A mid-size property and casualty insurer modernized its claims processing workflow to reduce settlement cycle time and improve fraud detection. The previous workflow was heavily manual: adjusters received claim packets by email, manually entered data into the claims management system, and escalated fraud concerns based on individual judgment. Average cycle time from first notice of loss to settlement was 18 days.

The Transformed Workflow

The new workflow is AI-assisted rather than fully automated. AI handles the data extraction and initial triage; adjusters focus on coverage decisions, negotiations, and cases requiring judgment.

Intake and extraction - Claim documents (FNOL form, photographs, police reports, repair estimates, medical records) are uploaded to a claims portal and stored in S3. An extraction pipeline using Amazon Textract and a Bedrock post-processing step produces a structured claim record: incident details, claimed amounts, supporting document inventory, and identified parties.

Policy matching - The structured claim record is matched against the policy database to pull coverage details, deductibles, endorsements, and claims history. This matching was previously manual and took 20-40 minutes per claim.

Automated liability scoring - A scoring model rates liability probability (0-1) and claim complexity (low/medium/high) based on incident type, supporting documentation completeness, and policy terms. Simple, low-complexity claims with high documentation completeness are flagged for straight-through processing. Complex or ambiguous claims are routed directly to senior adjusters.

Fraud detection - A separate model analyzes the claim for fraud indicators: provider patterns against known fraud networks, claimant claim frequency, document anomalies (metadata inconsistencies, image manipulation signals from Rekognition), and cross-claim network patterns. Flags are scored by severity and passed to the Special Investigations Unit queue.

Adjuster interface - Adjusters see an AI-generated claim summary, the structured extracted data, fraud flag details, and a pre-populated coverage analysis. Decision support, not decision replacement - the adjuster retains full authority over coverage decisions and settlement amounts.

Outcomes

After 12 months in production:

  • Average cycle time reduced from 18 days to 11 days for standard claims
  • Straight-through processing rate for low-complexity claims reached 34%
  • Fraud referral rate to SIU improved by 28% without increasing adjuster workload
  • Document data entry errors eliminated for extraction-processed claims

Key Lessons

Explainability is not optional in insurance. Coverage decisions must be defensible to regulators and policyholders. Every AI-assisted decision includes a structured explanation of which factors contributed to the scoring. Adjusters can see exactly what data was used and override any recommendation with a documented reason.

Feedback loops require explicit design. Initially, no mechanism captured whether adjuster overrides represented AI errors or legitimate judgment calls. Adding an override reason taxonomy and analysis process identified systematic extraction errors that were corrected in the next model update.

Change management takes longer than technical deployment. The AI system was technically complete four months before it reached full adoption. Adjuster training, process documentation, and trust-building through supervisor-accompanied use drove adoption more than technical features. Teams that received hands-on training alongside their supervisors adopted faster than those who received documentation alone.