9 月 1 日，GB/T 47746—2026《顾客联络服务 人工与智能客户服务协同要求》正式实施。这是国内第一个专门约束人机客服协同的国家标准，几条硬要求值得直接抄进需求文档：
- 必须设置清晰、醒目的人工服务入口，严禁隐藏人工通道；
- 用户明确要求转人工时，必须及时转接；
- 切换后要同步客户信息，不得重复询问；
- 涉及价格、折扣、退款、赔偿、合同变更等事项，AI 可以解释规则、收集材料、生成工单，最终确认应由人工或规范化流程完成。
现实是落地两个月，媒体实测仍在翻车：某购票平台记者三次说「转人工」才接通，有人等 40 分钟未接通导致票作废，甚至有机器人在识别到用户坚持后直接挂断。中消协上半年数据里，售后服务投诉占 26.79%，「AI 客服不实承诺、人工接入困难」是新增热点。
这些问题跟模型能力关系不大，几乎全是会话状态管理和产品策略的问题。下面给一份能直接跑的骨架。
技术要点
四个约束，决定了实现方式：
1. 转人工必须是规则，不是模型判断。 意图识别交给人，兜底交给规则引擎。标准里写的「明确要求」「超时」「高风险场景」，本质都是可枚举的触发条件。 2. 会话要有状态计数。 兜底话术连续命中几次、总轮次到了几轮，都得记。大多数线上事故就是没有这两个计数器。 3. 风险词优先于业务词。 「投诉」「监管」「12315」「人身安全」这类词必须走独立通道，不能被「自助流程」吃掉。 4. 转接要带上下文。 转过去之后人工看不到历史记录，等于用户要把委屈再讲一遍，体验二次崩塌。
可运行示例
纯标准库，复制即可跑。核心是一个 route() 函数，返回「继续聊」还是「转人工」。
# 三段代码块按顺序拼接后可直接运行
from dataclasses import dataclass
from typing import Callable
# --- 明确要求转人工：命中即转（国标硬要求）---
ESCALATE_PATTERNS = ("转人工", "人工客服", "找真人", "真人客服", "我要人工")
# --- 风险词：命中即升级，优先级最高 ---
RISK_PATTERNS = ("投诉", "监管", "12315", "起诉", "曝光", "受伤", "人身", "诈骗")
MAX_TURNS = 4
MAX_REPEAT = 2
# --- 会话状态：三个计数器全在这儿 ---
@dataclass
class Session:
sid: str
turns: int = 0
fail_hits: int = 0 # 兜底话术连续命中次数
last_reply: str = ""
history: list = None
def __post_init__(self):
self.history = []
def route(session: Session, text: str, reply_fn: Callable[[str], str]):
"""返回 (action, 回复文本)。"""
text = (text or "").strip()
session.turns += 1
session.history.append(("user", text))
# 1. 风险词最高优先级，不做任何自助挽留
if any(k in text for k in RISK_PATTERNS):
return "escalate", "已为您转接人工，请稍候。"
# 2. 明确要求转人工：一次就转，不挽留
if any(k in text for k in ESCALATE_PATTERNS):
return "escalate", "正在为您转接人工客服。"
# 3. 正常业务链路：模型只负责生成回复，永远不决定要不要转
reply = reply_fn(text)
if reply == session.last_reply:
session.fail_hits += 1
else:
session.fail_hits = 0
session.last_reply = reply
session.history.append(("bot", reply))
# 4. 兜底：重复话术 / 轮次超限，两个计数器任一命中
if session.fail_hits >= MAX_REPEAT:
return "escalate", "这个问题我处理不了，已为您转接人工。"
if session.turns >= MAX_TURNS:
return "escalate", "已经为您尝试多次，转接人工继续处理。"
return "reply", reply
if __name__ == "__main__":
FALLBACK = "非常理解您的心情，请您耐心等待。"
def fake_llm(text: str) -> str:
# 模拟一个只会答非所问的机器人
if "退" in text:
return FALLBACK
return "您可以在订单详情页自助操作。"
def run(cases, label):
s = Session(sid=label)
print(f"--- {label} ---")
for i, q in enumerate(cases, 1):
action, msg = route(s, q, fake_llm)
print(f"[{i}] {q} -> {action}: {msg}")
if action == "escalate":
print(f" 升级时携带 {len(s.history)} 条上下文")
break
# 三条路径各跑一遍
run(["我要退两张电影票", "退不了", "还是退不了", "算了"], "兜底话术重复触发")
run(["这个优惠券怎么用不了", "我要投诉"], "风险词触发")
run(["这个优惠券怎么用不了", "转人工"], "明确要求触发")
三条路径各跑一遍：兜底重复那条在第 3 句升级，风险词和明确要求都在第 2 句升级。也就是说，不需要模型多聪明，只要三个计数器有一个响了，人就能被捞出来。
如果要接大模型生成兜底话术，建议把「生成」和「决策」拆开：模型只负责 reply_fn，升级与否永远由 route() 决定。我平时调这类多模型兜底，统一入口切换的，省掉给每家单独对接 SDK 的功夫，也不用为了几个备用模型各开一份订阅。关键路径上别用单一供应商，这点比选哪个模型重要。
踩坑提示
坑一：把「人工」「0」当触发词。 很多系统对单字识别不准，「人工」两个字经常被当成业务词吃掉。要么匹配完整短语，要么维护一张同义词表，别指望模糊匹配。
坑二：情绪识别做反了。 有团队上了情绪打分，用户越急越触发「安抚话术」分支，结果把人困得更久。正确做法是情绪升级直接降权自助流程、提权转人工，不是加一段安慰。
坑三：把「转人工率」当负向 KPI。 这是根子上的问题。一旦转人工率高要被扣绩效，产品经理就会去藏入口，工程侧再漂亮也白搭。指标至少要成对看：转人工率 + 转人工后一次解决率 + 平均接通时长。
坑四：转接丢上下文。 国标明确要求切换后同步客户信息、不重复询问。实现上给工单附一段结构化摘要就够：用户诉求、已尝试动作、订单号、情绪标记。别把整段对话原样丢过去，人工读不完。
坑五：AI 的承诺没留痕。 中消协的态度很明确，经营者自己部署的 AI 客服，输出内容属于服务的一部分，「AI 说的不算」这种甩锅站不住。会话记录必须可检索、可导出，退款赔偿类承诺要落工单。
总结
转人工难的解法，八分在工程，两分在决心。工程部分是这几个计数器和一个优先级队列，一个下午能写完；决心部分是愿不愿意在「省人工」和「别把人逼疯」之间选后者。标准已经给了标尺，接下来谁先改，谁少挨骂。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
1 September, GB/T 47746-2026 `Customs Liaison Services `Physical and Intelligent Client Services Collaboration Requirements ' was officially implemented. This is the first national standard in the country to regulate the coordination of passenger flights of human aircraft, and several hard requirements deserve to be copied directly into the demand file:
- Clear and visible manual service entrances must be set up and the concealment of artificial access points must be strictly prohibited;
- When a user explicitly requests a transfer, it must be made in a timely manner;
- Synchronization of customer information after switching and not repeat queries;
- With regard to price, discount, refund, compensation, contract change, etc., AI may explain the rules, collect materials, generate work orders, and ultimately confirm that the process should be manual or standardized.
The reality is that for two months, the media is still turning over: the news reporter on a ticket buying platform said that he was connected three times, that someone waited 40 minutes to cancel the ticket, and even that robots hung up when they recognized the user’s insistence. In the first half of the year, after-sale service complaints accounted for 26.79 per cent, while "AI's false promises, manual access difficulties" were added.
These problems have little to do with modelling capabilities and are almost exclusively a matter of conversational management and product strategies. Here's a skeleton that can run straight.
Technical elements
Four constraints determine how to achieve them:
1. The transfer of labour must be the rule, not a model judgement. The intent is to identify the person, and the cover is to the rule engine. timeout, a high-risk scene, is essentially an itemable trigger. 2. Sessions shall be counted in a state. It is important to remember the number of hits and the total number of rounds. Most of the online accidents were the absence of these two counters. 3. The risk word prevails over the business word. The term "complaint" "supervisory" "12315" "personal security" must be used independently and cannot be eaten by the "self-help process". 4. The transfer is context. Turning around without seeing historical records means that users want to repeat their grievances and experience secondary collapses.
Runable Example
Pure library, copy and run. The core is a route() function that returns the words "continue chatting" or "transforming".
# Three segments of the code blocks can run directly after they are spelled in sequence
from dataclasses import dataclass
from typing import Callable
# -- Unequivocal request for man-to-man: hit-to-man (hard national requirement) ---
ESCALATE PATTERNS = ("Turn", "Advanced", "Find a real person", "Advanced a human person")
# -- risk word: promotion at impact, highest priority ---
RISK PATERNS = ("complaint", "supervisory", "12315", "prosecution", "discovery", "injured", "person", "fraud")
MAX_TURNS = 4
MAX_REPEAT = 2
# -- Session status: all three counters here ---
@dataclass
class Session:
sid: str
turns: int = 0
Fail hits: int = 0 # consecutive hits of a spin-down
last_reply: str = ""
history: list = None
def __post_init__(self):
self.history = []
def route(session: Session, text: str, reply_fn: Callable[[str], str]):
""Returns."
text = (text or "").strip()
session.turns += 1
session.history.append(("user", text))
# 1. Highest priority for risk terms, without any self-help retention
if any(k in text for k in RISK_PATTERNS):
Turn "escalate", "for your transfer, please wait."
# 2. Specific request for transfer: one turn, no retention
if any(k in text for k in ESCALATE_PATTERNS):
Turning "escalate", "is transferring you to the artificial passenger service."
# 3. Normal business link: the model is only responsible for generating responses and never decides whether to turn or not
reply = reply_fn(text)
if reply == session.last_reply:
session.fail_hits += 1
else:
session.fail_hits = 0
session.last_reply = reply
session.history.append(("bot", reply))
#4. Bottom: repeat / rotation exceeding limit, two counters having any life Medium
if session.fail_hits >= MAX_REPEAT:
Turn "escalate", "I can't handle this problem, I've switched to you."
if session.turns >= MAX_TURNS:
Turn "escalate", "have tried many times for you to switch to manual processing."
return "reply", reply
if __name__ == "__main__":
FALLBACK = "I understand your mood very well, please wait patiently."
def fake_llm(text: str) -> str:
# Simulate a machine that only answers questions People
If "go back" in text:
return FALLBACK
"You can do it on the order details page."
def run(cases, label):
s = Session(sid=label)
print(f"--- {label} ---")
for i, q in enumerate(cases, 1):
action, msg = route(s, q, fake_llm)
print(f"[{i}] {q} -> {action}: {msg}")
if action == "escalate":
print(f"Hystory)}Legend context")
break
# All three paths run together
Run.
Run.
Run.
The three tracks run through each other: repeat in the background that the article is upgraded in the third sentence, and the risk words and explicit requirements are upgraded in the second sentence. That is to say, there is no need for models to be smart, and if one of the three counters goes off, people can be pulled out.
If you want to use LLM as a spin-off, it is recommended to remove the words "generation" and "decision-making": the model is only responsible for reply fn, and promotion is always determined by route(). I usually do this multi-model bottom-up, single-entry switch, save each family the ability to connect to SDK alone, and do not have to sign up for a few alternative models. It is more important than choosing which model not to use a single supplier on a critical path.
Catapult hint
Tunnel One: "Manual" "0" as a trigger. Many systems fail to recognize words, and the word "manual" is often eaten as a business word. Either the full phrase is matched or a synonym is maintained, with no expectation of vague matching.
Pipe 2: Emotional recognition is the opposite. The more emotional the team scores, the more the user triggers the soothing branch, and the longer it takes. The correct approach is to upgrade the emotional power of the self-help process, transfer power, and not to add comfort.
Pipe Three: Treat the conversion rate as negative to KPI. That's the root problem. Once the turnover rate is high, the product manager will go to hide the entrance and the project side will be in vain. At a minimum, the indicator needs to be cross-referenced: the conversion rate + the subsequent resolution rate + the average connection time.
Pipe Four: context turned over. The national flag explicitly requires the synchronisation of customer information after switching, without repeat queries. A structured section of summary is sufficient to achieve a job offer: user claims, attempted actions, order numbers, emotional markers. Don't leave the whole conversation behind and read it manually.
The commitment of AI was not marked. It's clear that the operator's own AI service, the output of which is part of the service, does not count. Session records must be searchable, exportable and refundable commitments must be placed on the payroll.
Summary
The solution to the problem is eight-point construction and two-point determination. The project is based on these counters and a priority line, which can be completed in one afternoon; and the determination is to choose the latter between "saving people" and "don't drive people crazy". The standard has been given to the ruler, and whoever changes it first, is less abused.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.