• AWS-native AI integration · ships in 6–10 weeks

Brokerage cuts submission turnaround 75%

  • Agentic Workflows · Insurance
  • Independent commercial-lines brokerage — ~140 staff, 6 offices, $40M+ premium, Applied Epic

The problem

CSRs were spending 45–70 minutes per new commercial submission re-keying ACORD 125/126 data into Epic and matching against carrier appetite. Account executives were getting packets a day late, costing them on quote-by deadlines.

Our approach

What we built

  • Intake agent reads inbound ACORD PDFs and broker emails; extracts insured, schedule of locations, prior carrier, and loss runs.
  • Appetite-matching agent grounded in a Bedrock Knowledge Base of 14 carrier appetite guides; returns a ranked carrier shortlist with cited reasoning.
  • Draft-output agent producing the Epic activity note and carrier submission cover letter for CSR review.
  • Human-in-the-loop UI: every output goes to a CSR review queue before any external send — no autonomous external action in v1.
  • Eval rubric with Layer-1 unit checks and Layer-2 LLM-as-judge scorers, tuned against 80 historical submissions.

Caveats

The first three weeks exposed a long tail of non-standard schedule formats — Excel sheets emailed as inline images — that the agent couldn’t parse; we absorbed ~18 unplanned hours of Textract preprocessing. Workers’ comp appetite-match scores stayed below threshold on a subset, and now route to a human-only path. The brokerage’s owner pushed for auto-send to carriers; we declined and put a governance recommendation on file.

Stack. AWS Bedrock (Claude + Nova Lite for classification) · Bedrock Knowledge Base · Lambda · Applied Epic REST · Langfuse self-hosted.

Outcome

Median CSR time per submission fell from 58 minutes to 14. Same-day submission packets to AEs rose from 38% to 90%. Two CSR seats redirected to new-business outbound.

75%
reduction in CSR time per submission
90%
same-day packet delivery (was 38%)
92%
appetite-match agreement with senior brokers