The Elephant in the Room at an Anti-Fraud Chapter Event

A day with financial crime investigators at an ACAMS chapter event in New Jersey.

By Sweat AI

We recently attended an ACAMS (Association of Certified Anti-Money Laundering Specialists) chapter event in New Jersey. We spent the day with people whose actual job is to investigate suspicious activity, and we've been thinking about it since.

There were analysts just starting their careers and people who had been doing this for decades. They wanted to write better suspicious activity reports, or SARs. They wanted to hear directly from regulators about what makes a report useful. People taking their work seriously and trying to get better at it.

One analyst had the wide-eyed enthusiasm you get at the beginning of a career. She wanted to catch criminals. There was a little of the investigator-with-a-badge-and-a-gun energy about her. You could see why she had chosen this work.

Then came the sponsor presentations.

We would go from a discussion about the care required to write a good report, including pointed warnings that AI-written reports would be recognizable, straight into a pitch for automating the work. Sometimes the irony was hard to miss: a lot fewer people would be needed to do it.

It was a strange room to sit in as an AI vendor.

An analyst reads a case file on a monitor while a colleague looks on.
Photo: FBI.

A talk about healthcare fraud brought the stakes back into focus. Fraud hurts people, including older and vulnerable people who may have very little ability to recover. AI gives the people targeting them more tools. The analysts trying to stop it deserve better tools too.

But we owe those analysts some honesty about what we are selling. It is reasonable to be skeptical when someone promises to help you and then measures the opportunity in how many of you can be removed.

Our answer at Sweat is to build copilots and keep experienced people involved. AI can gather records, trace connections, extract documents, and help prepare a case. A human needs the time, evidence, and authority to question the result. Putting an approval button at the end of a report nobody can realistically check does very little.

We mean this as a lasting commitment. We believe this work will continue to need judgment, including as the models improve.

We keep coming back to that young analyst. We want them to become very good at catching fraud. If we build this properly, they should have a much better chance.