The question sounds like something lifted from a science fiction screenplay: could an artificial intelligence system anticipate a person's decision before they have even made it, including in situations they have never encountered before? An Israeli startup called 42AI is betting the answer is yes, and it has stepped out of stealth mode backed by $62.5 million in seed funding to make that case.

The company describes its core technology as a "behavioral world model" - a system designed to forecast how individual people will respond to circumstances before those responses occur. Its initial commercial deployment is focused on the financial sector, where the model is being used at a small number of brokerage firms to anticipate how traders will behave during periods of market volatility.

Probabilistic Behavior, Not Deterministic Prediction

Amit Gal-Or, the company's CEO and co-founder, is careful to frame what the model actually does. He does not claim it can read minds or predict individual choices with certainty. Instead, he argues that human beings carry persistent psychological tendencies that shape their decisions across changing circumstances, and that those tendencies can be measured, quantified, and fed into a machine learning system.

"Behavior is not deterministic; it's constantly probabilistic," Gal-Or has said when pressed on the scope of the company's claims. His argument is that while a person's situation may shift dramatically from one week to the next, their underlying behavioral patterns remain relatively stable - and it is those patterns that the model attempts to capture.

Co-founder and CTO Ehud Ezra expanded on this framing. "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," he explained. The model's goal, according to the two founders, is not to determine what someone will definitely do, but to assign probabilities to a range of possible decisions under a given set of conditions.

Internal Motivations Over Historical Patterns

What distinguishes 42AI's approach from conventional behavioral analytics, the company argues, is that it attempts to model a person's internal motivations rather than simply extrapolating from their past actions. To do this, the system draws on signals from a wide variety of interactions that may appear unrelated to the behavior being predicted.

In the context of a financial brokerage, those signals might include how a customer completes identity verification processes, how they engage during onboarding, the frequency and tone of their support calls, patterns in money transfers, and how they buy and sell financial instruments over time.

"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."

Ezra described the underlying philosophy this way: "We're trying to look at how they're looking at the world. In a world model, every individual is the world."

How the Model Works Technically

The observations gathered from those varied interactions are converted into mathematical representations of personality traits and behavioral tendencies. Those representations are then fed into a transformer-based world model, which generates probability distributions across possible decisions under different simulated conditions. A separate language model is used to produce human-readable explanations for the predictions the system generates.

According to Ezra, the model is specifically trained to maintain consistency in a person's underlying tendencies across simulations. If a person's behavioral profile indicates they are generally risk-averse, the model should continue to treat them that way even when testing them against hypothetical scenarios. If the model's simulation begins treating that person as a risk-taker, Ezra described that as a failure state - one that the model's training reward structure is designed to penalize.

Three Years of Development Across Multiple Disciplines

42AI spent three years building out this methodology before emerging publicly. The founders say the approach draws on research spanning psychology, neuroscience, and genetics, and that they consulted with specialists in psychology and post-traumatic stress disorder during development. One of their advisors is Dan Shechtman, the Israeli scientist awarded the Nobel Prize in Chemistry in 2011.

The team began by studying thousands of individuals through psychological assessments and structured behavioral scenarios. Over time, they incorporated government datasets and biobank data covering more than two million people. Gal-Or said that customer datasets accessed through the company's commercial deployments have since expanded that figure to nearly 30 million individuals.

In live deployments, the model works by analyzing existing customer records held by the brokerage, including transaction histories and recorded support calls. Gal-Or has argued that the system does not require personally identifiable information to function, stating that "the model never knows who the person is." However, the company does build individual behavioral profiles derived from linked customer activity, which raises questions about the practical distinction being drawn.

Performance Claims and Their Caveats

42AI has cited a 98% accuracy rate for predicting market trades, but Gal-Or was quick to clarify what that figure actually measures. The metric does not mean the system correctly predicted each individual trade. Rather, it refers to whether actual trading flows fell within a predicted probability range - a notably different and less stringent standard than trade-level accuracy.

The company also reports achieving an F1 score of between 0.8 and 0.85 when predicting trading direction and volume. The F1 score is a standard machine learning metric that evaluates how well a classification model balances accurate positive predictions against both false positives and false negatives. A score in that range is considered reasonably strong for a classification task, though context matters significantly in interpreting it.

The founders are also candid about the model's limitations. "There's a natural ceiling because you really don't have access to everything that a person experiences," Gal-Or acknowledged. Ezra pointed to a specific failure mode: when the model lacks important situational context - such as a trader who adjusts their behavior based on advice from a colleague - its accuracy degrades. Without visibility into that kind of external influence, the system cannot account for the resulting behavioral shift.

Commercial Traction and Revenue

42AI says it is currently working with 10 large brokerage firms, which are using the platform to manage liquidity, set margin requirements, and manage hedging exposure. The company is in the process of onboarding additional customers and says it is generating revenue in excess of seven figures annually, though Gal-Or declined to provide a more specific number.

Customers pay either a flat fee or a percentage of the measurable value the platform generates for them. Gal-Or cited one example in which a brokerage reportedly generated approximately $30 million in additional value through dynamic hedging strategies informed by the model's predictions.

Funding Structure and Investor Background

The $62.5 million in seed funding was raised over the course of three years and came from the founders themselves along with Infinity Group. Gal-Or noted that the total was not entirely sourced from external investors. Infinity Group was founded by Amir Gal-Or, Amit's father, who serves as both chairman and co-founder of 42AI. The company says it now intends to raise additional outside capital to support its next phase of growth.

Expansion Beyond Finance

While the current commercial focus is squarely on financial services, 42AI has outlined broader ambitions for its behavioral world model. The company is exploring applications in gaming, fraud prevention, and cybersecurity. Looking further ahead, the founders have expressed interest in using the technology to help robots and AI agents better understand and interact with human beings.

That longer-term vision comes with a significant practical caveat. Ezra acknowledged that the model cannot simply be transferred from one domain to another without substantial additional work. Each new application area requires the model to be trained on the specific rules, norms, and contextual signals relevant to that domain - a process that takes time and data.

Ethical Questions Around Behavioral Profiling

As with many AI systems that analyze human behavior at scale, 42AI's technology raises legitimate ethical questions that go beyond technical performance. If a financial brokerage can reliably anticipate what its customers are likely to do before they do it, that information represents a significant asymmetry of knowledge - one that could potentially be used to benefit the institution at the expense of the very customers being profiled.

When that concern was put to Ezra directly, he acknowledged it as "a valid concern." The founders pointed to regulatory frameworks, ongoing monitoring, and an internal ethics committee as safeguards. However, neither founder offered a detailed explanation of what specific mechanisms would prevent a customer of 42AI - a brokerage, for instance - from using behavioral predictions to gain an advantage over the retail traders whose data underpins those predictions.

Gal-Or expressed confidence that the technology itself could be part of the solution to such risks, rather than a source of them. "We are very proud that actually we are part of the solution; we are the solution to this problem," he said. Whether that confidence is well-founded will likely depend on how the technology is deployed in practice, what oversight structures are actually enforced, and how regulators in the financial sector choose to treat predictive behavioral profiling as the technology matures.