of labor: AI handles simple, standardized questions; complex disputes, high-risk scenarios, and explicit requests for a human are routed to agents promptly;
  3. Smooth human–AI handoff: when switching, the customer's identity, interaction history, and ticket status must carry over completely — customers should never have to repeat themselves;
  4. Enterprises are accountable for AI replies: companies cannot dodge responsibility with disclaimers like "AI-generated content is for reference only."

For enterprises evaluating customer service systems, these four red lines are effectively a procurement checklist. This post walks through the standard clause by clause and introduces the capabilities Bytedesk has already shipped — not compliance as a slogan, but as features and mechanisms you can find and verify in the product.

The Checklist at a Glance: Standard vs. Bytedesk​

Standard RequirementBytedesk Capability
Clear, prominent human entry (4.1 / 5.1)A persistent "human agent" button in the visitor chat window; human-handoff actions on every bot reply and FAQ bubble; trigger keywords, entry toggle, and button label all configurable in the admin console
Scientific division of labor (4.2)Three routable thread modes: bot-only / human-only / bot-first with human transfer, freely combined per business complexity
Smooth handoff, no repeated questions (4.3 / 5.2)Bot → queue → agent transitions happen within the same conversation thread, with full message history and customer context preserved
Accountable AI replies (4.4)Every bot reply can be rated up/down, evaluated, and quality-inspected; thumbs-down and negative emotion automatically trigger human transfer
Service closed loop (7.2)Post-session satisfaction surveys + automatic ticket creation on session close + ticket workflow
Continuous improvement (Chapter 9)Service statistics (resolution rate, response time, transfer count) + quality inspection + knowledge base analytics

Red Line 1: A Clear, Prominent Human Entry Point​

Bytedesk provides multiple layers of human entry in the visitor client, so customers can always reach a person:

  • Persistent button: the visitor chat window always shows a "human agent" button — no digging through nested menus;
  • Bubble-level actions: every bot reply and every FAQ answer carries a "transfer to human" action, so a customer dissatisfied with an AI answer can escalate with one click;
  • Keyword triggering: typing keywords such as "human agent" (in Chinese: 人工 / 转人工 / 人工客服 / 真人 / 转接人工) automatically transfers the conversation — the default keyword list works out of the box and can be extended or trimmed in the admin console;
  • Admin-configurable: administrators can decide whether visitors may manually transfer to a human at all, and customize the button label (e.g., "Talk to a real person"), keeping the entry point both prominent and on-brand.

Under the hood, the keyword listener applies strict trigger guards: it only fires for workgroup conversations that are currently in bot-serving state with no agent assigned yet, and it ensures the transfer status is persisted before the bot's reply is stored — preventing the experience failure of "the customer already asked for a human, but the bot keeps talking."

Red Line 2: Scientific Division of Labor​

The standard requires AI to handle standardized questions first, while humans prioritize complex ones. Bytedesk turns this division into three configurable routing modes:

Routing ModeBest ForHuman Transfer
Bot-onlyStandardized inquiries, 24/7 first responseNot supported (pure self-service scenarios)
Human-onlyHigh-value customers, complex business lines—
Bot-first + transferable (workgroup)Most customer service scenariosSupported: keyword, thumbs-down, emotion, and other multi-channel triggers

In workgroup mode, the AI answers standardized questions that have clear answers in the knowledge base; as soon as a complex issue or an explicit human request appears, the conversation is routed to a human agent. Bytedesk also tracks the "human transfer count" per FAQ entry — frequently transferred questions are a signal that the knowledge base needs enrichment, so the human–AI boundary keeps evolving with operational data.

Red Line 3: Smooth Handoff, No Repeated Questions​

The most painful handoff experience is "a new person takes over, and everything starts from scratch." Bytedesk's conversation model rules this out by design:

  • Same-thread transitions: bot serving, queueing, and agent service all happen in the same conversation thread, with state flowing from "bot serving → queuing → agent serving" — no new conversation, no lost context;
  • Complete context sync: when an agent joins, they immediately see the customer profile and the full AI conversation history, so the customer never has to repeat the problem;
  • Transparent queueing: customers are informed of their queue status while waiting — no "waiting in the dark";
  • Complete fallback paths: outside service hours, the conversation automatically becomes an offline message; when no agent is online, the system first tries a backup skill group before falling back to a message; when all agents are busy, the customer automatically enters the queue. Whatever happens, the customer's request always lands somewhere.

Red Line 4: Enterprises Are Accountable for AI Replies​

The standard is explicit: companies cannot use "AI-generated content is for reference only" to shirk responsibility. The precondition for being accountable is that every AI utterance is manageable, traceable, and correctable. Bytedesk provides the full accountability chain:

  • Per-message feedback: customers can rate every bot reply up or down;
  • Thumbs-down as safety net: after a thumbs-down, the system by default automatically tries to transfer to a human agent, and can simultaneously trigger a satisfaction survey and flag the reply for quality inspection — negative feedback never silently disappears;
  • Emotion awareness: customer messages are analyzed for emotion (satisfied / neutral / dissatisfied / angry / anxious, etc.); when negative emotion is detected, the system can automatically transfer to a human, escalate, or proactively offer help according to policy, with a configurable confidence threshold to prevent misjudgment;
  • End-to-end quality inspection: conversation records are fully retained, and inspection can pinpoint the exact AI reply — what the AI said, how the customer reacted, and how it was handled are all on the record.

Service Closed Loop: Satisfaction and Tickets​

Chapter 7 of the standard requires a closed service loop: proactively invite satisfaction evaluation, generate complete service records, and flow to-dos through tickets. Bytedesk's capabilities:

  • Satisfaction surveys: customers are automatically invited to rate the session when it ends, with two-way evaluation supported (agents rate customers, customers rate agents); the VOC module additionally supports systematic CSAT and NPS programs;
  • Automatic ticket creation: when a session closes, tickets can be created automatically based on close type (auto-timeout / agent-closed / visitor-closed), time windows, and other conditions; unresolved requests flow into the ticket pipeline, driven by the Flowable workflow engine — every complaint has a traceable ticket.

Continuous Improvement — and an Honest Note on Boundaries​

Chapter 9 requires continuous monitoring and improvement of collaboration efficiency and service quality. Bytedesk provides service statistics (resolution rate, response time, transfer-to-human count, daily served customers, dissatisfied session count, etc.), a quality-inspection console, and knowledge base analytics (FAQ transfer counts and thumbs-down data feed content optimization), enabling data-driven collaboration improvement.

We also state our current capability boundaries honestly:

  • Finer-grained auto-transfer policies (such as thresholds on interaction failure counts, or automatic transfer after a bot response timeout) are not yet available as standalone configuration items; they can be enhanced in future releases. Today, multi-channel auto-transfer via keywords, thumbs-down, and emotion detection covers the vast majority of human-handoff scenarios;
  • Data backup and classified protection (MLPS) are deployment and operations concerns: Bytedesk provides a security foundation of private deployment, fine-grained access control (RBAC), and configuration encryption, while tiered backup and MLPS assessment per GB/T 22239 are implemented by the deploying party on their own infrastructure.

Compliance is not a one-page declaration but an ongoing operational mechanism. Being explicit about boundaries is itself part of being responsible to customers.

Why Bytedesk​

Measured against GB/T 47746—2026, Bytedesk's advantages come down to three points:

  1. Open source and auditable: every mechanism — human entry, division of labor, conversation flow, accountability chain — has a public implementation you can verify during evaluation, rather than a vendor's verbal promise;
  2. Private deployment, data sovereignty: customer data never leaves your enterprise; security, backup, and compliance policies are under your control;
  3. Compliance as built-in capability: entry points, division of labor, handoff, accountability, and closed loops required by the standard are out-of-the-box product features in Bytedesk — not paid custom projects.

The new standard turns "customer-centricity" from a slogan into verifiable engineering requirements — which is exactly the direction Bytedesk has always pursued.