Open Legal AI Workflows: What Chinese Lawyers Should Really Consider
What should Chinese lawyers really worry about?
Anhui Jingwang Law Firm Yu Kun
Introduction: On May 12, 2026, the American artificial intelligence company Anthropic put the entire "Claude for Legal" set for law firms online at once, and anyone can download, modify, and commercially use it for free. This set of tools includes workflow templates for 12 practice areas, more than 80 intelligent assistants that can run in the background, and more than 20 integration interfaces with mainstream legal office software. Colleagues have already published several in-depth interpretations that break down its structure in detail. This article will not repeat the matter. From the perspective of a Chinese law firm, I would like to talk about three more important things: what does this release mean to Chinese lawyers, what are the shortcomings that cannot be copied, and what domestic counterparts should really do.
1. A real scene as the introduction to this article
At 9 a.m. on Monday, I opened my mailbox: a new supplier agreement from a client required a reply on the same day; a party to a certain case sent three new pieces of evidence compiled last night asking for evaluation; a trainee lawyer submitted a revised third draft of agency opinions for my review; a company was introduced by an old friend to discuss cooperation; a new regulation for an industry just released by the regulatory authorities required internal distribution and interpretation... This is a "Monday morning" that any partner who has been practicing for more than 10 years is familiar with.
Every job in this scene is not complicated at all. But together, they squeeze the same thing: attention. Gloria Mark, a cognitive scientist at the University of California, Irvine, has conducted years of follow-up research on knowledge workers in Silicon Valley offices, and has come to a widely cited conclusion: After being interrupted, it takes an average of 23 minutes to return to the complete state of mind of the original task. The level of complexity for lawyers only increases. In other words, after answering five calls in the morning and finding that no real work has been done, this is not because the lawyer is "inefficient", but because the brain itself works this way.
That’s the real context for this release from Anthropic. It is not another “faster contract review AI”, but a clear answer: when lawyers’ judgment is a scarce resource, at which layer of the workflow should AI be placed (the “workflow” here simply means how to do it step by step from receiving a thing to the final delivery).
2. What is this release?
The thing itself is not complicated. Anthropic has disclosed workflow templates in 12 practice areas on the code hosting platform: commercial contracts, corporate affairs (including M&A due diligence and delivery lists), labor and employment, privacy compliance, product compliance, regulatory compliance, artificial intelligence governance, intellectual property, dispute resolution, as well as legal clinics and law students for teaching scenarios, and finally a legal plug-in store. The complete set includes more than 80 intelligent assistants that can run automatically in the background, respectively responsible for specific tasks such as contract review, personal information query request response, and requirement comparison table drafting. There are also more than 20 integration interfaces, connecting Claude to mainstream software such as electronic signature, document management, contract management, legal search, and electronic evidence used daily by law firms.
To use an analogy that is easy for lawyers to understand: it is like having a set of "shortcut tools" classified by business areas on the computer desktop of a law firm (this is the so-called "plug-in"). Each tool comes with a secretary who is available 24/7 to help you run errands in the background (this is the so-called "intelligent assistant"). At the same time, these tools can directly read the contract library, search library and signature process used daily by the law firm (this is the so-called "docking channel").
But what deserves more attention than the plug-in itself are three things.
The first: a complete reversal of the relationship with the previous generation of legal AI companies. In the past, legal AI companies such as Harvey and Legora, which were valued at billions of dollars, were essentially Claude's "middlemen", buying interfaces from Anthropic, packaging them and then selling them to law firms. After this release, Harvey became a common plug-in in the Claude app market. Mark Pike, deputy general counsel of Anthropic, said something very interesting at Stanford Law School: The real division of the race is no longer which AI company beats which AI company, but which legal teams are willing to put in the effort to hand over the workflow and redesign it, and which teams are still pretending that nothing has happened.
The second thing: the entire set of things is open source, which means that anyone can use, modify, and sell it. Anthropic allows commercial use, modification, and redistribution. What’s more noteworthy is that this time it also cooperates with several judicial relief agencies in the United States, and the relevant integration interfaces are free and open to all users. Overseas media characterized it as an "efficiency tool that goes beyond hourly billing for law firms" with industrial significance: for the first time, it is possible for groups who cannot afford lawyers to get structured legal help.
Third thing: All plugin outputs are clearly defined as "drafts for attorney review" rather than legal opinions themselves. Every legal proposition is accompanied by a source of citation; uncertain matters are compulsorily marked as “requiring verification”; conservative default values are adopted when attorney privileges and subjective legal judgments are involved; assumptions about jurisdiction must be explicitly marked; and there is an explicit manual checkpoint before any “submission, issuance, and reliance”. This set of safety guardrails is stricter than many domestic legal AI products.
3. Flaws that cannot be copied: the four most acute ones
I scanned the workflows of the 12 plug-ins one by one, and there are indeed more than a dozen “lightning points” for Chinese lawyers. But among them, I think the following four are really likely to cause professional accidents. The first three of them come from differences in legal systems, and the last one comes from the compliance and confidentiality red lines in China. The rest are more translation-level differences in search tool names, citation formats, etc., which can only be localized.
[1] "Lawyer's Privilege" Trap: British and American privilege is not equal to China's lawyers' confidentiality obligations
This system regards "privilege of communications between attorneys and clients" as an insurmountable red line. However, Chinese law does not have this concept of evidence law under the common law system. China’s counterpart is the “lawyer’s obligation to maintain confidentiality” in Article 38 of the Lawyers Law, which is a lawyer’s professional obligation and is not an evidentiary right that the parties can use to fight against during the production of evidence. The legal consequences of the two are completely different: the British and American privileges can prevent the other party from accessing the information, but China's confidentiality obligations cannot directly prevent the court from accessing the information ex officio. If you follow the logic of "marked privileges mean security" to organize internal document management, serious accidents will occur when litigation occurs.
【2】The "evidence discovery" of US civil lawsuits does not exist in China, and the evidence investigation toolbox is completely different.
Concepts such as "out-of-court depositions," "written inquiries," "admission requests," and "document discovery" frequently appear in this workflow. These are the evidence collection procedures in American civil litigation. There is no corresponding system for civil complaints in China. Commonly used evidence enhancement tools by Chinese lawyers mainly include lawyer investigation orders, applications for the court to obtain evidence ex officio, applications for evidence preservation, notarization and preservation of evidence, applications for investigation and evidence collection, etc. If the "suggested next step of retrieval method" in the template output is not rewritten and given directly to the lawyer, the lawyer will be at a loss as to what to do or even miss the job.
【3】The “criminal blind spot” in trade secret cases
This is the one I personally worry about the most. The protection of trade secret infringement in the United States is mainly civil. China is different. Article 219 of the Criminal Law clearly stipulates that the crime of infringement of trade secrets, causing "significant losses", is punishable by up to three years in prison, and "particularly serious consequences" is punishable by three to seven years. When this system from the United States is used to divert cases, it will only recommend relief paths based on the civil framework and will not trigger criminal prompts. If a lawyer representing an accused party follows his or her "risk assessment" recommendations, the client may miss a critical response window before a criminal investigation is launched.
【4】Data compliance and confidentiality obligations: This is the red line that Chinese lawyers face every day
This is the most easily overlooked of the flaws that cannot be copied, but it happens every day. Uploading the client's case materials, contract text, and personal information of the parties to any overseas AI service triggers at least four domestic rules at the same time: the lawyer's confidentiality obligations in Article 38 of the Lawyers Law, the restrictions on the cross-border provision of personal information in the Personal Information Protection Law, the security assessment requirements of the Data Security Law on the export of important data, and the confidentiality clause that "no disclosure to third parties without the written consent of the client" is expressly written in most client entrustment contracts. Anthropic's system is designed for U.S. law firms. The default premise is that the law firm has a data processing agreement and enterprise-level data control arrangements. If Chinese lawyers directly run client contracts on overseas platforms, the risk of violation and professional discipline risk will appear before "whether it is used well or not." This one is more directly related to "dare to use it" than the first three.
The remaining nine articles (including citation specifications, exclusion of mediation and settlement communication evidence, handling of breach of contract in batches, prosecution acceptance standards, judgment precedent validity, fair use judgment methods, trademark confusion judgment factors, replacement of legal search tools, American Bar Association Law School Clinic Rules, etc.) are relatively low-risk paradigm differences that can be re-mapped by experienced Chinese lawyers. They are important, but not fatal.
The more fundamental judgment is: the legal layer of this system cannot be used directly, but the workflow layer can be learned, and should be learned seriously. This is what I really want to say in the next two sections.
4. Three designs that are truly worth learning
Taking this system apart and stripping away the specific legal provisions and search tools, the remaining three designs are universal across legal systems and are equally applicable to Chinese law firms, and are even rarer. I rank them in order of relevance from highest to lowest.
Design 1: Turn the firm's accumulated practical experience into explicit, editable guidance
When each Claude plug-in is first installed, it will first run a "cold start interview" program. This word may sound mysterious, but in fact, what they do is very simple: just like when a new colleague reports for work on the first day, the director will sit down with him and explain "how our office works, to what extent the contract is reviewed, what situations require instructions, and what red lines must not be stepped on." The cold start interview is to turn this conversation into a structured question and answer and precipitate the answers into a team file. After that, every time a lawyer uses any specific function, the first thing the system does is read the team profile.
The original words of Anthropic’s product principle are: Lawyers should never see the configuration file that engineers see, they should see a document about their team that can be directly edited in vernacular.
This design has an almost one-to-one correspondence in the context of Chinese law firms:
● Risk preferences of contract terms: liquidated damages clauses, dispute resolution methods, confidentiality periods, ownership of intellectual property rights... Each firm and even each partner has its own caliber, but today these calibers are mainly "hanging in the head of the director."
● Comparison table of who should be consulted for final decision: Which clauses deviate from the clauses can be judged by the trainee lawyer, which ones must be signed by the sponsoring lawyer, and which ones must be counter-signed by the general counsel or partners. All formal offices actually have "conventions" on this matter, but they are rarely documented.
● The preferred style of documents within the firm: whether it is offensive or restrained, whether the defense statement "refutes first and then breaks down" or "breaks first and then refutes", and whether the legal opinion must be accompanied by a risk level.
Making these things explicit, documented, and versioned is itself the most important institutional asset of a law firm. Its value does not depend on AI. Even if AI is not used, the cost of educating new lawyers will drop significantly. This profile is the true starting point for AI personalization once any AI tool is subsequently plugged into it.
Design 2: Clearly state “where AI is confident and where AI is unsure”
In this system, several labels are mandatory at the end of each functional output: which facts require manual verification, where there is uncertainty in reasoning, and where citation sources are missing. For weak arguments, it will be directly marked "Where is this point weak, and the alternative path available is A, B, or C." The entire output ends up giving the sponsoring attorney a 5-option “what to do next” decision tree.
The methodological value of this matter is greater than the legal value. Among the dozens of trainee lawyers I have taught in the past, the most common problem is not that "the level is not enough", but that "I am sure" and "I guess" are mixed together. Reviewers need to spend more than twice the time to identify which sentences are based on evidence and which sentences are speculation. This is the source of a lot of review conflicts and rework.
Using "trustworthy or untrustworthy markers" as a mandatory format for output is a design that makes both humans and AI "honest". Our firm's internal draft review has begun requiring trainee lawyers to mark their first drafts with three labels: "verified", "requires verification" and "pure inference". The effect is remarkable.
Design 3: Let the review manual iterate on its own
There is a little assistant running in the background called "Review Manual Supervisor" in the commercial contract plug-in. It tracks the deviation log within the firm over a long period of time: if the same type of clause is deviated 5 times in 3 months, a manual update proposal is automatically generated for review by the sponsoring partner. This is a self-iteration of the review manual.
This matter is of greater significance to Chinese law firms than to American law firms. Once the contract review guidelines and position manuals within Chinese law firms are finalized, it often takes one and a half to two years for the next complete revision. There were a large number of deviations from the original guidelines in the contracts that were internally audited during the period, but these were usually not recorded by the system, summarized and analyzed, or updated. During the final revision, the lead writing partner could only recall from memory and sporadic cases "what situations have exceeded the original guidelines in the past two years."
This is a common problem among Chinese law firms that “experience cannot be accumulated.” A considerable part of the judgments that a senior lawyer makes every day have not been settled. Once they are done, they will pass it over and think about it again next time they encounter it. This matter has nothing to do with AI. It is an old problem in law firm management. But this release gives a very specific solution: make deviation capture an automated background task, and make manual updates a regularly triggered counter-signature process.
5. Four points of judgment: written to Chinese law firms and corporate legal affairs
Putting this release of Anthropic in the context of the current situation of China's legal industry, it is worth making a comparison. According to a 2025 survey by LexisNexis, the proportion of lawyers in mainland China who have used generative AI in their work has reached 88%, an increase of 20 percentage points from 2024. The same report also shows that 83% of Chinese lawyers tend to use AI tools that are “based on authoritative legal resources and are professionally trustworthy”, 80% are worried about the accuracy of AI content, and 77% are worried about data security.
This set of figures illustrates a very specific reality: lawyers are already using it, but they are not really reassured. This state of “use it but don’t trust it” is the real problem with China’s legal AI at the moment. Based on this, I have four judgments and suggestions.
Judgment 1: AI does not replace lawyers, but it forces lawyers to redefine what “judgment” is
There is no illusion of "AI replacing lawyers" in the entire link of this system. The whole process is to place AI in the position of "allowing lawyers to make better judgments faster." Every substantive decision is left to the sponsoring attorney. But on the other hand, when structuring, grouping, drafting, and case triage can all be completed by AI, the rationality of attorney fees can only be based on "judgment." The impact of this incident on Chinese lawyers has just begun. Situations in which clients use AI-generated preliminary contract evaluations to seek confirmation from lawyers have gradually emerged among large clients and technology-savvy legal counsel. Even though most domestic law firms have long ago not only billed by the hour, but mixed risk agency, target amount ratio, annual consultant and other methods, as long as the rationality of any part of it is based on "I can write documents better than AI", this part will become increasingly untenable. The real moat for lawyers in the future is the depth and speed of judgment, not the ability to produce documents.
Judgment 2: The top priority of the law firm is not to buy tools, but to document the “practical experience of the firm”
This is what I most want to say to my domestic colleagues. Chinese law firms are discussing a lot about "should we use AI tools" and "which domestic legal model is better?" These discussions are important, but the real bottleneck is not in the tools, but in the fact that you don’t have a document that allows AI to understand your team’s judgment criteria. Mark Pike's first piece of advice to all new users is "Don't use it out of the box." The Chinese version should be: spend two to three months first to write down the firm’s contract review manual, risk red lines, comparison list of who should be approached, and document style preferences into a document, and then discuss what kind of AI to use. How to do it specifically? It is recommended to start with the most frequent and standardized contract types in the firm, such as supplier agreements, confidentiality agreements, and equity transfer agreements. For each contract, first make a 5- to 10-page internal position manual, and then spread it to other business lines. If you do the former, it is extremely valuable even if you don’t use AI; if you don’t do the latter, even if you buy the most expensive AI, it can only produce a draft that “looks right but cannot be signed.”
Judgment 3: Competition in the industry sinks from the product level to the workflow level
The industry significance of this matter is greater than Claude for Legal itself. Anthropic has open sourced the "legal business workflow", which means that the competition in legal AI in the future will no longer be "whose product has more functions", but whose workflow is more granular and more closely aligned with the practice of the firm. For the major domestic legal AI platforms (iCourt AlphaGPT, Fagongchang, PKULaw, Alpha, etc., I personally have not conducted in-depth tests one by one, only made observations at the industry level), this is an opportunity and a pressure: the differences in templates and prompt words will disappear quickly, and the key is whether they can undertake the law firm's own work manual. For law firms themselves, this means that “starting to structure your workflow early” has a real first-mover advantage.
Judgment 4: The localization of Chinese legal AI is not a matter of translation, but a redesign of institutional differences.
The open source agreement allows anyone to commercialize, modify, and redistribute this system. However, the four shortcomings listed in the third section above illustrate that localization is by no means as simple as replacing English with Chinese. Underlying system differences such as attorney privileges, criminal prosecution paths, and forensic toolboxes, as well as data compliance and confidentiality red lines in China, each specific function must be redesigned by lawyers who understand Chinese law, and cannot be left to non-legal engineers. This means that a truly usable Chinese version of legal AI requires in-depth joint construction by legal professionals, technical teams, and compliance experts, and cannot be simply outsourced. This is actually an opportunity for Chinese law firms to enter the industry.
6. Written at the end: Jingwang’s attitude
As a Anhui Jingwang Law Firm director practicing in central and southern Anhui and serving listed companies and local enterprises for a long time, my attitude towards this matter is very direct.
First, we pay attention, but we don’t copy it. The methodological contribution of this release is far greater than the tool itself. The combination of "Team Profile + Decision Checklist + Credibility Mark + Self-Iteration Manual" has nothing to do with the legal system. It is an internal preparation that any law firm that wants to truly incorporate AI into its workflow needs to make.
Second, what we are doing is to structure the firm’s internal review standards, case handling experience, contract review manual, and decision-making rules for each business department. The long-term value of this matter is far greater than "buying a better AI tool." A "Jingwang Practice File" that can be shared, edited, and versioned by every lawyer in the firm is itself the most important asset of the law firm. Even if I retire or take over my job, this asset will still belong to this firm.
Third, a specific piece of advice for corporate legal friends: Before your company fully rolls out any legal AI, conduct a "cold start interview" in two to three weeks. Write your department’s compliance red lines, contractual positions, upgrade paths, and external law firm cooperation rules into an editable document. This is extremely worth doing even if you don't use any AI tools. It only makes sense to choose AI after you have done it. If this document is not available, any “AI legal system” will output generalizations.
Anthropic’s release this time is not so much a new set of tools for the legal industry, but it is the first time that the issue of “what the legal workflow in the AI era should look like” is put on the table for discussion. The position of Chinese lawyers should not be observers, but participants and leaders in this redesign.
Sources of information for this article: Anthropic’s official announcement, code repository anthropics/claude-for-legal, Artificial Lawyer’s exclusive interview with Mark Peck, Stanford Law School’s Artificial Intelligence Action Public Event, Fortune, Bloomberg Law, and Technology Blog’s concurrent reports, LexisNexis’ 2025 China Legal AI Application Survey, Gloria Mark’s research on attention interruption and recovery at the University of California, Irvine, and other public information. The views expressed in this article are those of the author alone and do not represent the position of the firm.