The US Claims Stack — legacy, modern, AI-native
What US carriers run to receive, triage, estimate, and resolve claims, arranged by the generation of each tool in the workflow. The map is honest about the gaps: AI-native is not always better than modern; modern has sometimes not arrived yet; legacy is still doing most of the work in most places.
Legacy
Pre-cloud workflow — people, paper, on-prem software
Modern
Cloud SaaS, classical ML, template-based automation
AI-native
Built around deep learning or LLMs from day one
| Action | Legacy | Modern | AI-native |
|---|---|---|---|
| FNOL intake Receiving first notice of loss and opening a claim file. | Phone call to call-centre + paper form | ||
| Claims triage Deciding routing, severity, and complexity at intake. | Adjuster manual review of notes and photos | ||
| Auto damage estimation from photos Turning claimant photos into a line-level repair estimate. | In-person adjuster inspection with estimatics by hand | ||
| Property damage estimation from photos Turning claimant photos into a property repair estimate. | Adjuster site visit with ladder and measuring tape | ||
| Property pre-inspection imagery Assessing a property before binding, without a physical site visit. | Physical inspection, or no inspection at all for simple risks | ||
| Claim document extraction (IDP) Turning unstructured claim documents — medical records, police reports, claim forms, invoices — into structured data. | Offshore data-entry keying, or template-based OCR in OnBase / FileNet | ||
| Claims routing and shop assignment Assigning repair work to the right body shop or service provider. | Adjuster phone calls + paper shop lists | ||
| Claims fraud detection Flagging suspicious claims for SIU review. | Rule-based scoring + manual SIU review | ||
| Subrogation recovery Identifying and pursuing recovery of claim costs from responsible third parties. | Manual adjuster identification + external recovery vendors | — gap in our coverage |
How to read the gaps
A row marked “gap in our coverage” is a row where Phidea has not yet published a sourced fiche. FNOL intake also has Five Sigma. Fraud detection also has Sprout.ai. The methodology requires three publicly sourced signals per tool before it enters the map.
A public case study or press release speeds things up — send us one and the tool moves up the backlog.
How we classify generations
- Legacy
- Pre-cloud workflow — people, paper, on-prem software
- Rule: pre-cloud or on-prem workflow. Often a process more than a product.
- Modern
- Cloud SaaS, classical ML, template-based automation
- Rule: cloud SaaS, founded 1995-2015, classical ML or template-based automation at most.
- AI-native
- Built around deep learning or LLMs from day one
- Rule: built around deep learning or LLMs from day one; founded mostly post-2015.