
A $2M Anti-Fraud Agent: This One Point Decides If It's Worth It
I compared Dudao Tech's "AI + anti-fraud Agent" suite against the free "National Anti-Fraud AI" you can download on your phone — one sits in the backend, one sits in your hand. I poked at both. Whether it's worth it depends on the scenario; the two have very distinct use cases.
For provincial/municipal police anti-fraud centers, telecom operators' anti-fraud task forces, bank risk-control middle platforms — and those that already have a human persuasion agent team — this thing is worth considering. For small and mid-sized organizations, or those just wanting to do publicity and legal education, not worth it; the free phone one is enough.
What it actually is needs to be made clear. This is not an app installed on your phone. It's a set of agents installed in the backend of police, telecom operators, and banks: case-related numbers, call records, transfer behavior go in as input, and it does analysis, grading, generates persuasion scripts, and dispatches work orders to persuasion staff — the whole process in one pass. I hadn't encountered the Dudao Tech brand before; per the sales rep, it's already deployed in Jiangsu and Zhejiang, at 1.5–2 million RMB per customer.
How I actually got hands on it. An old security-industry client in Jiangsu deployed a similar system last year; I went on-site and watched for half a day. This time the sales rep opened a demo environment for me, and I ran through a few dozen scenarios I wrote myself — impersonating police/prosecutors, brush-order rebate scams, fake investments, impersonating customer service for refunds, all covered.
The interface has three main parts: lead pool, analysis, persuasion work orders. I threw in a chunk of dialogue text impersonating police/prosecutors; the pattern classification was correct, and it even pulled up similar cases. That part was better than I expected. The persuasion script generation is also done carefully — it tailors wording by group: one set for the elderly, one for students, one for corporate finance staff, so agents don't have to improvise. That's exactly what eats up human agents the most — hundreds of calls a day, all scripts relying on personal experience. This part genuinely saves manpower.
Where it gets stuck is data integration, not the model. The demo environment's data is clean; in a real project, call record formats, timestamps, number attribution — a pile of things to clean and align. The risk levels it outputs follow its own standard, while my work order system uses a different grading scheme, so a mapping layer has to be wrapped in between. I've been testing the work order system these past few days, and the more I test the more I feel this kind of integration work is the most time-consuming — nothing to do with how accurate the model is.
There's also the old problem of cross-department data access. Half the anti-fraud data is with telecom operators, half with police — that's not something technology can solve. Last month I wrote a line: whether a model is good isn't about the leaderboard, it's about whether it can be stuffed into existing systems without rework. If you want to stuff this thing in directly, rework is unavoidable.
Cost also needs to be calculated clearly. The 1.5–2 million is the product price, excluding data integration, agent training, and later model tuning. Actual first-year investment will add another chunk on top — how much depends on how bad the existing systems are; I might be overestimating.
Compare with the phone one. The National Anti-Fraud AI launched in September this year, guided by the Ministry of Public Security's Criminal Investigation Bureau and developed by Shanghai police themselves; input a suspicious scenario and it analyzes, breaks down the scam pattern, and recommends cases, plus there's a senior version and an international version. Media have tested the National Anti-Fraud Center APP's AI content detection, saying uploaded AI-generated images are basically all detected. This kind of thing faces ordinary people, is free, zero barrier to entry — but it passively waits for you to ask. The backend suite actively goes out to find people; the logic is different, they can't replace each other.
So who's it for. Those with human persuasion agents, local data access rights, and annual case-related losses they can keep under control — worth talking. Those just wanting to do publicity, without an agent team — don't spend this money yet.
Physix Frontier