Ertas for Insurance Claims Processing

The Challenge

Insurance claims processing is a document-intensive workflow where adjusters must review claim submissions, verify coverage, assess damages, and determine payout amounts, all while adhering to policy terms, regulatory requirements, and internal guidelines. A single auto insurance claim can involve police reports, medical records, repair estimates, photos, and policyholder statements, each requiring careful review and cross-referencing. The volume is staggering: large insurers process millions of claims annually, and any delay in processing directly impacts customer satisfaction and retention.

Generic AI tools cannot handle the complexity of claims adjudication because they lack understanding of insurance-specific concepts: coverage endorsements, deductible structures, subrogation rights, coordination of benefits, and state-specific regulatory requirements. A model that does not understand the difference between a comprehensive and collision claim, or between actual cash value and replacement cost, will produce useless output regardless of how fluent its language generation appears. Claims processing requires both domain knowledge and consistency: the same claim should receive the same assessment regardless of which adjuster or AI system processes it.

The Solution

Ertas enables insurance organizations to fine-tune claims processing models on their own adjudication history, policy terms and internal guidelines. The resulting model learns the organization's specific products, coverage structures and claims handling procedures. In Ertas Studio, claims operations teams train models on historical claims data: claim submissions paired with adjuster assessments, coverage determinations and payout calculations. Training runs on Ertas's cloud GPUs, so the training set is uploaded to Ertas; tokenize or remove policyholder PII before upload.

The fine-tuned model can support several stages of the claims pipeline. At intake, it triages claims by complexity and routes them to the right queue. During assessment, it extracts key information from claim documents and cross-references it against policy terms. For straightforward claims, it can draft a recommended decision with supporting rationale. Exported from Hub and run on the insurer's own servers, the model processes live claims inside the insurer's network, and record keeping stays in the claims systems the insurer already uses. Teams that already send claims tasks to a frontier model through API calls can also ask about Ertas's adaptive model endpoint: a custom model trained on their own requests takes over a growing share of them once it passes their evals, hard cases stay on the frontier model, and adopting it means pointing existing API calls at a new model name. Human adjusters review AI-recommended decisions and keep their expertise for complex cases.

Key Features

Example Workflow

A regional property and casualty insurer processes 50,000 claims annually with an average handling time of 14 days. The claims operations team exports 200,000 historical claims, including submission documents, adjuster notes, coverage determinations and payout records, tokenizes policyholder PII, and uploads the result to Ertas Studio as JSONL datasets. Senior adjusters review samples in Data Craft and approve the examples. In Studio, the team fine-tunes separate models for auto, homeowners and commercial lines, and compares runs against a held-out set of claims decided by its own adjusters. The models are exported from Hub and run on the insurer's own servers. At intake, the triage model classifies claims by complexity and routes them: simple auto glass claims go to the fast-track queue, while complex liability claims go to senior adjusters. For fast-track claims, the extraction model pulls key data points (date of loss, damage description, applicable coverage, deductible) and the adjudication model recommends a payout with supporting rationale. Adjusters review and approve these recommendations, and the team tracks handling time and override rates against its own baseline.