InsurTech Outsourcing · AI

AI-Assisted Claims Processing for a Property & Casualty Insurer

A regional P&C insurer was drowning in manual claims intake. Dillo built LLM-based document extraction for FNOL packets, fraud-signal scoring, and a human-in-the-loop adjuster workbench — cutting claims cycle time by more than half, with the first release in production in four months.

55%

faster claims cycle time

80%

of FNOL documents auto-extracted

+18 pts

NPS improvement on claims

4 months

to first production release

The Client

A regional property & casualty insurer serving homeowners and small commercial policyholders across several US states. Roughly 400 employees, a claims organization of about 60 adjusters, and tens of thousands of claims per year — with volume spiking sharply after every major weather event.

The Challenge

Claims intake was almost entirely manual. First notice of loss (FNOL) arrived as a mix of web forms, emailed PDFs, scanned police reports, contractor estimates and photos. Intake staff re-keyed everything into the core claims system, and adjusters spent the first hours of every claim just assembling a coherent file. During catastrophe surges, cycle times ballooned and complaint volume followed.

Fraud screening was equally manual — a rules checklist applied inconsistently under time pressure. Leadership wanted AI in the claims workflow, but with hard constraints: adjusters had to stay in control of every decision, every extracted field needed to be traceable back to its source document, and the whole thing had to satisfy state regulators and internal audit.

What We Did

Dillo delivered the system as a managed outsourcing engagement: a dedicated team of AI and full-stack engineers plus a delivery lead, working against a milestone plan the client's claims and IT leadership signed off on. We designed deliberately around human-in-the-loop review rather than full automation.

  • Built an LLM-based extraction pipeline on the Claude API that reads FNOL packets — forms, police reports, estimates, correspondence — and maps them to structured claim fields, each with a confidence score and a link to the exact source passage.
  • Added a fraud-signal scoring layer combining extracted facts with policy history and pattern checks, surfacing ranked signals for investigators instead of opaque verdicts.
  • Delivered an adjuster workbench in React: a side-by-side view of source documents and extracted data, one-click accept/correct on every field, and a full audit trail of who approved what.
  • Ran a shadow-mode pilot for six weeks, measuring extraction accuracy against adjusters' manual work before anything touched a live claim.
  • Deployed on the client's Azure tenancy with PostgreSQL, meeting their data-residency and audit-logging requirements from day one.

The first production release shipped in four months. After rollout, the client kept two Dillo AI engineers on as a staff-augmentation follow-on to extend extraction coverage to new document types and tune fraud signals — same engineers, same transparent flat-margin pricing.

The Results

  • End-to-end claims cycle time dropped 55%, with the biggest gains in the intake-to-assignment stage that had been almost entirely manual.
  • 80% of FNOL documents are now auto-extracted and only reviewed — not re-keyed — by intake staff and adjusters.
  • Claims-experience NPS improved by 18 points in the two quarters following rollout, driven largely by faster first contact and fewer requests for documents policyholders had already sent.
  • Fraud investigators report triaging referrals in a fraction of the previous time, working from ranked, evidence-linked signals instead of raw files.
  • The system held up through its first catastrophe surge without adding intake headcount.

Figures reflect outcomes reported for this engagement; they are project results, not audited benchmarks.

Tech Stack

Python Claude API PostgreSQL React Azure

Services Used

Software Outsourcing — an AI delivery team building the claims platform end to end, followed by IT Staff Augmentation for the ongoing AI engineering pod.

Ready to build your team?

Tell us what you need and get your first highly relevant candidate CVs within 24–48 hours.