42AI exits stealth with $62.5M for a world model that predicts how individual traders will behave

What 42AI says it built

The pitch sounds like a rejected science fiction premise: a model that estimates what a specific person will decide before they decide it, including in situations it has never seen them face. 42AI says that is what it has, and it came out of stealth on Monday with $62.5 million to sell it, TechCrunch reported.

"Behavior is not deterministic; it's constantly probabilistic," CEO and co-founder Amit Gal-Or said. His argument is that people carry underlying psychological tendencies that persist even when their circumstances change, and that those tendencies can be measured, encoded and turned into a distribution over likely actions.

"A week from now, you might be a different person, but your tendencies will always play out in certain areas to a probability or distribution," co-founder and CTO Ehud Ezra added. The model's job is to output the probability of a given person making a given decision under a given set of conditions.

The commercial version of this runs in financial brokerages. For a trader or investor, the inputs include seemingly unrelated interactions: how they completed KYC checks, how they behaved during onboarding, whether they called support and how often, how they spoke on those calls, when they moved money, and how they bought and sold as the market moved around them.

How the model works under the hood

"When the person is talking to support or sales in the initial process, we're looking at the sentiment, we're looking at the voice, we're looking at a lot of interactions that they're doing, how many times do they call, and so forth," Ezra said. "And all of those data points are going into the model. Later on, you have the interactions: how you buy and sell, and when, and we're looking at the market at that stage."

Those observations are converted into mathematical representations of traits and fed into a transformer-based world model that lays out probabilities for decisions under different scenarios. A separate language model generates plain-language explanations for the predictions, so a human can read why the system expects a particular behaviour.

"So we're trying to look at how they're looking at the world. In a world model, every individual is the world," Ezra said. The training objective is unusual: the model is rewarded for preserving a person's underlying tendencies across simulated scenarios. If someone is consistently risk-averse and the simulation starts treating them as a risk-taker, the target is wrong and the model is penalised.

The data story behind that is the part worth scrutinising. 42AI says it spent three years on the approach, combining psychology, neuroscience and genetics research with input from psychology and PTSD specialists and from Dan Shechtman, who shared the 2011 Nobel Prize in Chemistry. It says it first studied thousands of people through psychological assessments and behavioural scenarios, then added government and biobank datasets covering more than 2 million people. Customer datasets have since expanded that to almost 30 million people.

The accuracy numbers, and what they do not mean

42AI claims 98% accuracy in predicting market trades, a figure that collapses on inspection in a way the CEO is at least honest about. "It's more a measure of whether trading flows fall within a predicted range, not if each trade was predicted correctly," Gal-Or said. A range forecast that is right most of the time is useful for risk management and is not the same claim as knowing what a person will do next.

The company also says it consistently reaches an F1 score of 0.8 to 0.85 for predicting trading direction and volume. F1 balances precision against recall, so that number is a more meaningful statement about classification performance than the 98% headline, though it is still measured on the company's own data, reported by the company, without an independent audit.

The founders are upfront that the system has a ceiling. "There's a natural ceiling because you really don't have access to everything that a person experiences," Gal-Or said. Ezra added that accuracy degrades when the model misses situational context, for example when someone changes their trading behaviour because a friend or colleague gave them advice the system never saw.

That is the honest shape of behavioural prediction in 2026: good enough to be commercially valuable on aggregate flows, not good enough to be treated as a read of an individual's mind. The gap between those two things is where the marketing risk lives.

Who is paying, and for what

42AI says it is working with 10 large brokerages, using the model to manage liquidity, margin requirements and hedging exposure. It is onboarding more customers and generating more than "seven figures" in revenue, according to Gal-Or, who declined to give a specific number.

Pricing is either a flat fee or a percentage of the value the platform helps a customer capture. Gal-Or claimed one brokerage generated roughly $30 million in additional value through dynamic hedging — a number that, if true, explains why a seed-stage company can charge meaningful money without a consumer product.

The funding itself is unusual. The $62.5 million came from the founders and Infinity Group over three years, and was not entirely raised from external investors, Gal-Or said. Infinity Group was founded by his father, Amir Gal-Or, who is also 42AI's chairman and a co-founder. The company now plans to raise more money.

That structure explains how a three-year research programme survived without a priced round, and it also means the $62.5 million headline describes capital committed over time by insiders and a related fund rather than a competitive seed round with outside lead investors setting a valuation.

The conflict nobody has solved

If a brokerage can anticipate how its own customers will behave under stress, the obvious follow-up question is what stops that brokerage from trading against them. Ezra conceded the point was "a valid concern". The founders' answer was regulations, monitoring and an internal ethics committee, and they did not explain what, if anything, would stop a customer from using the technology to gain an advantage over the very people it profiles.

Gal-Or was more confident: "We are very proud that actually we are part of the solution; we are the solution to this problem." That is a positioning statement, not a control. An ethics committee reviews intent; it does not change the fact that the model's output is a per-person probability distribution held by the counterparty in a trade.

The data-protection angle is similarly thin. Gal-Or said the system does not need personally identifiable information and argued that "the model never knows who the person is". But the company also describes building individual behavioural profiles from linked customer activity, which is a profile of a person whether or not a name field is attached. Under GDPR-style regimes, a pseudonymous behavioural profile of an identifiable customer is still personal data, and the relevant question is what legal basis the brokerage claims for building it.

Where the company wants to go next

Beyond finance, 42AI says it is exploring gaming, fraud prevention and cybersecurity, and eventually wants to help robots and AI agents understand the people they work with. Ezra acknowledged that each new domain requires the model to be trained on that domain's rules and context, which makes expansion slower than the pitch implies.

The cybersecurity and fraud cases are the most plausible near-term second market, because both already run on behavioural signals and both have an established appetite for probabilistic risk scoring. The robotics case is the furthest out: predicting a person's next decision in a well-instrumented brokerage account is a different problem from predicting what a stranger will do in a physical room.

What 42AI has actually proven so far is narrower and more interesting than the headline claim. It has built a system that turns messy interaction data into usable per-person distributions, sold it to institutions that already own that data, and produced revenue without a consumer app, a foundation model, or a public demo. Whether that becomes behavioural infrastructure or a well-funded profiling engine with an ethics committee is now a question about its customers, not its code.