The Rise of Evaluative AI in Banking

Nearly every conversation about artificial intelligence in banking eventually turns to productivity. Institutions want to know how many hours can be saved, which tasks can be automated and how quickly work can move. That focus makes sense, but it misses a more urgent and increasingly a board-level question: Can a bank trust what it sees?

Javelin estimated identity fraud losses reached $27.3 billion in 2025, driven at least in part by fraudsters increasingly using AI to create more convincing scams and synthetic identities. Financial institutions are under growing pressure to determine who and what can be trusted before money moves or an account is opened.

According to Bank Director’s 2026 Risk Survey, 79% of CEOs, directors and senior executives identified fraud as a top risk for their institution, while 84% cited fraud and scams targeting customers as their primary AI-related concern. To address this risk, more banks are prioritizing AI built to make better decisions rather than simply faster work.

The Fraud Problem, Intensified by AI
Every day, an institution makes countless decisions about whom to trust, whether to open a new account, approve a payment or transfer, or place a hold on suspicious activity. Those decisions are becoming harder to get right. Fraudsters now use AI to create more convincing scams, synthetic identities and activity that can closely resemble legitimate customer behavior.

The Federal Bureau of Investigation’s Internet Crime Complaint Center reported that AI-related cybercrime appeared in its annual report for the first time in 2025, logging more than 22,000 complaints and nearly $900 million in associated losses. Fraud remains a persistent concern, now intensified by AI.

The problem is not catching red flags but seeing how they connect. A login from a new phone looks routine, so does a customer updating their address or a payment to someone they have never paid before. Each of those facts, what fraud teams call signals, looks ordinary alone. Seen together, and against the customer’s own history, they tell a different story. The insight rarely lives in any one fact, but in the relationships between them.

This helps explain why more data has not automatically produced better outcomes. Fraud prevention depends on understanding which signals matter, and what they reveal when viewed together.

AI as Both The Problem and The Solution
Much of the industry’s attention has focused on solving these concerns with generative AI and, more recently, agentic AI. But a third category — evaluative AI — addresses a different problem: assessing risk and trust in an era of increasingly sophisticated, AI-driven fraud.

While a generative model predicts the next word, an evaluative one anticipates the next attack and is designed to help organizations answer a fundamental question: Can this activity be trusted? By analyzing relationships across multiple signals rather than viewing them independently, evaluative AI can help organizations develop a more complete understanding of identities, transactions and customer behavior.

While many solutions analyze individual events, their effectiveness depends largely on the quality, depth and historical context of the data behind them. Understanding how signals connect over time requires a broader, more contextual view.

As institutions evaluate AI initiatives, they should consider whether the technology can provide that context. The ability to draw insights from years of historical information may reveal patterns that would otherwise go unnoticed, helping institutions make decisions about trust and potential risk with greater confidence.

Fraud prevention and customer experience are often treated as competing priorities. Reducing fraud is important, but so is ensuring that trusted customers can move through an institution’s channels with minimal friction. The strongest outcomes often come from improving both simultaneously rather than optimizing for one at the expense of the other.

Closing the gap between the pace of AI and the board’s ability to oversee it does not require directors to become data scientists, just that they ask sharper questions of management:

• What decision are we trying to improve?
• What does getting it wrong cost us, in both fraud and customer losses?
• Can the model explain its reasoning to our risk committee and our regulators?

Institutions that understand the problem they are solving are better positioned to tell genuine capability from an AI label.

The challenge is understanding which signals matter and what they reveal when viewed together. That’s where evaluative AI comes in, helping institutions more effectively assess relationships, context and intent. The next phase of AI adoption may be defined less by automation and more by an institution’s ability to make better decisions about risk, trust and customer relationships.