Wushi AI: Adopts a result-based charging model to resolve small-value disputes and is seeking angel round financing.
In China, tens of millions of civil disputes occur every year, yet less than 20 percent of them are brought to courts. With small subject amounts, long litigation cycles, and attorney fees often reaching thousands of yuan, a large number of parties involved in traffic accidents, unpaid labor wages, and rental disputes ultimately choose to let the issues go. These disputes are not worthless to resolve, but lack a layer of service infrastructure that can accommodate them at low cost.
Wushi AI (Chengdu Wandu Yijie Technology Co., Ltd.) is trying to build this layer of infrastructure through the collaborative delivery model of "AI Agent + Human Mentor". The project has recently launched its angel round of financing, planning to raise 3 million yuan in exchange for a 10% equity stake.
"Outcome Service Vacuum" for Small-amount Disputes
Tong Chen, founder of Wushi AI, defines this track as the "outcome service vacuum": litigation channels cannot accommodate the massive volume of small-amount disputes, law firms' model of charging based on the subject amount inherently disdains cases involving ten thousand yuan or less, general AI tools can generate legal documents but cannot deliver actual results, while the gray-area "rights protection" industry is harvesting this vacuum through means such as complaint bombing, threats and coercion — in July 2026, Wuhan police detained 37 people engaged in "fake rights protection, real extortion" in a single operation.
The market space is far from negligible. According to data from the Supreme People's Court, the national courts received more than 20 million first-instance civil and commercial cases in 2025, representing a year-on-year increase of 11%; for motor vehicle traffic accident disputes alone, the number of cases accepted in the first quarter of 2025 reached 330,000, up 59.21% year on year. Beyond cases accepted by courts, the volume of disputes that remain in pre-litigation mediation and private negotiation stages is even larger. Taking compulsory traffic insurance as an example, its annual compensation pool is about 226.2 billion yuan, but a large number of injured parties accept low quotations from insurance companies due to lack of professional assistance. It is worth noting that the core demand of this market has never been systematically met by any service layer before — what Wushi AI is facing is not a competition for share in an existing stock market, but building an incremental market from scratch.
Do not sell tools, deliver results — charge based on outcomes
As the first layer of capability of the infrastructure, the product logic of Wushi AI is "funnel-style delivery": users obtain case analysis, statutory compensation range calculation and solution deduction through free AI diagnosis; when the case is complex, certified professional mentors intervene one-on-one, responsible for negotiating boundaries with the opposing party and insurance companies until the compensation is credited to the account. The platform adopts the "charge by result, no charge if no result" model, where the compensation is directly transferred to the client's account, and the platform only collects the agreed service fee.
This pricing model directly changes the customer acquisition logic. Tong Chen himself has handled hundreds of disputes as a legal blogger, and his self-media account constitutes the initial customer acquisition channel of the project. Currently, the customer acquisition cost per order is about 10 yuan, while the reference price for industry paid traffic delivery is 200-500 yuan. Two months after launch, the project has completed more than 40 paid deliveries in total, with a month-on-month growth of 21% and a gross profit margin of about 65%.
Process data and "human as R&D": the growth path of a vertical agent
The real moat of the infrastructure lies in the data layer. Different from most legal AI products, Wushi AI places its moat bet on "process data". Tong Chen has an analogy for this: a written judgment is equivalent to the final line of a chess score that says "Black wins", while the negotiation process data is equivalent to every move of the chess game. If you only tell the AI "this game is won", it can learn nothing; only by telling it the situation and response of each step can it learn to play chess. Judicial documents have been made public for 20 years, yet no AI capable of negotiation has been trained, because what the training needs is the path rather than the end point.
In Tong Chen's view, process data has three layers of value. The first is unavailability: negotiations take place in claim settlement offices and chat windows, no regulation requires them to be made public, they cannot be crawled, purchased, or replaced by synthetic data — synthetic data can imitate statistical distribution, but cannot imitate the causal relationship in real games. The second is causal structure: outcome data only has correlation, while process data comes with the causal chain of "what action triggers what response at what node", which is exactly the whole of negotiation strategies. The third is compound interest attribute: legal provisions and judicial precedents will expire, but the game structure can be migrated across categories — most of the negotiation modes verified in traffic accident cases can be reused in work injury, unpaid wage and rental disputes, making process data a portable asset for category expansion. At present, the project has accumulated more than 500 real cases and over 15G of non-litigation game corpora.
Corresponding to this is a counter-intuitive view on cost: Wushi AI defines frontline human labor as R&D expenditure rather than service cost. The mechanism is that AI can currently independently solve only about 30% of conventional cases, and the remaining difficult cases are handled by humans. Every step of human handling of difficult cases — wording selection, evidence organization, negotiation rhythm — is structurally recorded and fed back to the training system, so that similar cases can be independently processed by AI next time. "Humans are not solving this single case, but realizing the permanent solution for this type of case." Tong Chen said, which means that while customers pay for the results, the project completes data collection and model training at the same time. Financially, this mechanism corresponds to a clear path: for every 10 percentage point increase in AI resolution, the proportion of cases handled by humans decreases correspondingly, and the gross profit margin rises from 65% to the 85% range for software products; the goal is to increase the independent AI resolution rate to 70% within 18-24 months.
This path has been verified overseas. EvenUp from the United States uses AI plus human support to enter the personal injury claims sector, with a valuation of more than 2 billion US dollars; Mercor has achieved a valuation of tens of billions of US dollars through the path of "human delivery precipitates process data, which in turn empowers AI"; while DoNotPay, which adopts a pure AI route, was punished by the FTC for false propaganda, which negatively confirms the necessity of "human backstop + compliance priority".
At present, the WeChat mini-program of Wushi AI has been launched on June 15, the APP is expected to be launched in late September, the algorithm filing has been completed, and the large model registration is under application. According to the plan, the funds raised in this round will be mainly invested in the replication of customer acquisition channel matrix, the improvement of AI resolution rate and category expansion — starting from traffic accident claim settlement, extending to work injury, unpaid wage and rental disputes. Outside the litigation doors of courts, this dispute market with a volume of tens of millions of cases per year may finally welcome its first layer of infrastructure.