Private AI for insurance claims processing without sending claims to a third party
A claims adjuster's inbox is medical records, repair estimates, police reports, bank details, and a claimant's account of what happened, all attached to a name and a policy number. Summarizing that intake into a consistent format, checking a submitted claim against policy terms, drafting the first response, all of it is a reasonable thing to ask an LLM to help with. Routing any of it through a third-party AI API means medical and financial data about a real person, who never agreed to their claim file being processed by that vendor, leaves your infrastructure.
Not legal advice. Insurance is a heavily regulated industry and claims data usually includes categories subject to specific protection rules depending on jurisdiction (health data especially). This post describes an infrastructure choice, not a compliance opinion. Check with your own counsel and compliance function before using AI anywhere in a claims workflow.
What makes a claim file a different category
A claim file combines several sensitive categories at once: health information for a medical or injury claim, financial account details for payout, and a first-person account that can include details about the claimant's home, family, or habits that go well beyond what's needed to process the claim itself. Health data in particular is treated as a special category under GDPR and under sector-specific rules in most jurisdictions with a regulated insurance market, which raises the bar for how it can be processed, by whom, and under what legal basis, well above the bar for ordinary business documents.
Third-party AI products aren't built around that bar. A general-purpose API has one data processing agreement covering every customer's use case, not one written for claim files with health data attached, and the claimant, who is not your organization's employee or customer in the sense of having agreed to a platform's AI terms, has even less visibility into where their file ends up.
What running the model yourself changes
Self-hosting the model on a dedicated Spark keeps claim documents on infrastructure your organization controls end to end: the model runs on hardware you rent or own, the documents never transit a third-party inference endpoint, and retention is whatever your claims system's existing schedule already specifies. That's the same architectural pattern as law firms handling case material or hospitals handling patient records, applied to claim files instead. It doesn't change who is allowed to see a given claim; wire the same access controls your claims platform already enforces into whatever sits in front of the model, so the AI layer doesn't become an unrestricted door around role-based access that took real work to set up.
A concrete example
An auto claim comes in with a police report, two repair estimates, and photos with a written description. A self-hosted model summarizes the intake into a structured form: date, parties, estimated damage range, whether the submitted documents are complete against the policy's required list, and a plain-language summary for the adjuster to read in thirty seconds instead of five minutes. The adjuster still makes every substantive decision, whether the claim is covered, what it's worth, whether anything about it warrants a closer look, the model just gets the file in front of them faster and in a more consistent shape.
Where this is not a drop-in replacement
An LLM summarizing and triaging claims is not a fraud detection system, and using it as one is a mistake worth naming directly. Real fraud detection in insurance relies on statistical models trained on historical claim and payout data, cross-referencing against known patterns, and specialized tooling built for that specific problem, often with regulatory oversight of its own. An LLM asked "does this claim look suspicious" will produce a confident-sounding answer based on surface features of the text, phrasing, inconsistencies it happens to notice, without the base rates or historical pattern-matching a real fraud model has. Use it to flag a claim for a human to look at more closely, or as one input alongside your actual fraud detection tooling, never as the tool that clears or flags a claim on its own.
Where the hardware fits
Claim files vary in size, a short auto claim versus a lengthy medical claim with imaging reports attached, but a single Spark's 128GB of unified memory handles a large instruct model with room for a long context window on the bigger files. A $0.79/hour on-demand instance covers most claims teams' volume; a two-node spark-2x cluster at $1.79/hour is worth considering if claim volume and document length both run high, for example a health insurer processing imaging reports alongside medical claim narratives.