You have a SaaS product with an AI agent inside it, and Chinese customers want it. The demo works. The problem is that the moment a Chinese user signs in, several regimes apply at once: telecom licensing for the website or app, security grading for the network, personal information rules for the data, and generative AI content rules if the agent produces text, images or voice. Each has its own regulator, its own filing and its own failure mode, and a gap in any one of them can take the service offline. This guide maps the licences, the data boundaries and the contract terms you need before you switch on production traffic in China.
Who this guide is for
- SaaS companies adding Chinese customers to a product built and hosted elsewhere.
- AI agent vendors embedding third-party models and unsure where responsibility sits.
- China general managers and product leads preparing a launch date.
Stage: Execute. You have a product and a market; the remaining question is what must be in place before launch.
The licence map
China separates the registration of a website from the licensing of a business, and both are often called “ICP” in English, which causes most of the confusion.
An ICP filing is a registration for a website or app hosted in mainland China. It is not a licence and does not authorise commercial activity; it records who operates the site and what it does. It requires a domestic entity, a Chinese domain and a hosting provider in China, and it is typically completed in a few weeks.
An ICP licence is a value-added telecom business licence for commercial internet information services. It is required when you charge users for online information, membership or advertising. Because foreign ownership in most value-added telecom categories is generally capped at 50 per cent, a wholly foreign-owned enterprise usually cannot hold one, and a joint venture with a Chinese shareholder is the usual route. Pilot programmes in selected free trade zones have relaxed some categories, and local practice differs, so confirm the position for your exact service.
An EDI licence covers online data processing and transaction processing. It applies where the platform processes transactions between third parties, such as marketplaces, booking engines and some B2B order platforms. A pure SaaS tool that does not handle third-party transactions usually does not need it.
| Requirement | Trigger | Authority | Typical timing | Foreign ownership |
|---|---|---|---|---|
| ICP filing | Website or app hosted in mainland China | Provincial communications administration (MIIT) | Two to four weeks | Domestic entity required |
| ICP licence | Charging for online information services | MIIT or provincial administration | Two to four months | Generally capped at 50 per cent |
| EDI licence | Platform processes third-party transactions | MIIT or provincial administration | Two to four months | Generally capped at 50 per cent, with pilot exceptions |
| App filing | App or mini-program distributed in China | MIIT | Weeks | Domestic entity required |
| Public security filing | Any website or app live in China | Public security bureau | Within 30 days of launch | Not applicable |
| MLPS grading | Any network operator, Level 2 and above | Public security bureau | Four to eight weeks | Not applicable |
| Algorithm filing | Recommendation algorithms with public opinion attributes | Cyberspace Administration of China (CAC) | Commonly around 30 working days | Domestic applicant |
| Generative AI filing | Public-facing generative AI service | CAC, provincial level | Commonly one to three months | Domestic applicant |
Two further items block launches in practice. Security grading under the Multi-Level Protection Scheme (MLPS) requires you to classify your system, typically Level 2 for ordinary SaaS and Level 3 for large or sensitive platforms, and file the grading with public security. Every app or mini-program must also be filed with MIIT before distribution, which app stores enforce.
Where SaaS ends and a generative AI service begins
The common miscalculation is assuming a SaaS product sits outside the generative AI regime because it is sold to businesses.
If your agent generates text, images, audio or video in response to open-ended user prompts, and the service is offered to users in China, it is likely a generative AI service under the Interim Measures for the Administration of Generative AI Services. That brings a security assessment and filing with CAC, a description of training data and content moderation arrangements, a complaint channel and, in practice, user identification by mobile number.
Embedding a model that is already filed does not move the obligation to the model vendor. You remain the service provider to your users, and provincial CAC offices have in practice required application-layer providers to complete their own filing or registration even where the base model carries a filing. Using an API transfers cost and latency, not responsibility.
A narrower route exists. If the AI function is not open to the public, is confined to named enterprise users under contract and does not produce publicly distributed content, some obligations may not be triggered. That boundary is fact-specific and is the single question most worth a written legal opinion before launch, because it determines months of work.
AI agents and who answers for what they do
An AI agent that books, pays, emails customers or changes records is not just generating content; it is acting. Chinese law has no dedicated statute for autonomous software agents, so responsibility is assembled from several sources.
The operator of the service bears the primary obligation to users and regulators. That means the entity holding the filings, not the model supplier, is the party that answers for harmful output, unlawful content or a transaction the agent executes incorrectly. Where the agent acts on a platform, platform liability rules may add obligations to verify, record and respond to complaints. Where the agent handles personal information, PIPL duties attach to whoever determines the purpose and means of processing.
Three practices reduce the exposure. Keep a human confirmation step, and log it, for any action with legal or financial effect. Record every tool call and model response so a disputed action can be reconstructed. And require the model supplier to warrant lawful training data, give change notices and indemnify third-party claims, knowing such clauses allocate cost rather than eliminate liability.
Data: user data, training data and the cross-border line
In a B2B SaaS deployment you usually wear two hats under PIPL. For your own account, billing and marketing data you are the personal information processor, the Chinese equivalent of a controller. For the end-user data your customer uploads, you are typically an entrusted party, the equivalent of a processor, acting on documented instructions.
Training is where the hats get mixed. Using customer data to improve a shared model is a different purpose from delivering the service, and it requires a separate lawful basis, usually explicit contractual permission with a genuine opt-out. Sensitive personal information, which includes biometrics, financial account data, health data and information about children under 14, requires separate consent and a personal information protection impact assessment. Aggregating or de-identifying the data before training reduces, but does not automatically remove, these duties.
Cross-border transfer is the decisive architectural question. Sending personal information from China to an overseas cloud or model endpoint is an outbound transfer under PIPL, and depending on volume and data type it requires a CAC security assessment, a standard-contract filing or certification. The 2024 cross-border rules raised the thresholds and simplified some routine flows, but the direction of travel remains: data about Chinese users that leaves China needs a route. Important data, and data held by critical information infrastructure operators, must generally stay in China.
The practical answer for most SaaS vendors is a China region or a China-operated instance, so storage and inference stay inside the border and export becomes the exception. Take that decision before the cloud contract, not after.
Content safety, labelling and minors
Content obligations run from input to output and continue after launch.
You need filtering on user prompts and on generated output, with keyword and semantic controls tuned to Chinese-language risk categories. You need a complaint channel with defined response times, a process for escalating serious content to the platform operator and, where required, reporting to CAC. Network logs must generally be retained for at least six months, and content records for a comparable period.
AI-generated content must be labelled under the measures in force since 2025, with a visible label for users and an implicit label in metadata. If the product synthesises faces or voices, the deep synthesis rules add consent, marking and anti-forgery obligations.
Where minors can access the service, PIPL treats information about children under 14 as sensitive and requires guardian consent, and minors’ protection and anti-addiction rules apply to products with social, gaming or livestream features. A business SaaS tool may be unaffected; a consumer-facing agent should assume they apply.
Contracts and launch conditions
The contract set matters as much as the licence set, because most China disputes in this sector are contractual rather than regulatory.
Your terms should identify the contracting entity, define the service boundary, and let you suspend or modify features on regulatory instruction without breaching the SLA. The SLA should exclude connectivity failures, regulator-ordered suspension and standard force majeure events.
Your data processing agreement should follow PIPL’s entrusted-processing requirements: purpose, duration, processing method, data categories, security measures, subcontractor approval, return and deletion, and audit rights. If the customer is the exporter of record, allocate the cross-border filing obligations explicitly, because ambiguity here is where deals stall.
Finally, allocate content and model risk. Cap liability at a level the contract can bear, carve out indemnities for customer-supplied data and unlawful instructions, require notice of material model changes, and provide for orderly data return on termination. Governing law and arbitration seat should be chosen with enforcement in mind, not by default.
What to do next
- Map every feature against the licence table and mark which triggers apply to your product, not to a generic SaaS.
- Decide whether the AI function is public-facing, because that single answer sets the filing route and the launch date.
- Confirm the foreign ownership position for the relevant telecom category before choosing between a WFOE and a joint venture.
- Choose the data architecture first: China region, China-operated instance or offshore with an export route.
- Complete MLPS grading, public security filing and app filing before any marketing launch.
- Stand up moderation staffing, complaint handling and six-month log retention as permanent functions.
- Paper the stack: terms of service, PIPL-compliant data processing agreement, model supplier warranties and liability caps.
- Recheck local CAC practice in your launch city, because provincial requirements for application-layer AI registration still vary.
This guide reflects the position as of late 2026, and both the telecom and AI filing rules continue to evolve. Confirm the current requirements for your specific service with counsel before launch. For the entity question behind the licence map, see What Is a WFOE for AI Companies.
Frequently Asked Questions↑ Back to article
Do we need an ICP licence to run a SaaS product in China?
Not always. A non-commercial website needs only an ICP filing, while charging users for online information services generally requires a commercial ICP licence, which is a value-added telecom licence subject to a foreign ownership cap. Many B2B SaaS products need the filing, security grading and an EDI licence if they process transactions.
Can a foreign company hold an ICP licence in China?
Generally no, because foreign ownership in most value-added telecom categories is capped at 50 per cent. A joint venture with a Chinese shareholder can hold the licence, and pilot programmes in selected cities have relaxed some categories, so confirm the current position for your service.
Does a SaaS product with an AI assistant need a generative AI filing?
If the assistant generates text, images or audio for users in China and the service is open to the public, it is likely a generative AI service under the Interim Measures and needs a security assessment and filing. Embedding an already-filed third-party model does not transfer that obligation to the model vendor.
Who is liable when an AI agent takes a wrong action?
The entity that provides the service to the user is the responsible operator in practice, and content and conduct obligations attach to it rather than to the underlying model vendor. Contract terms can allocate cost between you and your supplier, but they do not bind users or regulators.
Can we keep Chinese customer data on our overseas servers?
Only within limits. Sending personal information out of China is a cross-border transfer under PIPL, which can require a security assessment, standard-contract filing or certification depending on volume and data type. Important data and data held by critical information infrastructure operators must generally stay in China.
How long must we keep user logs and content records?
Network logs must generally be retained for at least six months under the Cybersecurity Law, and content and complaint records should be kept for a comparable period. Retention periods can be extended by sector rules, so set the schedule with counsel rather than by default.
What content moderation is required for an AI agent in China?
You need input and output filtering, a complaint channel with defined response times, escalation to the platform operator and reporting to the Cyberspace Administration of China where required. AI-generated content must also carry both visible and metadata labels under the 2025 labelling measures.