Predictive analytics usually works at the level of groups. It tells you which segment is likely to churn or which product will sell next month. A startup called 42AI says it can go further and predict what one specific person will do, even in a situation they have never faced. The company came out of stealth on November 2, 2026 with $62.5 million in seed funding, as TechCrunch reported.

I coordinate AI and software projects at Debugged Pro, and clients ask us about prediction features all the time. So I read this story closely. Below are the verified facts, the parts that need a pinch of salt, and what agencies in India and the USA can actually do with this trend.

The co-founders of 42AI, the startup behind a behavioral world model for predictive analytics — 42AI's co-founders. Image credit: 42AI, via TechCrunch.

What 42AI Announced

42AI is based in Israel. It describes its product as a "behavioral world model." The TechCrunch story was written by Ram Iyer, who interviewed CEO and co-founder Amit Gal-Or and CTO and co-founder Ehud Ezra.

Here are the main facts from the report:

Funding: $62.5 million in seed money, raised over three years.
Who funded it: the founders and Infinity Group. Gal-Or said it was not entirely raised from outside investors.
Family link: Infinity was founded by Gal-Or's father, Amir Gal-Or, who is also 42AI's chairman and a co-founder.
Customers: 10 large brokerages, using it for liquidity, margin requirements and hedging exposure.
Revenue: more than "seven figures," according to Gal-Or, with no exact number shared.
Development time: about three years of work, with input from psychology experts and Nobel laureate Dan Shechtman.

The company plans to raise more money. It is also onboarding more customers, Gal-Or said.

How the Behavioral World Model Works

Most models learn patterns from what people did before. 42AI says it tries to measure why they do it. The idea is that people carry stable tendencies, even when their situation changes.

"Behavior is not deterministic; it's constantly probabilistic," Gal-Or told TechCrunch. Ezra put it another way: "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."

Based on the interview, the process looks like this:

Collect signals: the system looks at reactions to seemingly unrelated events. For a trader, that includes KYC checks, onboarding, support calls, money transfers and trades.
Read the context: during support or sales calls, it looks at sentiment, voice and how often a person calls.
Convert to traits: those observations become mathematical representations of personal traits.
Predict: a transformer-based world model lays out the probability of different decisions under different conditions.
Explain: a separate language model writes explanations for each prediction.

Training rewards are built to keep a person's core tendencies stable. If someone is usually risk-averse and the model starts treating them as a risk-taker, "the target is off," Ezra said.

42AI behavioral world model banner with the line predicting human and AI agent behavior — 42AI's own description of its product. Image credit: 42AI.

Where the training data came from

Gal-Or said the team first studied thousands of people through psychological assessments and behavioral scenarios. It then added government and biobank datasets covering more than 2 million people. Customer datasets now cover almost 30 million people, he said.

The company says the model does not need personally identifiable information. "The model never knows who the person is," Gal-Or said. TechCrunch noted, though, that the company still builds individual behavioral profiles from linked customer activity.

Bar chart comparing 2 million people in government and biobank data with about 30 million in 42AI customer datasets — People covered by 42AI's datasets, as reported to TechCrunch. Chart: Debugged Pro.

Reading the Predictive Analytics Claims Carefully

Big numbers grab attention, so it helps to read them slowly. 42AI claims 98% accuracy in predicting market trades. Gal-Or clarified that this measures whether trading flows fall inside a predicted range. It does not mean each trade was predicted correctly.

The more useful number is the F1 score. He said the model consistently scores between 0.8 and 0.85 for predicting trading direction and volume. F1 balances correct predictions against false alarms and missed events.

A few other points stood out to me:

There is a ceiling. "There's a natural ceiling because you really don't have access to everything that a person experiences," Gal-Or said.
Context gaps hurt accuracy. Ezra said the model struggles when someone acts on advice from a friend or colleague.
Every domain needs new training. The model has to learn the rules and context of each new field.
Value claims are unverified. Gal-Or said one brokerage gained roughly $30 million in extra value through dynamic hedging. No outside audit was cited.

None of this makes the work unimportant. It just means a client deck should quote ranges and conditions, not a headline accuracy figure.

The Ethics Question Agencies Cannot Skip

TechCrunch raised a fair worry. If a brokerage can predict what its customers will do, what stops it from using that against them? Ezra called it "a valid concern."

The founders pointed to regulation, monitoring and an internal ethics committee. They did not explain what would stop customers from using the tool to gain an edge over the people being profiled. Gal-Or still said, "we are the solution to this problem."

For agencies, this is the core lesson. A model that predicts individual behavior can protect people, for example by spotting fraud. The same model can also be used to nudge or exploit them. The design choices around it matter as much as the model itself.

What This Means for Predictive Analytics Projects in India and the USA

Most agency clients will not buy a world model next quarter. But 42AI shows where client expectations are heading. Businesses will want prediction at the level of each customer, not only each segment.

On its official website, 42AI lists uses beyond trading, including fraud protection, behavioral cyber security and personalized AI. TechCrunch reported that gaming is also on the list. Those are fields where many Indian and US agencies already build software.

Use cases worth exploring

Fintech and lending apps: flag unusual activity early and route it to a human reviewer.
Customer support: predict which tickets may escalate so a senior agent steps in sooner.
SaaS products: spot accounts at risk of churn and offer help, not pressure.
Ecommerce: plan stock and staffing around likely demand swings.

Where to slow down

Personal prediction runs straight into privacy law. In India, the Digital Personal Data Protection Act, 2023 sets rules on consent and purpose. In the USA, rules vary by state and sector. Removing names does not end the risk if you still build a profile of one person.

Google's AI story feature in India raised similar questions about personal data. We covered that in our guide on AI personalization and Google Dreambeans.

A Practical Checklist Before You Build

Here is the process I would follow on any client project that predicts individual behavior:

Define one decision. Write down the exact action the prediction will support, such as "send this ticket to a senior agent."
Map your data and consent. List every data source, why you collect it, and whether users agreed to that use.
Pick honest metrics. Use precision, recall or F1 on held-out data. Avoid a single accuracy number.
Test the edges. Check how the model behaves for new users, rare events and missing context.
Keep a human in the loop. Let predictions suggest, not decide, wherever money or access is at stake.
Explain each output. Store the main factors behind a prediction so staff can answer "why me?"
Write a misuse policy. Agree with the client on uses that are off limits, and log who sees predictions.
Review on a schedule. People and markets change, so retest the model every few months.

For risk reviews, I point teams to the NIST AI Risk Management Framework. It is voluntary and free, and its four functions (Govern, Map, Measure, Manage) fit neatly into a project plan.

If your client wants a model that keeps learning from fresh data, our post on AI model training and continual learning explains the tradeoffs. And if you need help scoping a prediction feature, our software development consulting team can review the plan with you.

My Take as a Project Coordinator

I like that 42AI's founders were open about limits. They admitted the ceiling, the context gaps and the need to retrain for each domain. That honesty is rarer than it should be.

Still, the funding story is unusual. Much of the seed money came from the founders and a family-linked firm, not from outside investors. That is not a red flag by itself, but it is worth knowing when you read the $62.5 million figure.

For agency teams, the takeaway is practical. Clients will soon ask for "AI that knows what each customer will do." Our job is to deliver useful predictions while protecting the people being predicted.

Key Takeaways

42AI raised $62.5 million in seed funding, much of it from its founders and a family-linked firm.
Its model predicts individual behavior from traits, not only past actions, and serves 10 large brokerages.
The 98% figure is a range measure. The reported F1 score for direction and volume is 0.8 to 0.85.
Privacy and misuse are real risks, even without names in the data.
Agencies should start narrow, use consented data, keep humans in the loop and follow a framework like NIST's.

About the Author

Nitin Kumar is the Lead Project Coordinator at Debugged Pro. He manages AI, web and mobile projects for clients in India and the USA, with a focus on scoping, delivery and responsible use of data. He writes about how AI news changes the way agencies plan and build software.