Financial crime volumes keep climbing, investigation teams don't

We sat down with Oskar Dahl Aldeborn, AI Lead at Redeploy, who spends his weeks inside banks' financial crime teams, to hear where the time actually goes and what he thinks should change.

His work covers three areas that banks usually run as separate functions: know your customer (KYC), the checks a bank runs when it takes on a customer and reviews them later; anti-money laundering (AML), the work of finding and reporting laundered money; and transaction monitoring (TM), the alerts on unusual transactions that feed the AML investigators.

“Different teams and different tools, but the same shape of case every time,” Oskar says. “A trigger, evidence to assemble, a decision, and a record the regulator can follow.”

Most of a case is administration

We asked Oskar where the hours in a case go, and his answer begins with a screen.

“In the last process walkthrough I sat in, the analyst had the client lifecycle management system, the CLM, open and ten other windows. The CLM is where the case is built and signed off. None of the information the case needs lives there.”

It sits in the other ten, and much of it is unstructured: customer documents, scanned forms, annual reports, registry extracts, transaction history, screening results, negative news. The analyst opens each one, finds the detail, copies it into the CLM by hand, compares it with the next source, and chases the mismatches. That, Oskar says, is most of the working day, and it is why the case queue grows.

Take a periodic review of a trading company. The customer form names two owners, the registry extract names three, and one of those is a holding company in another country. The annual report arrives as a scanned document. An hour into the copying, the analyst notices that the ownership percentages do not add up.

“That last observation is the investigation. Everything before it is what an experienced investigator was not hired to do.”

Transaction monitoring is no different, he adds. The judgment is whether a pattern of transactions fits the business the customer claims to run. The hours go into assembling the pattern.

Why hiring does not close the gap

The obvious response to a growing queue is more people, and Oskar’s point is that this option is weaker than it looks. The budget is hard to defend when the driver sits outside the bank and keeps growing. Even with the money approved, an experienced investigator takes months to recruit and longer to become effective.

“Meanwhile the backlog is what the regulator sees, and overdue reviews are what the board hears about.”

The shortcut most banks have already tried

Most banks, he says, have reached for the off-the-shelf AI assistant first, the kind that now sits in every office suite. The investigator pastes case text into a chat window and gets a summary back.

“It helps with the writing and changes nothing about the queue. The assistant does not know how the bank defines suspicious behavior, it cannot read the source systems, and everything it produces has to be checked against the sources anyway. It adds a step to the case instead of removing one.”

Agents do the administration, people do the investigation

The split Oskar argues for is simple to state. AI agents take the administrative steps: read the documents, pull the data from the source systems, fill in the CLM, compare every detail across sources, screen the parties, and reason against the bank’s own rulebook rather than general knowledge. The investigator opens a finished file where every statement points to the source it came from, and gaps are flagged. The decision is theirs.

“The hours do not disappear. They move from copy-paste into the CLM to investigation,” Oskar says. “And that is why the quality argument is stronger than the speed argument. A case that gets a full hour of investigation instead of a full hour of data entry is a better case.”

Where agents still fail

He is equally clear about the limits.

“Bad scans. Handwritten forms. Ownership chains that run through jurisdictions with weak registries. Two sources that disagree and no third to settle it. Those cases go to a person, as they always did. The difference is that the person receives them with everything else already assembled and the disagreement already marked.”

How a bank should test it

His advice to a bank considering this starts before any technology.

“If you have not measured today’s time per case, you are not ready to start. Measure first.”

From there it stays deliberately small. One case type, run all the way through over a few weeks on anonymized test cases, never live data. Decide what good looks like before anything gets built, and keep a person reviewing every case. At the end there is one decision, go or no go, and the bank makes it.

"A capability that works in a controlled test has absorbed nothing. It counts when it runs inside the daily flow, where the volume sits."

The number worth watching

Asked what he would measure at the end of the year, he is quick to answer:

“How much faster the queue moves, with the same people making the calls. When the board asks what the bank is doing about AI in financial crime, that is the answer that survives the follow-up questions.”