FICO Executive: How We’re Teaching AI Human Judgment

THE BIG IDEA
FICO believes the next generation of enterprise AI won’t simply generate answers — it will think more like experienced professionals. By combining probabilistic AI that investigates with human-like curiosity and deterministic AI that enforces business rules, FICO aims to create AI agents that are both more capable and more trustworthy. The approach offers a blueprint for how enterprises can deploy increasingly autonomous AI without giving up governance over critical decisions.

For decades, the financial industry has worked to remove human judgment from routine decisions.

Algorithms replaced manual reviews. Predictive models replaced intuition. Credit scores and fraud systems became faster, more consistent and less susceptible to human bias.

Now one of the pioneers of that transformation believes the next leap in artificial intelligence depends on teaching machines something decidedly human: Curiosity.

“The interesting thing when someone asks me about AI, I say it is inherently human because we have to train the agent to be human, to be curious, to look at things,” said Rachael Hadaway, vice president of AI product management at FICO, in an interview with The AI Innovator.

That approach is shaping the next generation of FICO’s decision platform as the company expands beyond predictive AI into generative and agentic AI.

Instead of asking whether AI can analyze more data than people – a question largely settled years ago – FICO is tackling a different challenge: Can AI learn to think the way experienced professionals think?

Take a look at how FICO applied its approach of training AI to be more human to one of banking’s most difficult – and crucial – jobs.

When a suspicious transaction cannot immediately be explained, it normally lands with a fraud investigator.

The work isn’t simply looking at an established framework to determine if a purchase is fraudulent. Experienced investigators spend years developing an instinct for patterns that don’t appear in any rule book. They compare merchants, locations, timing of purchases, transaction histories and other factors, looking for subtle signs that might reveal a fraud ring.

Those decisions come from experience and gut feeling as much as explicit logic; human investigators are “curious” when they spot anomalies and look for relationships that aren’t immediately obvious, she said.

FICO’s newest AI agents are being trained to develop that same curiosity.

The company trains its AI models on both successful and unsuccessful fraud cases. The AI agents learn how experienced investigators reached conclusions, which clues mattered, which turned out to be false leads and why certain cases required additional scrutiny.

“We’re now training AI agents to look and act and be like a fraud investigator,” Hadaway said. “Instead of having this really long queue of cases they have to investigate, we’re allowing the fraud investigator agent to go do that, and then all the human has to do is review the case.”

The goal is to preserve decades of institutional expertise while allowing people to spend their time on the cases that still require human judgment.

Criminals continually adapt

Meanwhile, criminals are continuously adapting their techniques, forcing investigators to recognize patterns that didn’t exist a year earlier.

In one example, Hadaway said the AI detected unusual increases in fraud activity in certain countries before investigators had fully connected those patterns to changes in local regulations that fraudsters were exploiting. By recognizing emerging geographic trends early, banks were able to update fraud rules and investigations more quickly.

In another, the system uncovered new spending patterns designed specifically to evade traditional fraud detection.

For example, rather than making a single purchase large enough to trigger alerts – such as a transaction of more than $10,000 – criminals began making multiple smaller purchases in rapid succession.

“We were seeing people swipe their cards five, six, seven times,” Hadaway said. “They were getting really creative with their swipes and time patterns.”

Those discoveries become part of the next generation of training, allowing both investigators and AI to improve together.

However, FICO isn’t just teaching its AI to think like experienced fraud investigators. It’s also teaching it to anticipate how increasingly AI-enabled criminals might behave.

Hadaway said the company deliberately creates adversarial AI during training, effectively asking one AI system to outsmart another before those tactics appear in the real world.

Digital twins

FICO is applying its AI for other use cases as well, such as creating digital twins that let financial institutions simulate the impact of business decisions before actual implementation.

While banks have long used simulations to model business outcomes, generative AI dramatically expands the number of scenarios they can explore and the speed at which they can evaluate them.

Instead of changing lending policies and waiting months to measure the results, for example, banks can simulate thousands of possible outcomes inside virtual environments – and also get recommendations.

“You can now test thousands of simulations and get back not just ‘here’s what happened,’ but ‘here’s the best path forward,’” Hadaway said.

“We’re seeing a lot of banks wanting to do that so that they can test lots of different thresholds, different approaches to their portfolio without having to put A/B tests in market, wait a long time, and then maybe not even get back to some of those scenarios that those A/B tests represented,” she said. “So it’s a way to do A/B tests on steroids without impacting their customers.”

A lender considering different pricing for a particular customer segment can model how those changes might influence profitability, customer retention or default rates without exposing actual customers to experimentation.

“You’re learning from life consequences without impacting your customers,” Hadaway said.

Deterministic and probabilistic

For decades, FICO has depended on predictive AI. Machine learning models evaluated data and produced deterministic decisions: Was this transaction fraudulent? Should this loan be approved? How risky is this customer?

Those decisions remain at the core of the platform. What’s changed is what surrounds them.

Generative AI now explains why models reached particular conclusions, translating complex analytical output into language investigators and business users can more easily understand. The FICO AI assistant helps facilitate conversations in natural language.

Agentic AI performs much of the surrounding work — gathering information, investigating cases and handling the simpler cases while preparing the complex ones for human review.

This setup is critical to help avoid hallucinations in financial decisions as the AI agents are being trained to handle or route cases.

“It takes a really long time to train an agent to act like a human,” she said. “And it gets it wrong. It’s going to take working with a human hand-in-hand for months and months before it starts to pick up on those things.”

Importantly, FICO’s generative AI models never make the final financial decision.

“We’re still relying on our tried-and-true fraud models,” Hadaway said. “It’s all about enabling the decision versus we’re using something that can hallucinate to make the decision. The decision is still deterministic.”