42AI raises $62.5M to predict human behavior before decisions are made
An Israeli startup claims its AI can predict what you will do next by mapping your psychology, but the 98 percent accuracy claim has a catch.
Models · Source: TechCrunch
What happened
Israeli startup 42AI just came out of a three-year stealth period with $62.5 million in seed funding. The company built a behavioral world model designed to predict human decisions. Instead of just copying past actions, the model maps underlying psychological tendencies to calculate the probability of what a person will do next. The funding came from the founders and Infinity Group, a firm founded by the CEO's father.
The system is already live at ten large financial brokerages. It monitors how individual traders react to volatile market conditions. The model ingests data from support calls, KYC processes, and trading history to build mathematical representations of user traits. It then uses a transformer-based architecture to predict trading direction and volume with response times as low as one to ten milliseconds.
The company claims 98 percent accuracy, but there is a catch. CEO Amit Gal-Or clarified this number only means trading flows fall within a predicted range, not that the AI predicts every exact trade. The model actually scores between 0.8 and 0.85 on the F1 metric for predicting trading direction and volume. One brokerage reportedly generated $30 million in additional value using the tool for dynamic hedging.
Key facts
- $62.5 million — Seed funding raised by 42AI over three years
- 30 million — Number of people covered in their current customer datasets
- 0.8 to 0.85 — F1 score for predicting trading direction and volume
- 1 to 10 milliseconds — Response time for the world model
- $30 million — Additional value generated by one brokerage customer
Why it matters
We are moving from AI that predicts text to AI that predicts human psychology. Builders usually train models on what users clicked or bought in the past. 42AI is trying to model the internal motivations driving those clicks. If this works, applications in finance, gaming, and fraud prevention will shift from reacting to user behavior to preempting it.
This creates a massive privacy and ethics gray area. The company says the model does not need personally identifiable information. However, it still builds highly specific behavioral profiles from linked customer activity. If brokerages know what their clients will do before the clients do, the risk of exploitation is high. The founders pointed to regulations and an ethics committee, but offered no hard technical limits to stop customers from weaponizing these predictions.
For builders
Sell outcomes instead of software subscriptions
42AI charges customers a flat fee or a percentage of the financial value the platform creates. If your AI tool directly increases revenue, tie your pricing to that upside. Brokerages are willing to pay a cut if you can prove you made them millions.
Build models on unconventional data signals
The startup does not just look at trading data. They analyze voice sentiment on support calls and behavior during onboarding to predict financial risk tolerance. Look for adjacent, seemingly unrelated user actions to feed your predictive models.
Define your accuracy metrics clearly
A 98 percent accuracy claim gets headlines, but enterprise buyers will dig into the details. 42AI admits their real metric for trading direction is an F1 score of 0.8 to 0.85. Be ready to defend your math when selling to technical teams.
My take
Predicting human behavior is not mind reading. It is just pattern matching at a massive scale. I respect the ambition to model psychology, but letting financial brokers front-run human tendencies based on shadow profiles is a regulatory nightmare waiting to happen. You cannot claim privacy just because you stripped out a name while keeping the exact blueprint of a person's brain.
Original reporting: TechCrunch. This is my rewrite and opinion.