The Challenge: A Critical Bottleneck in High-Volume Quoting
Amwins serves as benefits program manager for one of the largest Professional Employment Organizations (PEOs) in the United States, generating over 8,000 SMB group medical insurance quotes annually. A core feature of our underwriting process is providing detailed, defensible "apples-to-apples" pricing comparisons between a group's current medical coverage and our proposed replacement coverage.
For years, this critical capability relied on a manual, time-intensive process: underwriting staff reviewing PDF carrier and TPA invoice documents and manually copying subscriber details into our census tools. This reconciliation step consumed 15-30% of total quote time—anywhere from 20 minutes to 2 hours per quote, depending on group size and complexity.
With quote volume increasing 30% year-over-year and no corresponding headcount growth, this bottleneck was becoming unsustainable.
The Search: Why Traditional Solutions Fell Short
We spent years evaluating potential solutions:
Business Process Outsourcing (BPO): Failed on two critical dimensions. Round-trip workflow times actually exceeded our underwriters' manual processing, and quality issues created additional downstream work rather than eliminating it.
OCR Solutions (with and without AI): Required extensive "care and feeding" to manage the wide and constantly changing range of invoice formats we receive. The maintenance burden outweighed the efficiency gains.
The fundamental challenge: we process invoices from over 80 unique carrier formats (and growing daily), each with its own structural variations. We needed a solution that could handle this complexity without constant retraining.
The Solution: Document-Centric AI That Actually Works
Upstage's Information Extraction was the first solution that felt like it had cracked the code on reliable schema extraction from highly variable document sources.
The difference? Their document-centric LLM with integrated OCR handled edge cases and novel invoice structures with minimal prompt tuning—better than anything we'd evaluated over several years of searching.
Our implementation approach was straightforward:
1. Defined our target schema with inline descriptive prompts
2. Let the model do the heavy lifting
3. Minimal tuning during evaluation yielded 90%+ success rates with 100% accuracy on our test set
The ease of integration into our Azure/M365 platform meant we built and deployed a full MVP in under a month.
The Results: Immediate and Measurable Impact
Production metrics speak for themselves:
● 1,100+ invoices processed in the first month
● 200+ invoices processed daily by our team of 18 underwriters
● Processing time reduced from 20 minutes-2 hours to under 5 minutes
● 1.5 FTE worth of capacity reclaimed per week across the team
● 3-month payback period on build costs
● 80+ unique carrier invoice formats successfully processed (and counting) The Adoption Story: When Users Won't Let Go
We initially planned a limited pilot with select underwriters, intending to incorporate their feedback before rolling out a fully-featured version 1.0.
Reality had other plans.
Once the solution was in our underwriters' hands, there was no taking it away. The tool landed at exactly the right moment—as our team was working to keep up with 30% volume growth without additional headcount. Our VP of Underwriting made it clear: she wasn't giving up the invoice extraction tool "over her dead body."
Within days, all 18 underwriters were benefiting from the solution.
The Business Impact: Beyond Time Savings
The efficiency gains enabled us to:
● Meet SLA commitments without headcount growth
● Absorb 30% YoY volume increases with existing team capacity
● Deliver faster quote turnaround to our PEO client
● Redeploy underwriter time to higher-value activities
What's Next: Expanding the Partnership
Based on our success with invoice extraction, we're exploring additional use cases across Amwins Group Benefits, including:
● Automated extraction from Summary of Benefit and Coverage (SBC) documents
● Data extraction from carrier benefit guides and plan grids
Upstage is proving their worth as a proactive partner, working to understand and prioritize our product roadmap while expanding their Information Extraction capabilities to meet our evolving needs.
The Takeaway: Choosing the Right AI for the Job
This success story reinforces a critical lesson in AI adoption: specialized solutions often outperform general-purpose tools for complex, domain-specific challenges.
After years of evaluating alternatives, we found that Upstage's document-centric approach delivered what we actually needed:
● Reliable performance across highly variable document formats
● Minimal ongoing maintenance burden
● Fast time-to-value
● Seamless integration into our existing infrastructure
When you're processing 8,000+ quotes annually with 80+ invoice formats, you need AI that's purpose-built for the complexity. Upstage delivered exactly that.
