Anticipating what a person will choose to do before they have even made up their mind remains one of the most elusive frontiers of artificial intelligence. Tel Aviv based startup 42AI believes it has built a framework to crack that puzzle, exiting a three-year stealth phase backed by $62.5 million in seed capital. Rather than analyzing human actions in a vacuum, the company is deploying what it terms a behavioral world model, currently operating inside ten financial brokerages to forecast how individual traders react to volatile market swings.
Human Decisions as Probabilities Rather Than Fixed Outlines
Human decision-making rarely follows rigid, deterministic paths. As co-founder and Chief Executive Officer Amit Gal-Or explains, personal choices exist across a spectrum of probabilities. Individuals carry underlying psychological traits that guide their actions even when the surrounding environment completely transforms. The startup aims to quantify these persistent internal tendencies into structured mathematical parameters that algorithms can parse. Co-founder and Chief Technology Officer Ehud Ezra notes that while life circumstances may alter an individual over time, core behavioral distributions remain remarkably stable across comparable situations, allowing the system to project the likelihood of specific actions rather than claiming absolute certainty.
Mapping Deep Motivations Over Surface Past Actions
Traditional predictive models frequently rely on extrapolating patterns directly from an individual's historical actions. In contrast, 42AI focuses on identifying latent motivations by evaluating how people respond to seemingly disconnected touchpoints. In a trading context, these indicators span everyday customer journeys, including onboarding timelines, completion of KYC procedures, support interactions, deposit mechanisms, and trade execution styles.
During early interactions with customer service or sales teams, the platform monitors conversational tone, voice cadences, call frequency, and general sentiment. These disparate signals are synthesized into the model alongside actual market participation, evaluating when and how assets are bought or sold against prevailing market volatility. Ezra describes the methodology as treating each individual as their own self-contained world within the broader architecture, decoding how each user perceives risk and opportunity under pressure.
Transformer Foundations and Three Years of Interdisciplinary Research
Once captured, these observations are translated into mathematical representations of human traits and routed through a transformer-based world model to simulate decision probabilities under varied scenarios. A complementary language model generates natural language explanations detailing why specific outcomes were forecasted. To maintain fidelity, the training framework penalizes deviations from baseline temperaments; if an inherently risk-averse profile is simulated as taking reckless positions, the training loss indicates a structural error.
Building the platform required three years of research bridging psychology, neuroscience, and genetics. The founders collaborated with specialists in post-traumatic stress disorder and cognitive science, alongside Dan Shechtman, recipient of the 2011 Nobel Prize in Chemistry. The developmental runway began with psychometric evaluations and scenario testing across thousands of participants, later integrating government and biobank datasets representing more than 2 million individuals. Customer datasets have since scaled to encompass nearly 30 million people.
Deployment Metrics and Inherent Operational Limits
In commercial environments, the algorithm examines existing enterprise transaction histories and call logs without collecting direct personally identifiable information, according to Gal-Or, though the system maintains linked behavioral profiles across user activities. In market forecasting, 42AI points to a 98% accuracy benchmark, though Gal-Or clarifies this reflects trading volumes staying within modeled boundary ranges rather than predicting every single order. The model consistently logs an F1 score between 0.8 and 0.85 for direction and volume classification, balancing true positives against false positives and missed events.
The system faces clear boundaries. Gal-Or acknowledges a natural ceiling given that an algorithm cannot observe an individual's total lived reality. Sudden external influences, such as a trader acting on private advice from a colleague or personal confidant, create blind spots that degrade predictive performance.
Commercial Traction, Capital Structure, and Ethical Questions
The company currently serves 10 major brokerage houses, using predictive profiling to optimize margin rules, liquidity planning, and hedging exposures. Generating seven-figure annual revenue through recurring software fees or value-share arrangements, Gal-Or reported that dynamic hedging powered by their software helped one client secure an estimated $30 million in incremental value. The $62.5 million seed round was assembled over three years through the founders and Infinity Group, an investment vehicle established by Amir Gal-Or, who also serves as 42AI's chairman and co-founder. The firm is preparing to solicit fresh investment capital.
Looking ahead, 42AI envisions deploying its architecture across gaming ecosystems, cybersecurity defense, fraud detection, and autonomous robotics. However, Ezra concedes that entering each new domain demands extensive retraining on sector-specific rules and contextual constraints. The platform also brings notable ethical questions: brokerages armed with foresight into client reactions could theoretically leverage information asymmetries to their own financial benefit. While the leadership points to internal ethics committees, monitoring protocols, and external regulations, structural barriers preventing client exploitation remain an ongoing industry challenge that Gal-Or maintains their technology can ultimately help solve.
