42AI comes out of stealth with $62.5M and a world model that can ‘predict human behavior’

Israeli startup 42AI has emerged with $62.5 million in seed funding and a pitch it calls a behavioral world model: a system that tries to estimate how a specific person is likely to decide, even in situations they haven't faced before. It is already running at brokerages, which the company says use it to manage liquidity, margin requirements and hedging.

That puts a profile of each customer's tendencies in the hands of the firm on the other side of their trades. The founders told TechCrunch the model doesn't need to know who a person is, but it still builds individual profiles from linked activity like support calls, transfers and trading history. That's the part listeners should sit with.

We'll give the technical idea its due. Modeling stable tendencies rather than just extrapolating last week's clicks is an interesting bet, and a brokerage that can anticipate a wave of panic selling could manage its own risk more safely. Firms that hedge better in a volatile market can be steadier counterparties. That's a real benefit.

But the claims need careful reading. The headline 98% accuracy, by the CEO's own clarification, measures whether trading flows land inside a predicted range, not whether individual trades were called correctly. The more meaningful number offered is an F1 score of 0.8 to 0.85 on trading direction and volume. The $30 million in added value at one brokerage is the company's claim, not an audited figure. And the seed round came from the founders and Infinity Group, founded by the CEO's father, who is also 42AI's chairman, over three years. That's a family-backed build, not a broad market vote of confidence.

The harder issue is incentives. A broker that can predict a customer's next move under stress has an obvious temptation to profit from it. When TechCrunch raised this, the co-founder called it a valid concern, and the company pointed to regulation, monitoring and an internal ethics committee without explaining what would actually stop a client from using the profiles against the people being profiled. We don't think 'trust our ethics committee' is an answer at this scale.

The raw material matters too. The model reads voice and sentiment from support and sales calls, and the company says its customer datasets now cover almost 30 million people, after early work that drew on government and biobank data covering more than 2 million. Most of those people never signed up to be modeled this way.

If this spreads into gaming, fraud prevention, cybersecurity and eventually agents and robots, as the company hopes, the pattern will be the same: institutions that know your tendencies better than you do, deployed by parties whose interests may not match yours. We want to see disclosure to the people modeled, clear limits on use, and regulators asking questions now, while it is still ten brokerages.

The company says it spent three years developing its approach, drawing on psychology, neuroscience and genetics research, with advice that included Nobel laureate Dan Shechtman. It converts observed traits into mathematical representations fed to a transformer-based model, and uses a separate language model to explain predictions. The founders acknowledge limits: accuracy drops when the model misses situational context, such as a trader acting on a friend's tip, and each new domain requires training on its own rules.