The financial crime block
Counterparty risk, run as one loop - every alert read against the whole relationship, the case assembled before the analyst arrives, every action governed.
The pain
“Analysts drown in alerts that go nowhere, and real risk hides in the queue.”
Financial crime is the bottleneck on growth - and it isn’t a people problem. AML, KYC, sanctions and fraud each run as their own queue, each reading its own slice of the customer. An alert is judged on the transaction in front of it, because the systems that hold the rest of the relationship - the history, the counterparties, the cause that cleared three times before - were never built to be read together. So the queue fills with things that were never risk, and the things that are risk sit in it, waiting their turn. In wealth and legal this is the AML and KYC book; in distribution, sanctions and export controls; in insurance, broker and insured due diligence.
What the block is
The object is the counterparty’s risk state - one continuous picture per counterparty, from onboarding through monitoring to exit, with financial crime run as one loop rather than four queues.
Everything the relationship produces feeds one live record: screening results, transaction patterns, adverse media, list updates, documentation, the outcomes of every past alert. A model reads it permanently. Your screening and monitoring systems stay exactly where they are - the block reads from them, it does not replace them.
The loop in action
Each output states what it saw, why it matters, what it recommends, and how confident it is - every fact linked back to its source, in an order an investigator can read and a regulator can examine.
You set the dial
Every capability has four positions: observe - it reads silently and measures its own accuracy; recommend - it surfaces the case and a proposed action to a named owner; draft - it prepares the action for a human signature; act - it executes within agreed bounds and logs everything.
Everything starts at observe. Promotion is earned on the measured record and reversed instantly if it doesn’t hold. Escalation decisions and exits stay human - by policy, permanently if that is your policy. The mechanics are the immune gate described in our framework.
How it lands
It starts narrow: one segment of the book, the block at observe against your own past outcomes. Nothing is replaced, nothing is migrated - the first signals flow from the systems you already run. From there it is a bounded engagement - blueprint, build, handover - and your team owns the result outright: the signal map, the judgement rules, the dial settings, the working loop.
The backtest
Take three past alerts: one that was real, one false positive that burned a week, and one that surprised you late. What did your own records know, and when did you find out?
That is the first question the diagnostic answers. Take the ten-minute diagnostic, or write to hello@somai.studio.
Runs on our framework · related reading: Most AI spend never reaches the P&L