🤯 We automated an insurer’s entire AML review process with AI agents. Full breakdown of how we built it below ⬇️ 1. A government alert comes in flagging a suspected money launderer. 2. Workflow executes internal database search + web search → AI agent issues a pre-screening decision 3. Relevant SOPs pulled from SharePoint 4. Compliance memo drafted + sent to human for review 5. If approved, a browser-based agent logs into the portal and closes the account ⏰ 10mins start to finish (down from 3-4 hours per case). No account closed without human sign-off. This is one of the StackAI use cases I’m proudest of- because it shows that with the proper guardrails and gates, we can make agentic AI run right in the world’s most regulated industries. If you're in risk/compliance and want to see this in depth, shoot me a message!
Curious where the 10 minutes goes when the pre-screening decision and the memo disagree with each other. Database search says one thing, adverse media says another, and the reviewer now needs to see which source produced which claim rather than a clean summary. Does the memo carry per-field provenance, or does the contradiction get resolved before it reaches the human?
Really impressive Bernard Aceituno! And the thing worth noticing isn’t 3-4 hours down to 10 minutes — it’s that nothing irreversible happens without a human. Pre-screening, SOPs at runtime, memo drafted for review, sign-off before the account closes. That’s the part everyone else skips. Now the honest half. We’re building an AI-native claims company — same shape of problem, different regulator. Carmack liability instead of OFAC, denial letters instead of SARs. Exactly the kind of thing your architecture is built for. But there’s no door for a company my size. It’s a free tier with 500 runs, or “get a custom quote” and a procurement cycle. Nothing in between for a founder with real regulated workflows and a small budget. And when I reached out, I never heard anything back. Not a no — just nothing. Which is the part I keep turning over. StackAI went through YC and raised a $3M seed once. Today’s startups are the enterprises you’d be selling to in three years, and there’s no on-ramp for them. So when a post ends with “shoot me a message” — what should someone actually expect?
Cool build! Sign-off gate is still crucial and could only be relaxed over time. One other question: where did the regulatory input come from? Looks like the customer’s own SOPs out of SharePoint. Did you ever look at licensing regulatory content from a vendor, or was customer-supplied always the assumption?