A foreign AI company reaches the point where a Chinese filing is due, asks three providers for proposals, and receives three documents that cannot be compared. One comes from a law firm, one from a consultancy with a Beijing office and a well-maintained public account, one from a candidate the company is considering hiring. All three describe the same scope of work in the same vocabulary, and the prices differ by a factor of four.

The difference that matters is not the expertise each claims but what each can put its name to. A law firm can sign an engagement, issue an opinion under its seal, appear for you when a matter turns adversarial, and answer for the work through a professional liability regime. A consultancy can sign a services contract and deliver a report, and its liability is whatever the contract says, usually capped at the fee. An employee signs nothing externally, and the company absorbs the consequence of the work. These are three different products. The market prices them differently because they are different.

The practical conclusion is to buy the licence where liability attaches and the operations where volume accumulates. Foreign companies get this wrong in two directions. They use a consultancy for a filing whose legal position later has to be defended, or they use a law firm to run a training cycle and a quarterly policy refresh that an operations team could run for a fraction of the cost. What follows sets out how to tell the three apart inside a proposal, how to write a scope that makes the comparison real, and where each model breaks under pressure.

The obligations you are buying help with

AI compliance is not one obligation. It is four layers of obligation with different triggers, different document sets and different consequences, plus a horizontal overlay on data movement. A proposal that prices “AI compliance” as a single line has not mapped your product, and the mapping is the first thing you are paying for.

LayerWhat triggers itWhat you must produceWho administers it
Data and personal information baseAny processing of personal information of people in China, and any processing of important dataPrivacy notices, consent and separate-consent records, internal rules, security measures, incident response, a record of processingCAC and the sector regulators
Algorithm rulesAn algorithm that recommends content or generates synthetic media, where the service has public-opinion attributes or social mobilisation capacityAlgorithm filing, a self-assessment report, disclosure of the algorithm’s basic detailsCAC
Generative AI servicesA generative service offered to the public in ChinaSecurity assessment, algorithm filing, lawful training data, content moderation, user identity verificationCAC
Content labellingSynthetic output shown to usersVisible labels, embedded metadata, disclosure at the point of use, under rules in force since 2025CAC, with MIIT involved in distribution
Cross-border transferAny transfer of personal information out of China, including remote access from an overseas head officeSecurity assessment, standard contract filing, or certification, on top of a transfer mapCAC, provincial level

Each layer attaches to a different feature. A recommendation feed on an e-commerce app is an algorithm matter even if nothing is generated. A generative assistant that writes marketing copy is a generative service matter even if it recommends nothing. An internal tool used by forty employees in a Shanghai office may fall outside the public-service measures and still sit squarely inside the personal information rules, because the employees are people in China. The first deliverable from any provider should be a table of this kind, populated with your product’s actual features and marked where the answer is uncertain.

Two misconceptions cost foreign companies time. The first is that filing is an approval. It is record-filing: the authority registers the algorithm, publishes a summary, and retains the power to require rectification afterwards. A filing number is not a licence to operate, and it does not insulate you from a later finding that the service is non-compliant. The second is that the filing is confidential. The published summary identifies the algorithm, its type and the product it runs in, which is readable by competitors and by journalists. Companies that treat the filing as a private matter write descriptions they later regret.

The overlay matters more than foreign teams expect. Remote access to a Chinese subsidiary’s customer database from a Singapore or Frankfurt head office is a transfer, even if no file is copied. Support tickets, HR records, and the logs your engineers pull during an incident all move data. This is the layer where the provider’s judgement is worth the most, because the thresholds are volume-based and were relaxed in 2024, and the correct route depends on numbers that only your systems team can produce.

What each provider can sign, file and be liable for

The engagement with a law firm is with the firm, not with the lawyer you met. The firm holds a practice licence issued by a provincial justice department; the individual holds a practice certificate; both numbers can be checked against the public register. The firm issues opinions on its letterhead under its seal, and that document is what an investor’s counsel, a lender or a counterparty will accept during diligence. The firm appears for you in an administrative hearing and in court. It carries professional liability insurance, and its lawyers answer to the bar association and to the justice department for conduct, which is a sanction that no consultancy faces.

What the firm cannot do is be the applicant. The applicant in an algorithm filing is your Chinese entity, the filing account is opened in the entity’s name, and the entity signs. A law firm prepares the file, drafts the self-assessment, writes the technical description, and handles the correspondence that follows, but the legal personality on the record is yours. This is not a technicality. It is the reason the quality of the drafting determines your exposure rather than the firm’s.

A consultancy is a registered company whose business scope typically reads “enterprise management consulting” or “information technology consulting”. It signs a services contract, produces policies, training, process design and often the technical annexes of a filing. It cannot issue a legal opinion, cannot appear for you, and is not subject to the lawyers’ disciplinary regime. Its liability is contractual, and the contract usually caps it at the fee and excludes consequential loss, which means the clause that matters is the one your procurement team normally skips.

There is a workable test. Ask who signs the document that goes to the regulator, and ask what happens if the authority rejects it. If the answer is that the consultancy prepares and you sign, the consultancy is doing staff work and you are carrying the legal risk. That can be a rational purchase at the right price. It is a bad purchase when it is presented as advice.

An in-house hire is an employee, and the company is liable for the work while being unable to claim ignorance of it. A Chinese national holding a practising certificate can be registered as a company lawyer, which permits practice for the employer and nothing else: not for the founders personally, not for a sister company, not for a joint venture without a separate arrangement. A foreign national cannot hold a PRC lawyer licence at all, so an in-house legal hire who is not a Chinese national is a contracts and operations professional rather than a licensed adviser, whatever the job title says.

FunctionLaw firmCompliance consultancyIn-house hire
Contract counterpartyThe firm, under a practice licenceA company, under a business licenceAn employment contract
Can issue a legal opinionYes, under seal and signatureNo; can deliver a reportNo, unless registered as a company lawyer and only for the employer
Applicant or agent in a filingPrepares; the entity applies; can be authorised to correspondPrepares; has no standing before the authorityPrepares and applies as the entity’s employee
Receives rectification noticesYes, if appointed to receive serviceRarely; notices go to the entityYes, as the entity’s own staff
Appears in litigation or a hearingYesNoOnly if also qualified and separately engaged
Professional insuranceYesGenerally not; contractual caps insteadNone beyond employment
Disciplinary exposureBar association and justice departmentConsumer complaint, civil claimInternal discipline only
Who bears the loss if the advice is wrongThe firm, within the insurance and the retainerWhoever the contract says, usually youThe company, and then the individual internally
Time to startDaysDaysThree to six months to hire

A word on confidentiality, because boards arrive with the wrong assumption. Chinese law places a statutory confidentiality duty on lawyers and gives a limited right not to testify about client secrets. It does not create a common-law style privilege that allows a company to withhold advice from a regulator or resist production of the file. Plan accordingly: keep factual records separate from advice, avoid asking a lawyer to warehouse engineering data you would not want disclosed, and do not assume that marking a document “privileged” changes how a Chinese authority will treat it. The confidentiality duty still has commercial value, and it is materially stronger than a consultancy’s, which lasts only as long as the contract does.

Who is in scope, and why the applicant is always your Chinese entity

The trigger is the service, not the company’s address. If people in China can reach the product, the Chinese rules apply to the operator, and the operator is normally expected to be an entity established in China. That entity needs a business licence whose scope covers the activity. A scope that reads “technology development, technical consulting” and says nothing about an internet information service will need amendment, and amendment takes weeks you may not have.

Distribution adds a second gate. App stores in China require an app filing tied to the domestic entity and to the domain or package, and they enforce it. Companies that plan to launch through a Chinese app store and only then look at filings discover that the store will not list the app, which is a commercial failure before it is a regulatory one. Confirm the current filing route with the access provider and the local authority, because practice at the counter moves faster than published guidance.

The hard case is the foreign company with no Chinese entity that wants to serve users in China directly. There is no filing account to open and no applicant to name. The realistic routes are to establish an entity, to appoint a domestic operator who becomes the service provider of record, or to restructure the product so that the Chinese-facing service is run by a partner. Each has commercial consequences beyond compliance, and the choice should be made with a lawyer rather than a consultant, because it allocates liability between you and the partner.

Sector overlays change the answer again. A model sold to banks, hospitals, schools or state-owned buyers meets procurement requirements and sector rules that sit on top of the general stack, and the buyer’s own compliance team will audit you. A model embedded in a hardware device brings product certification and radio approvals into the same launch plan. None of this appears in a generic compliance proposal, and its absence is a useful signal.

Some work is genuinely out of scope. An internal tool that no member of the public can reach, used only by employees, is unlikely to be a public-facing generative service, though the personal information rules still govern the employee data it touches. A model licensed from an upstream provider and used inside a closed enterprise workflow may sit outside the public-service measures, depending on how the regulator reads the arrangement. These boundaries are drawn in practice. The correct move is to put the question to the district authority through your lawyer before launch, not to reason it out from the text and hope.

Scoping the engagement

A scope that cannot be compared is not a scope. The proposal should contain ten elements, and a firm that cannot produce them at proposal stage has probably not done the work before.

  1. The entity and the product, named, with the version and the launch date the work is planned around.
  2. The obligations in scope, mapped to the instruments in the table above, with anything excluded listed as excluded.
  3. Each deliverable as a defined output: a memo of a stated length, a policy set with named documents, a filed application, a training session with a stated audience.
  4. The assumptions the fee rests on: whether the model is self-trained or licensed, whether personal information is processed, whether any data leaves China, whether the service is public or closed.
  5. The exclusions: translation, notarisation, official fees, technical testing, model evaluation, code review, engineering implementation of labels.
  6. Your obligations, with dates, and the name of the person on your side who supplies each item.
  7. The number of regulator rounds included. One round of comments is not the same product as three.
  8. Change control: what triggers a re-quote and at what rate.
  9. Acceptance: what makes a deliverable complete, and who decides.
  10. Named staffing, with the proportion of time from the partner and from the associate who actually drafts.

The document list is the fastest test of experience. Ask, before signing, for the list of documents the provider will need from you. A firm that has run filings will produce it from memory in a day: business licence, articles, the algorithm’s design description, the training data provenance record, the content moderation rules, screenshots of the user interface, the privacy policy, the user agreement, the security assessment materials. A provider that asks you to send “whatever you have” is learning on your file.

The deliverable set is equally standard, and its absence from a proposal is informative. A proper engagement produces a gap assessment that maps each feature to the obligation it triggers, with a rating and an owner; a data map that covers the product, the CRM, the support desk and HR, with the transfer points marked; a transfer memo that selects the route and states the volume assumptions behind it; a policy set comprising the privacy notice, the user agreement, internal data rules, incident response, a retention schedule and vendor terms; the filing pack, meaning the technical description, the self-assessment, the disclosure text and the evidence annex; a labelling specification that engineering can implement, covering the visible label and the embedded metadata; a launch readiness note, signed by the firm, that lists what remains open and who accepted the risk; and a training record with a monitoring calendar. A proposal that offers “compliance documentation” without naming these items is selling an unknown quantity, and the price tells you nothing about it.

Define acceptance in writing and tie payment to it. A compliance memo that restates the regulation is not a deliverable; a memo that lists, feature by feature, what must change before launch, who owns each change, and what the regulator will ask about it is. If the proposal cannot describe the difference, the price is the only variable you will be able to negotiate, and price is the least informative variable in this market.

Pricing: four fee models and what drives them

Chinese providers price the same work in four ways, and the model tells you what the provider expects the matter to become.

ModelHow it is calculatedSuitsWhere it goes wrong
Fixed project feeOne number for a defined outputA filing, an assessment, a policy setAssumptions are unwritten, so every surprise becomes a variation
Staged feeMilestones, for example assessment, submission, acceptanceWork with a regulatory gateThe final stage may carry a success element the provider cannot control
HourlyTime at rates that vary by seniorityInvestigations, disputes, unclear scopeTransfers the risk of the provider’s inefficiency to you
RetainerA monthly amount for availabilityOngoing monitoring and ad hoc questionsThe excluded list matters more than the number

Published ranges in this market, as professional fees and excluding third-party costs, run from roughly RMB 20,000 to 30,000 for a basic assessment, RMB 30,000 to 80,000 for filing support depending on complexity, and RMB 50,000 to 150,000 for a full programme around a new product launch. Treat those as anchors, not quotes. The spread inside each band is driven by factors you can inspect.

The number of products and algorithms is the first driver. One recommendation feed and one generative assistant are two filings with two technical descriptions, even if they sit in the same app. Whether the model is self-developed or licensed is the second: a licensed model shifts some obligations upstream, but you remain the service provider to your users, and the upstream contract has to be read before the filing is drafted. Whether personal information is processed, and at what volume, determines whether a cross-border route is needed at all, and which one.

Translation is a cost line that surprises foreign clients. Filings are in Chinese and the Chinese text governs. A technical description written in English by your engineers and translated by a vendor, then reviewed by a paralegal, will produce a supplementation notice. Budget for the lawyer who will defend the text to be the person who reviews the translation, and expect that to be priced.

Third-party costs sit outside the professional fee and are routinely omitted from the headline number: notarisation and apostille of parent company documents, certified translation, searches, travel to the authority, and the engineering time your own team spends producing screenshots and test sets. A proposal that quotes one figure and excludes all of these is not cheaper. It is incomplete.

The commercial consequence of choosing on price alone is measurable in one place: rework. A filing prepared from a template is rejected or, worse, accepted and then challenged when the product changes. The second attempt is paid at the second provider’s rate, the launch slips, and the delay usually costs more than the difference between the two quotes. Companies rarely repeat the mistake, which is why the second engagement is almost always with a firm.

Documents, filings and timeline

StepAuthorityTypical timingWhat foreign companies miss
Entity and business scopeDistrict market supervision administrationTwo to four weeks for a new entity or a scope amendmentThe scope must name the service; “technology development” does not cover an information service
App or website filingMIIT, through the filing systemTwo to six weeksApp stores will not distribute an unfiled app, and the filing is tied to the domestic entity
Data mapping and personal information buildInternal, no filingThree to six weeksTeams map the product and forget HR, CRM and customer support systems
Cross-border transfer routeCAC at provincial levelFour to twelve weeks, depending on routeThresholds are volume-based and were relaxed in 2024; recalculate rather than reuse last year’s memo
Algorithm filingCACTen to thirty business days of review after submission, plus supplementationThe technical description is assessed as well as the form; vagueness produces a notice
Generative AI security assessmentCACBefore launch, running alongside the filingThe assessment tests content safety, and the test set has to be built by someone
Labelling implementationInternal, with product and designTwo to four weeks of engineeringLabels are visible and embedded; the metadata side is the one that gets missed
Change filingCACOn a material changeA model swap or a new use case can require a fresh filing; teams treat the first one as permanent

Two timing observations from practice. A company with clean documentation and a product team that can answer questions quickly can be filing-ready in four to six weeks. A company starting from nothing needs two to three months to build the programme before any submission, and the submission itself then takes the review period on top. Plan the launch date backwards from the review period, not from the drafting.

Where each model fails

A law firm fails in a predictable pattern. The partner who wins the engagement attends the first two meetings and then hands the file to an associate carrying four others. The opinion is written to be defensible rather than usable, so it recites the rules and declines to make the commercial judgement the client needed. Fees creep because the scope said “filing support” without saying how many supplementation rounds were included, and the third round is billed. None of this is misconduct. It is what happens when a firm is bought for its brand rather than for the person who will do the work, and the remedy is named staffing in the engagement letter with a stated proportion of partner time.

A consultancy fails in a different pattern. The policy set is drafted from a template that does not match the product, so the privacy notice describes data flows that do not exist and omits the one that does. When the authority issues a rectification notice, the consultancy has no standing to answer it and the client discovers that the person who wrote the documents cannot sign the response. The consultancy also carries a structural conflict that is rarely disclosed: many are commercial partners of a cloud or platform vendor and steer implementation towards that vendor’s stack, which may be the right answer commercially and is not an independent one.

An in-house hire fails in ways that are organisational rather than personal. Hiring takes three to six months, which is longer than the compliance runway of most launches. The first hire is usually a mid-level generalist given a mandate that spans filings, contracts, employment and data, and no one in the building can benchmark the work. The compliance lead reports to the commercial team whose launch is being delayed, which is a reporting line that produces optimistic assessments. And the company lawyer cannot act for the founders or for affiliates, so the group still needs external counsel at exactly the moments that matter.

The fourth model, which foreign companies do not always recognise as a model, is the vendor package. A cloud provider offers a filing template set, a labelling module and a data residency option, and the package is genuinely useful engineering. It is not advice. The judgement about whether your service is public, whether the model change is material, and whether the transfer route is correct remains with you, and a vendor’s support team will not accept it.

The last failure is the hybrid without a boundary. A firm builds the programme, a consultancy runs it, and nobody owns the file. The consultancy changes a policy the firm drafted, the firm is not told, and the version submitted to the authority differs from the version the engineers implemented. Fix this with a named owner and a single document register before the second provider is engaged, not after.

Sequencing the work between providers

The work has a natural order, and each stage has one party who should lead it.

StageWorkLeadYour sideOutput
ScopingMap features to obligations, identify gapsLaw firmProduct and engineering leadsA gap table with owners
DesignDecide the entity, the operator, the transfer routeLaw firmFinance and corporate developmentA structure memo and a decision
RemediationChange the product, the notices, the retentionEngineeringProduct, with legal reviewImplemented controls
DocumentationPolicies, user agreement, internal rulesLaw firm drafts; consultancy may operationaliseHR, customer supportA controlled document set
FilingTechnical description, self-assessment, submissionLaw firm prepares; the entity submitsEngineering evidenceAn accepted filing
LabellingVisible and embedded labelsEngineeringDesign and contentA tested release
TrainingStaff awareness and process drillsConsultancyManagersTrained staff and a record
MonitoringRegulatory tracking, periodic reviewConsultancy, with the firm on callCompliance ownerA quarterly review note
In-house buildRecruitment and handoverYou, with the firm advisingFounders and HRA functioning team

Two rules make the sequence work. Nothing is filed before remediation is complete, because a filing that describes controls you have not built is a false statement on the record. And nothing is handed to a consultancy until the firm has signed off the document set, because the consultancy will edit what it is given and the edits are not reviewed by anyone unless the sequence requires it.

Practical checklist

The choice between a firm, a consultancy and a hire is really a choice about who carries which risk. Liability-heavy work, meaning filings, opinions, regulator correspondence and anything that may become a dispute, belongs with a licensed firm. Volume work, meaning training, monitoring and process maintenance, belongs with an operations provider once the framework exists. The in-house hire belongs at the point where the volume justifies a salary and the company has enough external history to benchmark the person it recruits.

Our AI compliance practice covers the assessment, the document set, the filing and the correspondence that follows, and our algorithm filing service is built around the technical description that the authority actually reads. For the surrounding questions, the guide to choosing a China AI lawyer sets out how to test a firm’s record, the note on AI data compliance risks covers the transfer analysis in more detail, and the comparison of factory lawyers, sourcing agents and law firms applies the same framework to manufacturing engagements. If you would like the scope table for a filing drafted against your product, contact us with the feature list and the launch date.

Frequently Asked Questions↑ Back to article

What criteria should I look for in an AI compliance provider in China?

Key criteria include deep AI regulatory expertise across the full regulatory stack including cybersecurity, data protection, generative AI measures and algorithm filing rules; a practical, business-oriented approach that balances rigor with operational feasibility; bilingual capability among the staff actually doing the work; a proven track record with regulators including completed algorithm filings; transparent fixed-fee pricing with defined deliverables; ongoing support for evolving regulations; integrated legal and compliance capability; and cross-border understanding for foreign company needs.

What are the common mistakes foreign companies make with AI compliance in China?

Common mistakes include treating AI compliance as a check-the-box paperwork exercise rather than substantive regulatory compliance; hiring generalist data protection consultants instead of AI specialists; underestimating the time required (three to six months is typical for a proper program); choosing cost over quality, which leads to enforcement risks far exceeding the savings; and not getting legal advice, since compliance consultants cannot provide legal opinions or represent you before regulators.

What are the differences between a law firm, compliance consultant and in-house team?

Law firms provide licensed legal expertise, can represent you before regulators, and offer legal privilege but cost more per project. Compliance consultants focus on operational implementation at lower cost but cannot provide legal opinions or represent you. In-house teams are fully integrated but have the highest total cost, take 3-6 months to hire, and have limited external perspective.

When should I choose a law firm over a compliance consultant for AI compliance?

Choose a law firm when you need algorithm filing or other regulatory submissions, when you are launching a new AI product and need a compliance program built from scratch, when you have high legal risk and want legally privileged advice, when you want the security of working with licensed professionals, or when you anticipate potential legal issues or disputes.

What is the recommended approach for foreign AI companies entering China?

The recommended approach is a hybrid model: start with a specialized AI law firm to build the compliance program, handle regulatory filings, and ensure legal compliance from day one; add compliance consultants if needed for implementation, training, or ongoing monitoring once the program is built; and build an in-house team as you scale and your compliance volume justifies the cost, with external firms supplementing for specialized or high-stakes work.

How much does AI compliance typically cost in China?

Costs vary by scope: a basic compliance assessment typically starts at RMB 20,000-30,000; algorithm filing support ranges from RMB 30,000 to 80,000 depending on complexity; and a full compliance program for a new product launch usually ranges from RMB 50,000 to 150,000. Fixed-fee arrangements with clear deliverables provide the best predictability.

How long does it take to get AI compliant in China?

The timeline depends on your product and starting point. A company with good existing data practices and documentation might be ready to file in 4-6 weeks. A company starting from scratch typically needs 2-3 months to build a complete compliance program. Algorithm filing itself takes an additional 10-30 business days for review once submitted.

Geason Law Offices AI Industry Research Group | Reviewed by He Qianggao (Lawyer License No. 14403201310092337, 10+ years of practice)