Custom AI Models for Claims Processing and Underwriting
Ertas lets insurers train custom models on their own claims, policies and adjuster decisions, then run them on their own servers, where the policyholder data processed in production stays with you, or through the adaptive model endpoint.
The Challenges You Face
Claims Processing Is Document-Heavy and Slow
Every claim involves medical records, police reports, repair estimates, correspondence and policy documents. Manually extracting the relevant information, verifying coverage and routing claims consumes enormous adjuster time, and the delays frustrate policyholders.
Policyholder Data Is Heavily Regulated
State insurance regulations, GLBA and, for health claims, HIPAA put strict conditions on how policyholder information is processed. Sending every claim to an external AI API brings vendor reviews and data agreements that slow adoption.
Underwriting Models Need Institutional Knowledge
Effective underwriting depends on decades of institutional knowledge about risk factors, loss patterns and pricing strategies. Generic AI models lack this domain expertise, and the proprietary nature of underwriting data makes external training partnerships risky.
Fraud Detection Requires Nuanced Pattern Recognition
Insurance fraud comes in many forms, from staged accidents and inflated claims to provider billing schemes, each with subtle indicators that differ by line of business. Generic anomaly detection misses the domain-specific patterns that experienced adjusters recognize.
How Ertas Solves This
Ertas Studio trains a custom model on your own claims history: first notice of loss reports with the fields your adjusters extract, claims with the routing decisions your team made, policy wording with the coverage questions it answers. Senior adjusters and underwriters can review, correct and approve those examples in Data Craft, Studio's guided dataset builder, so their judgment becomes training data. Training runs on open models through a visual canvas, with several runs in parallel to compare.
Training runs on Ertas's cloud GPUs, so the training set is uploaded to Ertas. Remove policyholder identifiers the task does not need before you upload, and delete datasets when you no longer need them.
The trained model is yours to download as GGUF or full safetensors and run on your own servers with llama.cpp, Ollama or vLLM. Claims processed in production stay in your data center. Meeting state data security rules and GLBA remains your own work, and running the model in-house keeps production claims data out of another vendor's systems.
For a high-volume task that runs on a frontier API today, the adaptive model endpoint moves a growing share of requests to a custom model once it passes your evals, with hard cases left on the frontier model. It runs Ertas-managed, in your own cloud, or on-premise.
Key Features for Insurance Companies
Adjuster-Reviewed Training Data
Data Craft lets adjusters and underwriters edit example exchanges, version them and mark them approved, encoding their knowledge of coverage nuances, red flags and risk factors into the training set.
Policy-Specific Model Training
Fine-tune open models on your actual policy language, coverage structures and claims outcomes. The resulting model learns your products specifically, so its extractions and routing suggestions sit closer to what your adjusters would do.
Run It in Your Data Center
Trained models are listed in Hub and download as GGUF or full safetensors. Serve them on your own hardware with llama.cpp, Ollama or vLLM, with no Ertas dependency at runtime.
Adaptive Model Endpoint
For one high-volume task, the adaptive model endpoint gives you a single API. A custom model trained on your own requests and your team's feedback takes over a growing share once it passes your evals, while hard cases stay on the frontier model you already use.
Why It Works
- On the QMSum meeting-summarization benchmark, a 1.2B model trained on Ertas scored 35.41 ROUGE-1 against 32.43 for the best of five hosted frontier models: a small model trained for one task can beat far larger general models on that task.
- Exported models run on llama.cpp, Ollama, LM Studio or vLLM on hardware you control, so production inference needs no connection to Ertas.
- Adjusters contribute through Data Craft, a conversational editor for building and approving training examples, so their expertise reaches the model without anyone writing code.
Example Workflow
A property and casualty insurer wants a model that extracts key data from first notice of loss (FNOL) reports and suggests a loss type and severity for routing. The claims operations team exports 5,000 historical FNOL reports with the fields adjusters recorded, strips policyholder names and contact details, and formats them as a JSONL dataset.
Senior adjusters review the examples in Data Craft, fix inconsistent loss-type labels and approve the set. The team trains runs on two base models in parallel in Ertas Studio and compares them on a held-out set of recent reports.
The insurer's IT team downloads the better model as a GGUF file and serves it on internal servers. Intake adjusters see the extracted fields and suggested routing alongside each new report, confirm or correct them, and the corrections feed the next training run.
Related Resources
Ship AI that runs on your users' devices.
Free plan with 30 credits/mo, no card required. Paid plans from $10/mo USD.