We built a system that automatically monitors these government sources, determines what is relevant to each customer, and drafts a clear summary of what changed using Claude 3.5 Sonnet on AWS Bedrock. The system translates the legalese into a strict data model:
- Obligation: The legal requirement the provider must follow.
- Policy: The provider's internal document that explains how they will comply.
- Action: The specific task assigned to fix or update a policy.
The Delta Engine: We integrated live APIs where available, but many government sites required custom HTML scrapers. When a page is scanned, it creates a hashed snapshot. On the next scan, it compares the hash against the previous version, isolates the exact HTML delta, and flags it for the LLM to process.
The Human-in-the-Loop Rule: The LLM drafts the policy and impact, but it never publishes automatically. It routes to a human reviewer who sees a dashboard with the AI's confidence score and must explicitly click Approve, Reject, or Approve with edits.