Signal from Noise: An LLM Triage Engine for Compliance Mail
Finds the 1-3 real regulatory notices a month hidden in 10–300 emails a day, with zero missed since launch. Claude classifies every incoming email in a Power Automate flow, extracts the key details and suggests next steps, then opens and updates the right Jira issue until the matter is closed.
The problem
The compliance inbox gets anywhere from 10 to 300 emails a day, but only 1 to 3 of them a month are real regulatory notices. The rest are fraud attempts and noise. A notice that slips through the cracks is a missed legal deadline, so the goal was never to save time. It was to make sure no notice is ever missed.
What I built
A Power Automate pipeline that sends each incoming email to the Claude API, which sorts it into a notice, a fraud attempt or noise and suggests next steps. For notices, it extracts the issuing body and the US state, then opens a Jira issue with the original attachments.
How it works
- Classification: Claude weighs several signals in each email to label it as a notice, a fraud attempt or noise, and generates suggested next steps.
- Extraction with Claude: an HTTP action in Power Automate sends the email to the Claude API, which returns a fixed JSON format (issuing body, US state and the other ticket fields). The flow uses that JSON to create the Jira issue directly.
- Escalation matrix: follow-up emails on an open matter update the issue's priority, assignee and deadline.
- Closing the loop: resolution notices close the matching issue without anyone touching it.
Moving the AI steps to Claude
The flow first used Microsoft's built-in AI in Power Automate. I replaced those steps with direct HTTP calls to the Claude API, running my own prompts for classification, extraction and next steps. Two reasons drove the switch. We wanted a fuller read of each email: Claude reasons over everything in the notice and brings general compliance knowledge to it, instead of only pulling out fixed fields. And token-based pricing made the running cost easier to predict and budget.
Result
Since launch, the flow has caught every notice. Monthly audits of the inbox check that every compliance notice was processed, and none has been missed. That's 1 to 3 tickets a month found among 10 to 300 emails a day.
My role
I built the whole thing solo, from prompts to Power Automate flows to the Jira integration. I designed the escalation matrix from the client's existing process. Fraud attempts had always been handled by hand, and several came through during testing, so I added a separate path for fraud, and another for spam and marketing mail, to keep both out of the ticket queue.