The number, corrected by the person who quoted it
Start with the clarification, because it arrives in the same interview as the claim. 42AI says its behavioral world model, running at brokerages, predicts market trades with 98% accuracy. Asked to justify that claim, CEO Amit Gal-Or explained what the number counts: whether trading flows fall within a range the model predicted, not whether each trade was predicted correctly. Those are different measures. Predicting individual trades can be evaluated trade by trade. Measuring whether trading flows land inside a predicted range depends heavily on how wide that range is. A sufficiently broad range can capture more outcomes while offering less useful precision. 42AI has not published the range width, its coverage methodology or the test period, so the 98% figure cannot establish how precisely the model predicts trading behavior.
The second number Gal-Or offered measures something different. For predicting trading direction and volume, he said the model consistently achieves an F1 score of 0.8 to 0.85. Unlike the 98% range-coverage figure, F1 is a classification metric that balances precision and recall. But the company has not published the underlying dataset, test period, classification methodology or baseline, making the result impossible to independently evaluate or compare with competing systems. The company's website goes further: it promises to "know what every individual client will do before they act," up to 24 hours ahead. The interview's own fine print does not support that certainty.
What the model reads before it predicts
The method deserves a slower read. 42AI says it does not simply extrapolate a trader's past trades. It builds a representation of underlying traits from how a person behaves across unrelated parts of their life with a brokerage: the KYC process, onboarding, money transfers, and support calls, including the sentiment and sound of the person's voice and how often they call. Those observations become trait representations feeding a transformer-based model that assigns probabilities to decisions; a separate language model writes the explanation for each prediction.
Gal-Or says the system does not need personally identifiable information and that "the model never knows who the person is." That does not remove the underlying privacy question. The system still builds individual behavioral profiles from linked customer activity, including financial transactions and interactions with support staff. Removing a person's name from the model is not the same as eliminating the sensitivity of the behavioral profile. The interview does not explain what customers are told, what consent or authorization is obtained, or how brokerages govern the use of these predictions.
A $62.5 million seed funded by insiders and the founders
The $62.5 million funding figure needs its structure attached. Gal-Or described it as seed funding assembled over three years from the founders and Infinity Group, and said it was not entirely raised from external investors. Infinity Group was founded by Amir Gal-Or, Amit's father, who is also 42AI's chairman and a co-founder. The company has not disclosed how much came from each source. It now plans to raise additional capital, which could provide a clearer indication of external investor demand.
The commercial claims sit at the same evidence level. Gal-Or said 42AI is working with 10 large brokerages on liquidity, margin and hedging decisions. He described revenue as seven figures or more, without giving an exact amount. Gal-Or cited one brokerage that he said generated approximately $30 million in additional value through dynamic hedging. He did not provide the calculation methodology, comparison baseline or independent customer confirmation, so the figure should be treated as the company's reported estimate rather than a verified financial result. No brokerage is named, and the interview provides no customer-level evidence with which to verify those claims.
The question the founders were asked, and what answered it
TechCrunch put the obvious question to the founders: if a brokerage can anticipate what its own customers will do, what stops it from using that knowledge against them? Ezra called it a valid concern. The founders pointed to regulations, monitoring and an internal ethics committee, but the interview did not explain how those safeguards operate in practice. It did not establish whether brokerages are restricted from using individual predictions to trade against their own clients, how customers are informed, or whether an independent party audits the system's use. Those are important unanswered questions for a product that claims to forecast the behavior of a brokerage's own customers.
None of this makes the underlying idea empty. Behavior does carry stable tendencies, the founders are candid that the model degrades when it misses context such as a friend's advice, and a probability distribution over a client's next trades is a plausible risk tool even below the advertised precision. What launch day establishes is narrower: a family-backed team, a live deployment claim at 10 unnamed brokerages, and one headline number its own CEO had to redefine within the interview that introduced it. Judge the company by the F1 figure it volunteered second, and wait for a named customer or a published evaluation before judging it by anything else.
Sources
- [1] TechCrunch — "42AI comes out of stealth with $62.5M and a world model that can 'predict human behavior'," November 2, 2026 (launch interview with the two founders: behavioral world model description, trait representations and transformer architecture, separate language model for explanations, input types including KYC, onboarding, support-call sentiment and voice, transfers and trades, study of thousands and government/biobank datasets covering 2M+ people, customer datasets covering almost 30M, three years of development, Dan Shechtman's advice, the 98% claim and Gal-Or's clarification, F1 0.8–0.85 for direction and volume, 10 large brokerages, revenue above seven figures, one brokerage's claimed ~$30M in additional value, $62.5M from the founders and Infinity Group over three years with no allocation disclosed, the misuse question and the regulations/monitoring/ethics-committee answer; read first-hand)Read source
- [2] 42AI — official website ("the first Behavioral World Model"; "Know what every individual client will do before they act"; predicting each client's trades up to 24 hours ahead by instrument, direction and volume; response times of 1–10ms and 2–12ms; read first-hand)Read source