Dear product manager,
You’ve been asked to write requirements for “AI in AML”. It’s an easy brief to get wrong, because the phrase hides two different products.
The first is detection AI. It sits upstream and decides which transactions get flagged. The second is investigation AI, which sits where an analyst opens the alert and decides what to do. The distinction is the core of our piece on the two types of AI in AML. Most vendors only sell the first.
Write requirements for the second one too
First screen. The analyst sees customer history, prior dispositions and linked accounts without opening other tabs. That’s the gap behind alerts that arrive without context.
Traceability. Every sentence in a summary points to a record. Read why evidence-grounded AI beats answer-generating AI.
Human decision. The tool recommends, a named person decides, and the log shows who.
Memory. When a case closes, reasoning is kept in a form the next case can use.
Data hygiene. FATF warns that poor or biased data fed into a model can be learned by it, so say how data quality is checked.
sanctionscanner
One requirement to resist
Don’t write “reduce false positives” as the only success measure. It describes detection. If the goal is better decisions, measure what analysts see and how well they can explain the result. That is the argument of the next phase of AML is better decisions.
We work on the investigation layer at TechnoXander, and you can see how at the platform page.
Over to you
What is the first line of your AI requirements doc, and does it mention the analyst?
Should traceability be a hard requirement or a nice-to-have?
Hot take: a requirement that can’t be tested in a demo isn’t a requirement. Agree, or fight me?
Warm regards,
Sonal
Top comments (0)