I get asked this question several times a month.
The conclusion first: there is no procedure called an "AI customer service filing" — but before launch you do have to clear three gates of compliance.
This piece lays out what the three gates are, and then gives you a checklist.
(Everything below is a summary of the current state of play; always confirm against the latest requirements of your local cyberspace and market-regulation authorities.)
Gate 1: Identity and disclosure
When an AI is talking to a customer, the customer has the right to know the other party is not a human.
Many owners think that is unnecessary — "if we say it's a bot, the customer will just stop engaging". The reality is the opposite. State the identity clearly, expectations line up, and the bad interactions later on drop sharply.
Three self-checks:
- At the start of the conversation, is it clearly disclosed that this is an intelligent customer service?
- Is there a human-agent entry available at any time? How hard is it to reach?
- After the transfer, does the customer have to repeat the whole story from the beginning?
The third one is the most easily overlooked — and the most damaging to experience.
Gate 2: Content compliance
What customer service may and may not say has boundaries.
Five self-checks:
- Is there a red-line word list? (extreme words, commitment-type phrasing, tone landmines)
- Are the red lines written into the system, or into the employee handbook?
- Are there hard trigger conditions for automatic transfer to a human?
- For pricing, delivery times and after-sales wording, can you guarantee consistency across every channel?
- If something goes wrong, can you replay the complete conversation at the time?
The last one is the floor. In the era of full-volume quality checks, "what was said at the time" and "do you have evidence of it" are two different things.
There are two ready-made references you can align your content compliance against:
First, the national standard GB/T 47746—2026 ("Customer contact services — Requirements for collaboration between human and intelligent customer service" (《顾客联络服务 人工与智能客户服务协同要求》), effective 1 September 2026). It names five scenarios where AI must hand over to a human automatically; the standard runs to 61 requirements — 48 shall, 4 should, 9 may, including 5 deal-breakers. It is a recommended standard, but when a platform or a customer asks how you measure up, it is currently the most direct reference table.
Second, the platform-side customer-service quality-check rules. E-commerce platforms now run full-volume inspection with per-order penalties — 200 to 2,000 yuan per order. The exact rule details follow what each platform publishes in its seller dashboard.
Gate 3: Data and security
Customer service conversations are full of customer data: addresses, phone numbers, order numbers.
Four self-checks:
- Is customer information collected on a minimum-necessary basis? Do not ask for fields you do not need.
- Can it be deleted? When a customer asks to have their data deleted, do you actually have a way to clean it out?
- Are the permissions on sensitive knowledge-base content (price lists, suppliers, internal scripts) under control?
- Are you using a third-party large model? Will the data cross a border?
That last one deserves particular attention.
A third-party large model — one of the general-purpose AIs you have heard of, the ChatGPT or ERNIE type — boils down to a single question: does your conversation data leave your control, and does it leave the country?
For generative AI services offered to the public, whether algorithm or generative-AI-related filings or registrations are required is a clearly defined area. Confirm against the standards of your local competent authority — do not decide it yourself.
The Checklist (12 items, take it and use it)
| # | Check | Status | |----|----------------------------------------------------|--------| | 1 | AI identity disclosed at conversation start | ☐ | | 2 | Human transfer available anytime, entry prominent | ☐ | | 3 | After transfer, the customer is not asked again | ☐ | | 4 | Red-line word list exists and lives in the system | ☐ | | 5 | Hard trigger conditions for mandatory handover | ☐ | | 6 | External wording (price / delivery / after-sales) unified | ☐ | | 7 | Full conversation replayable as evidence | ☐ | | 8 | Customer information collected on a minimum basis | ☐ | | 9 | Customer data deletable on request | ☐ | | 10 | Knowledge-base permissions tiered | ☐ | | 11 | Data flow clear (whether it crosses a border) | ☐ | | 12 | Someone owns periodic review and updates | ☐ |
If all 12 are ticked, you can launch with a clear conscience.
If you are only halfway there, do not rush it.
Final Word
Compliance is not a luxury for big companies.
The fewer people you have, the less you can absorb a single fine, a single complaint, a single data incident.
Get the wording straight first, then make the system run.
How many of the 12 items did you tick? Where are you stuck? Tell us in the comments.
If this was useful, give it a like so more business owners see it.
Who we are: we help small and micro businesses build enterprise knowledge bases and AI customer service that actually ships — turning the wording, rules and scripts scattered in employees' heads into a traceable base the AI can look up, plus red-line interception and human-handover judgment.