Financial Services / Mortgage Capital Markets · AI/ML Engineering · Agentic AI on Amazon Bedrock · BUILDSTR
Challenge
Scale underwriting without scaling headcount
MAXEX runs a mortgage loan exchange in the secondary market, where originators and investors buy and sell loans. Its underwriting team fields 100–400 routine emails a day, 30,000 to more than 100,000 a year: status checks, document requests, and general questions from sellers and investors. Each one was read, sorted, looked up, drafted, and checked by hand. Routine email took 60–70% of an underwriter’s day, and pulling details by hand had a 15–20% error rate.
- About 15 minutes of underwriter time per routine email, with replies taking anywhere from 2 hours to 3 days
- Response quality depended on who picked the email up and how busy they were
- Institutional know-how lived in individual underwriters’ heads and was at risk with every departure
- Slow, uneven replies put seller and investor SLAs at risk
- Keeping up with growth meant hiring in step with volume, at about $8.50 fully loaded per email
Solution
An agentic inbox with a human in the loop
In less than six weeks, BUILDSTR designed and deployed a serverless, agentic workflow on AWS. The solution orchestrates a complex and data-heavy workflow leveraging Strands agents on Amazon Bedrock AgentCore. It reads each email, works out what’s being asked, gathers the right context, and drafts a reply for underwriter review. Nothing goes out without a person approving it.
- Ingest: email arrives through Amazon SES and is parsed and threaded, so the agent sees the full conversation history
- Understand: Amazon Bedrock with Claude Sonnet 4.5 classifies intent (status query, document request, complaint, general inquiry) and extracts loan numbers, addresses, dates, and amounts using a hybrid of rules and the LLM
- Ground: two Amazon Bedrock Knowledge Bases on Amazon OpenSearch Serverless are queried in parallel, one over past resolved cases and one over current underwriting policy and SOPs. Keeping them separate improved relevance, so replies reflect both what worked before and what policy requires now.
- Draft and check: the Strands agent on Amazon Bedrock AgentCore drafts a reply, then scores it for confidence, completeness, and tone before it reaches a reviewer
- Review: underwriters approve, edit, or reject every draft in a React review console behind Amazon Cognito and Amazon CloudFront
- Audit: every step is logged in Amazon DynamoDB for compliance, and the review decisions feed prompt tuning
- Run: AWS Lambda orchestrates the flow, with AWS KMS encryption and Amazon CloudWatch monitoring. About 177 resources are delivered as Terraform, aligned to the AWS Well-Architected Framework.
Result
Four times the capacity, in production
The inbox went into production within the engagement window and now handles live underwriting email for the team. Iterating on the prompts during the build raised extraction accuracy from 87% to 96%.
- Each underwriter handles about four times as many emails a day: 150–200, up from 30–40
- Processing time per email down 83%, from about 15 minutes to 2.5
- 92% of AI-drafted replies approved without changes, 6% edited, 2% rejected
- Cost per email down about 94%, from $8.50 to about $0.45, roughly $288K a year in cost avoidance
- Data-entry errors down from 15–20% to under 4%
- Volume growth absorbed without hiring in step; underwriters spend their time on judgment calls
“What used to take 15 minutes per email now takes less than 3 minutes, and the AI-generated responses are remarkably accurate — over 90% are approved without modification.” — Brian Ruzycki, CTO of MAXEX
