From 4 Hours to 15 Minutes: How an Insurance Company Automated Claims Triage

Overview

Keystone Mutual Insurance, a mid-market property and casualty insurer serving 200,000 policyholders, was spending an average of 4 hours per claim on initial triage — reading the submission, classifying the claim type, routing it to the appropriate adjuster, and screening for fraud indicators. After deploying the Dynamic Decision Router (Blackboard Architecture), claims are now analyzed, classified, routed, and screened automatically, reducing triage time to 15 minutes per claim while improving fraud detection accuracy by 38%.

The Challenge

Keystone Mutual underwrites homeowner, auto, and small commercial policies across 14 states, processing roughly 4,200 claims per month through 65 adjusters organized into specialty groups. The 800-person company had grown its policyholder base 40% over three years, but its claims operation hadn't scaled to match.

Triage was almost entirely manual. A coordinator would read each submission — structured form, scanned document, or phone transcript — classify it by type and severity, route it to the right adjuster, and review it against a 47-item fraud checklist. This took 4 hours per claim. Simple claims still took 90 minutes because the coordinator followed the same process as complex multi-vehicle accidents. The system had no concept of proportionality.

Misrouting compounded the problem. Twenty-nine percent of claims went to the wrong adjuster specialty. Each misroute added 11 days to resolution and required a full re-review. A property claim with hidden liability exposure might land with the property team, only to be re-routed weeks later when an attorney surfaced a slip-and-fall allegation.

Fraud screening was the weakest link. The 47-item checklist was binary and linear — no weighting, no pattern recognition. Three low-significance flags triggered the same SIU referral as one high-significance flag. Only 14% of SIU escalations resulted in confirmed fraud; the remaining 86% consumed investigator time on legitimate claims.

"We were drowning in volume," said David Kowalski, Keystone's VP of Claims. "Every claim got the same 4-hour treatment whether it was an $800 windshield replacement or a $400,000 house fire. And our adjusters re-did the triage coordinator's work because they didn't trust the initial classification."

The Solution

Dynamic Decision Router (Blackboard Architecture)

The Blackboard Architecture is built around a shared knowledge space — the "blackboard" — that multiple specialist agents read from and write to as they progressively build understanding of a problem. For claims triage, this architecture maps naturally to the multi-dimensional assessment that each claim requires.

When a new claim enters the system, it's posted to the blackboard as a raw submission. Four specialist agents then analyze it in parallel, each contributing a different dimension of understanding.

The Classification Agent determines claim type, sub-type (e.g., water damage vs. fire vs. wind), and severity tier from the narrative, documents, and form data — writing its classification with a confidence score and supporting evidence.

The Routing Agent matches the classification against adjuster specialties, workloads, geographic assignments, and expertise. A commercial property claim in Louisiana goes to an adjuster who understands Louisiana's direct-action statute, not just the next name in the queue. Claims requiring multi-specialty coordination get team assignments.

The Fraud Screening Agent replaces the binary checklist with weighted pattern analysis. It cross-references claim data against policyholder history and known fraud typologies. A claim filed 30 days after inception with photo metadata inconsistencies and an address matching three recent claims scores very differently from one that simply triggered the "new policy" flag. The risk assessment goes to the blackboard, where the routing agent factors it into assignment decisions.

The Completeness Agent checks for required documentation by claim type. A bodily injury claim without medical records gets flagged immediately with a specific request list — eliminating the back-and-forth that previously added days.

The blackboard is the key element. Each agent sees what the others have written and can revise its assessment. A fraud flag causes the Routing Agent to assign a senior adjuster or escalate to SIU. Missing documentation causes the Classification Agent to mark its confidence as preliminary. This collaborative analysis happens in minutes because the agents work in parallel through the shared blackboard rather than sequential handoffs.

Keystone deployed in three phases over eight weeks — auto claims first, then property, then commercial — each with a two-week parallel run comparing system decisions against human coordinators.

The Results

Keystone measured results over the first 120 days of full production deployment across all claim types.

• Triage time dropped from 4 hours to 15 minutes per claim. Simple claims processed in under 5 minutes; the average reflects the full mix including complex multi-type claims.
• Adjuster routing accuracy improved from 71% to 96%. The remaining 4% misroutes involved claim types that evolved after filing — not classification errors.
• Fraud detection improved 2.3x while false SIU referrals dropped 61%. The SIU team's confirmed fraud rate rose from 14% to 47% of referred claims.
• Claims per adjuster increased 3x. With accurate routing and upfront documentation requests, adjusters spent time on evaluation rather than re-triage.
• Average resolution time decreased by 8 days, driven by eliminating misrouting delays and catching incomplete documentation upfront.

The system reached target performance within six weeks, matching or exceeding human triage accuracy on 94% of claims by week two of each phase.

"The system routes claims the way our best senior adjusters would — if they had time to read every document, check every history, and cross-reference every pattern before making a decision. Our triage coordinators used to be experienced adjusters themselves, and they were the first to admit they couldn't hold 47 fraud indicators in their heads while simultaneously classifying a claim and matching it to the right adjuster. The system doesn't have that limitation." — David Kowalski, VP of Claims, Keystone Mutual Insurance

Key Takeaways

• Claims triage is a multi-dimensional problem that benefits from parallel analysis. Classification, routing, fraud screening, and completeness checking are distinct tasks that inform each other. The Blackboard Architecture allows specialist agents to work simultaneously and share insights through a common knowledge space, rather than processing sequentially.
• Proportional processing matters. The old system spent the same 4 hours on an $800 claim as a $400,000 claim. Automated triage allocates attention proportionally — simple claims flow through in minutes, while complex claims receive deeper analysis.
• Fraud detection improves when you analyze patterns, not checklists. The weighted, multi-signal approach caught more actual fraud while dramatically reducing false referrals. The SIU team's time is now spent on genuinely suspicious claims rather than clearing false positives.
• Routing accuracy is an underrated driver of claims efficiency. The 25-percentage-point improvement in routing accuracy (71% to 96%) eliminated thousands of hours of re-routing delays and duplicate reviews per year — a compounding efficiency gain that improved every downstream metric.

Ready to Explore Intelligent Routing for Your Claims Operations?

If your claims triage process treats every submission the same regardless of complexity, or if misrouting adds days to your resolution timeline, the bottleneck is in your decision architecture. Agentica's Dynamic Decision Router integrates with existing claims management systems and can be configured for your specific claim types, adjuster specialties, and fraud typologies. Schedule a consultation to discuss how intelligent routing applies to your claims operation.