42AI emerged from stealth on November 2, 2026, saying 10 large brokerages already use its AI system to manage liquidity, margin requirements and hedging exposure. The Israeli startup calls the technology a behavioral world model: software designed to estimate how individual people will respond under different circumstances.
The consequential detail is its access to linked customer activity. According to TechCrunch’s launch report, 42AI analyzes transactions, support calls and other interactions to build representations of individual decision tendencies. Brokerages can then use those predictions in financial decisions.
Its headline performance claim needs careful reading. The company’s reported 98% accuracy describes whether trading flows fell within a predicted range, not whether it correctly anticipated 98% of individual trades. That distinction separates a potentially useful forecasting tool from the much stronger suggestion that software can reliably read a trader’s next move.
42AI also disclosed $62.5 million in funding accumulated over three years from its founders and Infinity Group. That was development financing, not entirely a newly raised external seed round. The deployed profiling system deserves more attention than the funding label.
How the Behavioral World Model Works
CEO Amit Gal-Or and CTO Ehud Ezra told TechCrunch that 42AI tries to infer underlying psychological tendencies from behavior across different situations. For a brokerage customer, those observations can include:
- Know-your-customer checks and onboarding activity.
- Support and sales calls, including sentiment, voice and contact frequency.
- Money transfers.
- Buying and selling activity, interpreted alongside market conditions.
The company converts observations into mathematical representations of traits and feeds them into a transformer-based model. Its output assigns probabilities to possible decisions under different circumstances.
Ezra described training that rewards consistency with inferred tendencies. If the system represents someone as generally risk-averse, a simulation that abruptly treats that person as a risk-taker would miss the training target.
That is a useful description of the intended behavior, but it leaves a scientific question open: how accurately do those inferred traits capture a person’s decision-making? Encouraging consistency can make simulations coherent. A person can also change strategy, learn something new or behave differently under pressure. Coherence and predictive validity need separate evaluation.
The LLM Explains the Prediction
42AI uses a separate language model to generate explanations for its predictions, according to TechCrunch. The behavioral model produces the forecast; the LLM supplies the written explanation.
A fluent explanation can make a score easier to understand, but readers still need evidence that it faithfully describes the forecast’s basis. The published account does not describe that validation.
The company also acknowledges missing context. A customer might change a trading decision after advice from a friend or colleague that the model never observes. Ezra said expansion into another domain requires training on that domain’s rules and context. The brokerage deployment therefore offers limited evidence for broader ambitions in gaming, cybersecurity, robotics or AI agents.
What the Accuracy Figures Actually Measure
42AI’s two reported performance figures answer different questions. Neither should be read as a blanket probability that an individual prediction will be correct.
| Company-reported metric | What the report says it measures | What readers still need |
|---|---|---|
| 98% accuracy | Whether trading flows fall within a predicted range | Range width, forecast horizon, aggregation level and comparison with simpler forecasts |
| F1 score of 0.8–0.85 | Performance predicting trading direction and volume | Class definitions, thresholds, averaging method and evaluation conditions |
These figures, the customer count and the reported financial results are company claims relayed by TechCrunch. The cited coverage provides no independent audit or technical evaluation, and insufficient methodological detail to compare performance with competing systems.
Range Coverage Depends on Range Width
A forecast interval specifies a range of possible outcomes. Coverage measures how often the observed result lands inside it.
A wide range is generally easier to hit than a narrow one. A forecast can achieve high coverage while offering little guidance if its interval includes most plausible outcomes. A useful evaluation therefore needs coverage and interval width together, measured over the same forecast horizon.
For a brokerage, timing also changes the practical value. A forecast covering the next few minutes serves a different purpose from one covering a trading day. The available report does not specify the horizon behind the 98% figure.
Aggregate forecasts can still be valuable even when individual predictions are unreliable. Errors involving different customers may offset one another, producing a reasonable total-flow estimate. That would support a liquidity-planning use case without demonstrating equally strong account-level prediction.
My view is that interval width, forecast horizon and a comparable historical-flow baseline would be more informative than the headline percentage alone. Those details would show whether 42AI narrows uncertainty enough to improve an actual decision.
An F1 Score Is Not a Percentage of Correct Trades
F1 is the harmonic mean of precision and recall. Precision measures how often a model’s positive predictions are correct; recall measures how many actual positive cases it catches.
An F1 score of 0.8–0.85 can indicate useful classification performance, but its meaning depends on the task. Predicting trading direction requires defined categories. Applying F1 to volume requires a classification or thresholding setup, which the report does not explain.
A reader cannot safely translate the score into “80–85% of trades predicted correctly.” The missing class definitions and averaging method also make it difficult to judge whether strong results in common situations conceal weaker performance in rare, costly ones.
For brokerage risk management, losses from mistakes matter alongside prediction scores. A model could perform well across ordinary trading periods yet miss the unusual behavior that creates the largest exposure.
Linked Profiles Remain a Privacy Issue
Gal-Or told TechCrunch that 42AI does not need personally identifiable information, saying “the model never knows who the person is.”
That claim concerns identifying information available to the model. The described system still links interactions and transactions into individual behavioral profiles.
A brokerage can potentially act on an account-level prediction without the prediction engine knowing the account holder’s legal name. If a profile remains connected to an account through an identifier, it can influence decisions affecting that customer. Removing direct identifiers therefore leaves questions about how the profile is used.
This is the central privacy issue in the reported design. Re-identification is one concern, but even a successfully pseudonymized profile can support consequential treatment of the person behind it.
According to TechCrunch, Gal-Or said development began with psychological assessments and behavioral scenarios involving thousands of people. The company subsequently added government and biobank datasets covering more than two million people. He said customer datasets had expanded to cover almost 30 million people.
Those figures describe the reported scale of the data. They do not show that nearly 30 million people are continuously monitored, or establish what each person understood about the processing.
The practical safeguards remain unclear: which party holds the account-to-profile link, how long profiles persist, whether raw recordings remain available, and how customers can challenge an inference. Those details determine whether a claim about identity protection translates into meaningful limits on profiling.
Frequently Asked Questions
3 questions
1Does 42AI Predict Individual Trades With 98% Accuracy?
No. The company says the figure measures whether trading flows fall within predicted ranges. It is not a reported per-trade success rate.
