Will AI replace fraud analysts? | Frank McKenna & Marc Evans

Frank McKenna is Chief Fraud Strategist at Point Predictive and author of the Frank on Fraud blog. Marc Evans is a detective, Certified Fraud Examiner, and founder of Fraud Hero, with 14 years in law enforcement. Frank tracks fraud across trillions of dollars in originations. Marc shows up at the dealership. 

In this interview from Fraud Fight Club III, we explore their different vantage points about how AI will impact fraud detection, and why they still arrive at the same conclusion about where humans fit in.

The first-party fraud misclassification problem

The number I haven't stopped thinking about: Frank told me that up to 70% of early payment defaults in auto lending contain fraud in the application — usually falsified income or employment. And most of it never gets called fraud. It gets coded as credit risk and written off. "It's never called fraud," he said. "It's called early payment default." The industry is absorbing billions in losses that aren't showing up in any fraud report.

This matters because the defense against first-party fraud is fundamentally different from identity theft or synthetic identity fraud. By the time a lender recognizes the pattern, the vehicle is gone.

Frank's hot take: generative AI will expand the scope of what fraud teams can do, which will create more demand for analysts, not less. "There's a lot of fear that generative AI is going to eliminate the fraud analyst. I think it's going to do the exact opposite. What we're going to see is a lot more get hired."

Marc arrives at the same conclusion from a different direction. The bad guys get access to new tools before the good guys do, and they have no limits on how they use them. No single tool closes that gap permanently. What does: knowledge, education, and understanding how fraud actually happens at the human level. "Fraud is not just a tech problem. It's a human problem."

Neither of them is arguing against technology. They're both arguing that technology without human judgment leaves something critical on the table.

“A lot of fraud gets misclassified as default. Up to 70% of early payment defaults — there's fraud in the application, and it's never called fraud.”

Frank McKenna

Chief Strategist, Point Predictive

“Fraud is not just a tech problem. It's a human problem. If you can solve part of the human problem with knowledge and education, you can disrupt a lot of fraud.”

Marc Evans

Founder, Fraud Hero

Real-time intelligence sharing between lenders and law enforcement

The most functional version of lender-law enforcement coordination Frank and Marc described is not a platform. It's a phone number. Frank pointed to a program in Houston where a detective networked directly with car dealerships and would dispatch a squad car on a personal call if someone suspicious was in the showroom. Manual, but it worked.

Marc has been on the other side of that call. His department would show up at dealerships, sometimes before the buyer arrived, to verify documents, cross-check IDs against DMV records, and stop vehicles from leaving the lot. The recurring pattern: if someone is committing fraud at one dealership, they're likely doing it at several.

What blocks a more systematic version? Frank cited two things: data privacy constraints around sharing PII on someone who might not actually be committing fraud, and the liability exposure of a false positive resulting in an arrest. The informal networks work precisely because they're built on trust and direct verification, not automation.

“If they're doing it to one dealership, they're doing it to multiple dealerships.”

Marc Evans

Founder, Fraud Hero

Model explainability and what holds up in court

The explainability problem in fraud is partly technical, partly a communication challenge. Marc's framing: to explain a fraud determination in court to someone with no fraud background, you need to point to specific, concrete discrepancies. A font inconsistency. A dollar amount that doesn't reconcile across documents. A driver's license number with no DMV record. The more specific, the better.

Frank's diagnosis of why legacy fraud scores fall short: most are limited to three pre-programmed reason codes written to cover thousands of use cases in seven words or fewer. The result is language too generic to be useful in a legal or compliance context.

Generative AI changes this. A model that can articulate why a specific application looks risky, in plain language calibrated to the audience, produces a different kind of evidence than a three-character reason code ever could.

“With generative AI, I think explainability in fraud is going to change a lot. It was so limited before — reason codes that need to be seven words only and fit 10,000 use cases. You have to make them fairly generic.”

Frank McKenna

Chief Strategist, Point Predictive

Speaker bios

Frank McKenna is Chief Fraud Strategist at Point Predictive and author of the Frank on Fraud blog. He has spent his career building data-driven defenses against first-party fraud and application fraud in consumer lending.

Marc Evans is a Certified Fraud Examiner, active law enforcement detective, and founder of Fraud Hero. With 14 years in the field, he brings a street-level view of how fraud is actually committed and how knowledge transfer can disrupt it.

Frequently Asked Questions

What is first-party fraud in auto lending?

First-party fraud occurs when a real person uses their own identity to fraudulently obtain a loan they wouldn't otherwise qualify for — typically by falsifying income, employment, or other application details. Unlike third-party fraud (identity theft), the applicant is a real person presenting real ID. According to Frank McKenna at Point Predictive, up to 70% of early payment defaults in auto lending contain first-party fraud in the application, and most of it is never classified as fraud.

How does AI-enabled fraud differ from traditional fraud?

AI-enabled fraud uses generative AI tools to create convincing fake documents, voices, images, and videos at scale. The key difference from traditional fraud is speed and quality: forgeries that previously required skill and time can now be generated in minutes for near-zero cost.

What should fraud analysts expect from AI adoption?

Frank McKenna's argument is that generative AI will expand the scope of what fraud teams can do — more coverage, more complexity — which will increase demand for skilled analysts rather than reduce it. The human judgment layer doesn't go away; it shifts.

Why is it hard for lenders and law enforcement to share data in real time?

Two main barriers: data privacy laws restrict sharing PII on individuals who may not actually be committing fraud, and the liability risk of a false positive resulting in an arrest. The informal networks that do work — like the Houston program Frank described — are built on personal relationships and direct verification, which doesn't scale easily.

What makes fraud model outputs hard to explain in court?

Most legacy fraud scores are limited to a small number of pre-programmed reason codes written to cover thousands of scenarios in as few words as possible. The result is language too generic to be useful in a legal context. Frank McKenna's view is that generative AI will change this by enabling model outputs that explain specific findings in plain, contextual language.

Related Reading

• 2026 State of Document Fraud Report — Inscribe's annual data on AI-enabled fraud trends, including the 5x rise in AI-generated documents Frank referenced.
• From forgeries to deepfakes: Document fraud in the age of generative AI — How document fraud has crossed the threshold where manual review is no longer sufficient.
• AI Fraud Detection for Document & Credit Risk — How Inscribe approaches the detection problem Frank and Marc described.

Subscribe to Inside Inscribe for monthly updates on fraud, AI, and what I'm learning from the people fighting it.