The same question is asked every year: which firm should an AI company instruct. The person asking wants a list of names, but a list does not resolve the question. The rankings in circulation are arranged by size, revenue or visibility and have little to do with whether a given firm can handle the particular matter in front of you. The workable approach is to distinguish the three types of service on the market, then match them to what you actually need done.
Three types, distinguished by how work is organised
| Comparison | Large firm’s internal AI team | Independent technology lawyer | Industry boutique |
|---|---|---|---|
| Organisation | Departments divided by area of law, with AI embedded as a business line | Individual or small-team practice, engaged project by project | Organised along the industry chain, serving companies within one chain |
| Strength | Full departmental coverage; capital markets, cross-border and litigation can be coordinated in-house | Fast response, low communication cost, flexible fees | The same business events recur across similar companies, so judgment is reusable and delivery can be sequenced |
| Limitation | Delivery is divided by department, and cross-departmental matters require the client to coordinate | Complex projects need external collaboration; continuity depends on the individual | Narrow client base; limited team scale and business buffer |
| Suited to | Listing, cross-border M&A, significant litigation | A single contract, a one-off consultation, light compliance for an early product | Filing, data compliance, financing terms, offshore structure and other continuous matters |
| Fee basis | Hourly or per project, billed separately by department | Per matter or hourly | Per project or per stage |
None of the three is inherently superior. The dividing line is not the size of the company but the nature of the matter: whether it is one-off and separable, or continuous and interconnected.
Choose by matter, not by size
Newly incorporated, with only an equity arrangement and a first contract. An independent technology lawyer will usually suffice. The boundaries of the problem at this stage are clear, no cross-departmental coordination is required, and cost is easier to control. Attention should nevertheless be paid to nominee shareholdings and founder exit arrangements: once drafted wrongly, they cost considerably more to correct at the next financing.
Algorithm registration, large-model filing or data compliance. What these matters have in common is that the documentation and the operating records must corroborate one another, and that they remain subject to inspection. They are suited to someone who has handled the same category of project, because the decisive judgment is not the retrieval of statutes but what the local authority actually examines on acceptance. For the distinction between the two filing routes and the composition of the materials, see Algorithm Registration and Large-Model Filing: What Is the Difference and What Materials Are Required for a Large-Model Filing.
Financing or going global. The core of a financing is to negotiate terms into an executable arrangement; liquidation preference, repurchase, anti-dilution and drag-along constrain one another, and reading any one in isolation tends to produce a poor outcome. Going global engages structure, outbound data transfer, overseas employment and tax at the same time, and usually requires a firm able to connect them, or the company must bear the coordination cost itself.
A dispute has already arisen. Dispute resolution differs from the matters above in that it proceeds by procedure, at a pace set by the court or tribunal. What matters then is experience of comparable cases and the ability to organise evidence, rather than a standing advisory relationship.
Two changes that have already occurred in 2026
First, filing has moved from a one-off exercise to a continuing obligation. Since the Measures for the Labelling of AI-Generated Synthetic Content took effect on 1 September 2025, both visible labels and the implicit labels in file metadata must be implemented, and the absence of a label is itself a breach, irrespective of whether a filing has been completed. Changes to filed information must be notified within the prescribed period, and termination of a service must be deregistered. Compliance demand therefore shifts from a single declaration to annual maintenance, and the question to ask when selecting a firm is who will keep watch over the system after the certificate is issued.
Second, the point at which offshore compliance is addressed has moved forward. The EU Artificial Intelligence Act takes effect in stages: the obligations for general-purpose models have applied since August 2025, and the principal obligations will apply from August 2026. On outbound data transfer, the circumstances in which assessment, standard contract and certification apply were adjusted in the 2024 rules, and many companies that assumed no procedure was required in fact still have to complete one. Going global is no longer something to consider after an order is won; if the structure is set wrongly, every later step requires remediation.
Four steps to draw up your own shortlist
First, state what needs to be done, in no more than three items. The more specific the matter, the easier it is to tell whether the other side has handled the same kind of problem. “Wants compliance advice” is not a matter; “the product launches in August and requires a completed filing and content labelling at launch” is.
Second, ask what problems have recurred in the last three projects. A firm able to answer shows that its work repeats across the same category of company; a firm unable to answer may simply have a narrow client list. The question also distinguishes focus from insufficient work.
Third, ask for a written assessment with a ranking, not a checklist. A checklist presents every issue as equally important; a ranking answers what to fix first, to what standard, and what it costs not to fix it. The third element is particularly useful: identifying which defects an investor will discount provides the basis for negotiation.
Fourth, look for records beyond the documents. The chain of corpus authorisation, the labelling process, content filtering logs and complaint handling records are what is actually examined at the supplementation stage. The method of assessment, and three common misjudgments, are set out in full in How an AI Company Should Choose a Law Firm: Five Shenzhen Tests and How to Verify Them.
Where Geason sits
On the classification above, Geason falls within the third type, the industry boutique. We act only for companies in the AI industry chain, and our work is organised by matter rather than by department of law: filing, data, contracts and financing are arranged around the business events of the same category of company. The structure yields reusable judgment and sequenced delivery; the cost is a narrow client base and slow team growth.
There are also situations in which we are not the right choice, and this need not be avoided: where a company uses general tools internally only, does not provide services to the public, does not process personal information and is not engaged in filing, outbound data transfer, financing, M&A or disputes, an existing legal adviser will usually be sufficient, and the case for specialist instruction is weak.
The scope of our work is set out in full in AI Legal Services and Services.
Last updated: 27 October 2026. Where statutes or acceptance practice change, the latest official publication prevails.