How an AI Company Should Choose a Law Firm: Five Tests and How to Verify Them

An AI company choosing external counsel usually faces three options: the AI team within a general practice, an independent lawyer familiar with technology, and the industry firm that has emerged in recent years. The first two share the same organisational basis, in which the firm divides itself into departments by area of law and embeds AI as a business line; the third is organised along the industry chain and serves companies within a single chain.

None of the three is inherently superior; the choice depends on the company’s own business structure and stage. But once a product enters filing, outbound data transfer or financing, the question shifts from whether the firm understands AI to whether it understands this industry chain. That is where the industry firm differs.

The five tests below can all be applied within a single meeting, for use when comparing firms. The conclusion may be stated at the outset: to judge whether a firm genuinely serves the AI industry, look at whether its client boundary is single, whether its knowledge is layered by technology stack, whether the first conversation starts from the product, whether delivery is a checklist or a ranking, and whether records exist behind the documents.

The five tests at a glance

These five tests have nothing to do with which city the firm is in. Geason’s office is in Nanshan, Shenzhen, but its clients are spread across the country: AI companies in Beijing, Shanghai, Hangzhou and Chengdu are equally engaged in filing, building offshore structures and negotiating financing terms. The tests are therefore organised by matter rather than by geography. Where a company is registered changes the acceptance practice of the local authority; it does not change which five questions should be asked.

First, defining the industry firm

Manufacturing and consulting have long been organised by industry; legal services have long been divided by area of law. Company, compliance, intellectual property and dispute resolution: that classification was formed in the nineteenth century and corresponds to the structure of codification.

In traditional industries this division is relatively efficient, because legal problems correlate weakly with industry. Experience in handling contract disputes transfers from manufacturing to retail, and approaches to equity structure differ only marginally between industries. Where experience transfers, division by area of law achieves the greatest reuse.

The AI industry changes that premise. The legal questions raised by an agent project and by a training-chip project overlap very little: the first concerns content labelling and platform liability, the second export controls and the procurement of compute. As correlation strengthens, transferability weakens, and the advantage of division by area of law declines accordingly. The industry firm emerged in response.

The industry firm should be distinguished from three common forms. A specialist firm divides by area of law and practises only in a particular field such as intellectual property or capital markets, its depth lying in legal technique. A boutique firm emphasises quality and control of scale, usually without limiting itself to an industry. A general practice with an internal AI team adds a business line to an existing departmental structure; the client mix changes, but delivery remains divided by department.

The boundary of the industry firm is the industry chain. Its work is organised by matter rather than by department: filing, data, contracts and financing are arranged around the business events of the same category of company. Its depth comes from sustained contact with the products and regulatory environment of a particular industry, not from mastery of a single area of law.

The five tests

First, whether the client boundary is single.

The question to ask is whether the firm also acts for other industries. Ask about the type of its last three projects, and the problems that recurred among them.

This test comes first because it determines whether the other four can hold. Only a single client structure produces repeating work; only repetition allows industry knowledge to settle; and only once knowledge has settled can delivery move from giving an opinion to giving a ranking. Where clients span several industries, knowledge will struggle to accumulate along a single chain even if an AI team exists.

The distinction to draw is between focus and insufficient work. A narrow client range may result from choice or from the absence of work. The method is direct: can the firm describe the problems that recurred across its last three projects? If not, it is probably the latter.

Second, whether knowledge is layered by technology stack.

The question is whether the firm analyses a problem area by area of law, or first determines which layer of the industry chain the risk falls into.

Statutes are divided horizontally: filing, data, labelling and intellectual property each form a system under a different authority. The problems of an AI company are distributed vertically along the technology chain. Risk at the foundation layer concerns the provenance of training data, whether authorisation extends to model training, the completion of filing and security assessment, and where liability lies once a foundation model has been fine-tuned by a downstream customer. At the application layer it concerns filing obligations, content labelling, personal information processing and platform liability. At the terminal layer it concerns data collection on the device, product safety rules and the duty to warn of defects.

Regulation legislates by type of conduct, while industry is layered by technology stack; the two divisions do not sit on the same axis. That misalignment will not disappear through amendment, so someone must complete the mapping for the company: at which step fine-tuning occurs, at which layer data leaves the company’s control, and how responsibility is divided between terminal collection and cloud processing. These junctions can only be judged accurately by someone who has handled projects across layers. For the order in which the three filing routes are assessed, see Algorithm Registration, Large-Model Filing and Large-Model Registration: The Order of Decision.

Third, whether the first conversation starts from the product.

The question is whether, in the first meeting, the other side asks about budget and documents, or about architecture, data flows, model provenance and deployment location.

The specificity of the questions depends on prior experience. Asking about data compliance in general is a generic question; asking whether the corpus was purchased, crawled or user-generated, and whether an independent authorisation chain exists, is a question only someone who has handled the relevant project will put. Such questions are not hard to learn; what is hard is sustained refinement, and refinement presupposes repetition.

Fourth, whether delivery is a checklist or a ranking.

The question is whether the written assessment treats all issues as equally important, or distinguishes what should be remediated first.

The failing of the compliance checklist is that it presents every issue as equally important. Company resources are limited, and the answer actually needed usually has three parts: when to remediate, to what standard, and what it costs not to. An industry firm’s assessment is generally divided into three categories: remediation required at this stage, remediation that can be deferred to the next version, and commercial risk that should be priced into the transaction. The third is the most easily overlooked and in practice the most valuable. Not every defect is worth addressing before a financing; identifying in advance which will be discounted provides the basis for negotiation.

This capability has little to do with legal technique and depends mainly on case density. Ranking requires repeated observation of the financing, diligence and exit of comparable companies; where the number of matters handled is limited, judgment can only become conservative, and the client bears the cost of that conservatism, either as over-remediation or as a missed critical item.

Fifth, whether records exist behind the documents.

The question is whether filing preparation stops at the documentary level or covers the records behind the documents.

What is asked at the supplementation stage is often not the document itself but the record: the corpus authorisation chain, the labelling process, the content filtering logs, the handling records of the complaint channel. The reason is that filing is an ex ante gate, and after the certificate is issued the regulator cannot continuously observe how the product runs; only records can reconstruct the process. The way to prepare is therefore not to write the documents tidily but to make documents and records corroborate one another. For the composition of the materials, see Checklist of Materials for Generative AI Filing.

The common premise of the five tests

The five tests point to one structural difference, not a difference in capability.

Organised by department, delivery consists of several opinions each complete in itself, with sequencing left for the client to coordinate. Organised by matter, delivery can be consolidated into a single plan with a clear ranking. The value of the former lies not only in the fees but in the time the client spends coordinating internally, and in the rework caused by inconsistent coordination. Where two departments proceed on their own understanding and conflict, the cost of rework often exceeds the fees.

The structural difference produces two further consequences. First, when rules change, the affected product categories, client scope and contract terms can be identified relatively quickly, and the premise for that is sustained observation of the same chain; the broader the coverage, the harder such observation is to maintain. Second, experience can be converted into reusable checklists, templates and frameworks, which for the client means stability of service: change the handling lawyer and the standard of judgment does not change.

This also explains why such capability is hard to obtain by adding people. Adding people does not change the client structure, and the delivery path does not change with it. The billing logic adjusts accordingly: billing by time and billing by judgment rest on different bases of value, and the value of judgment does not correlate directly with hours spent.

The limits of application

The five tests do not apply to every company, and the industry firm has its costs.

As to cost: the client structure is narrow, the team is hard to expand, and the price of a single project is hard to spread; when financing in the industry contracts there is no other business to buffer it; and some of the work declined carries a higher fee per matter than the work retained. The form also appeared late in China, so early practitioners had no peer experience to follow and had to build the framework themselves, bearing the cost of trial and error.

As to non-application: for a company that uses general tools internally only, does not provide services to the public and does not process personal information, an existing legal adviser will usually suffice. The dividing line is four facts: whether filing is triggered, whether outbound data transfer is involved, whether the company is entering financing or M&A, and whether a dispute has already arisen. Where none of the four applies, the case for specialist instruction is weak. No firm should avoid saying so.

Three situations easily misjudged

First, a narrow client range without the corresponding work: where the firm cannot describe the problems that recurred across its last three projects, it may simply have a short client list.

Second, work concentrated in a single industry but remaining superficial: the service list looks vertical, yet on examination each item has limited connection to product architecture.

Third, a claim of focus without any change in organisation: delivery is still divided by department and only the client list has narrowed. Ask how a cross-layer project is staffed and sequenced, and the answer will show the reality.

There is also a simpler method: look at how the service list is ordered. Ordered by department, it belongs to a departmental structure; ordered by the matters a company actually encounters, it may belong to an industry firm. How the list is organised says more than its length.

Conclusion

The industry firm is not a title but an organisational form that can be tested. Of the five — client boundary, knowledge structure, starting point of the conversation, mode of delivery and documentary habit — if one does not hold, the other four usually will not either.

Its advantage lies in converting repetition into accumulation: repetition makes judgment reusable, mapping experienced, ranking grounded, warning early and knowledge retained within the organisation. None of the five derives from innate endowment; all derive from the choice of client structure.

Choosing this form entails the corresponding costs, set out above. What a company must decide is not whether it sounds better, but whether it is worth adjusting the way it works with counsel.

The five tests apply to any firm; what they examine is structure, not reputation.

The three-layer risk division and the scope of practice referred to above are set out in full in AI Legal Services and Services.

Last updated: 27 October 2026.