It’s 2:13 a.m. on an invented Tuesday, and a transaction pattern just crossed a threshold. Follow the alert and you’ll see where AI belongs in casino AML work, and where it absolutely doesn’t.

The system that raised the score did honest work: grouped related events across three days, connected two instruments to one patron, noticed the sequence resembled structuring. FinCEN’s casino guidance publishes red flags for exactly this reason, and a risk-based program is expected to watch for them. So far, computation is doing what computation does.

Morning arrives, and here’s the fork in the road.

In the deployment we build, the analyst opens a case file the machine already assembled: the transactions in sequence, the relationships it inferred with each inference labeled, the patron context, the prior alerts and their dispositions. Every claim links back to source rows. She corrects one entity link the system got wrong, the correction persists alongside the original, and she makes the call: reasonable explanation, continued monitoring, or escalation under the institution’s standard. Her reasoning’s documented, and the record’s built to survive an examiner.

In the deployment we won’t build, the score itself is treated as the finding. Alerts multiply because more alerts look like diligence, reviewers drown, and dismissal becomes a reflex. That program looks busier and sees less every quarter. False positives were never free; they spend analyst attention, the scarcest resource on the compliance floor.

The difference between the two is where judgment sits, not model quality. An unusual transaction and suspicious intent are different findings, and only one of them comes out of a database. The analyst knows the floor, the season, the machine that’s been paying out oddly, the difference between a nervous tourist and a practiced smurf. Columns don’t carry that, and I’ve stopped pretending they might.

So our rule for this work is plain: computation assembles context, humans assign meaning, and the queue gets measured on quality of review rather than volume of detection. Back at 2:13 a.m., the system’s job was to make the story visible by morning. The verdict always belonged to the person drinking the coffee.