Insurance claims: from a claim to a decision your experts can verify

A claim is rarely one question. Handlers move between policy wording, repair estimates, case history and internal systems. A fluent AI answer is not enough if the evidence or the settlement amount is wrong.

The failure mode is specific. A general policy limit can be correctly cited and still produce the wrong decision, because a glass sub limit applies to this particular claim. A missing estimate has to become a request to the claimant, never a confident guess. Any system that prepares claims has to surface the governing evidence and make the proposed decision inspectable.

The governance bar is rising at the same time. The EU AI Act, DORA and EIOPA's expectations of risk based AI governance mean every decision needs a reviewable trail, and claim data has to stay inside the insurer's walls. The question is no longer whether to use AI on claims, but how to run it under control.

Eigen turns the way your claims specialists work into coordinated AI workflows. It gathers case facts, checks policy rules, verifies evidence and prepares a documented recommendation for human approval. Four stages:

1. 01
   Capture the method
   A claims specialist describes how the claim should be handled, in ordinary language. No workflow to draw, no script to write.
2. 02
   Plan the work
   Eigen compiles the method into a plan of dependent and parallel steps and gives each step only the tools and the model it needs.
3. 03
   Execute and verify
   Agents retrieve policy and claim data, check evidence against the source, apply limits and the deductible and calculate the payable amount. Each step records what it read and used.
4. 04
   Prepare for review
   A proposal arrives with source references, the calculation and exception flags. A human releases it. Nothing is paid or communicated by the system.

“Check that the loss falls in the policy period and no exclusion applies. Get the repair estimate; if it is missing, request it. Apply the deductible and any sublimit, check the overall limit, and send anything unusual to a reviewer with the evidence.”

From those few sentences Eigen builds the network: one coordinator and twelve units of work. Retrieve claim, retrieve policy, check policy period, check exclusions, get repair estimate, verify evidence, check prior claims, apply deductible, apply overall limit, calculate payment, draft decision, reviewer handoff. Independent steps run in parallel, and the expert writes zero lines of code.

Every approved method is kept in EigenBrain and reused, so the judgement of your most experienced handlers stays with you when they move on, and the next claim type is faster to build. Claims and corporate KYC share most of their structure: retrieve evidence, verify against the source, apply rules and limits, route exceptions.

The nine scenarios cover a clean settlement, a missing invoice, a coverage sub limit, an estimate that is unavailable and must not be invented, an estimate inside a document, inconsistent history, a loss above the overall limit, an exclusion and a loss before cover began. Clause references follow the published German motor conditions. Every system under test receives the same claims and tools, and a deterministic grader scores the final state. Realistic in its procedures, synthetic in its data.

The modeled economics

On 50,000 claims a year, a single frontier agent costs between 5,290 and 26,870 US dollars a year in AI processing at October 2026 list prices, depending on the model. The model choice alone moves the cost fivefold. Eigen on three small open models on the insurer's own infrastructure is modeled at 17,213 US dollars a year, 66 percent below GPT-6 Astra, with 88 percent less GPU spend than self hosting a large open model.

Against the cheapest API the case is decided by reviewer time. Fourteen seconds of reviewer time saved per claim breaks even, and at 60 US dollars an hour every reviewer minute saved per claim is worth 50,000 US dollars a year. If a 20 minute manual claim becomes a 5 minute review, that is 12,500 expert hours freed a year.

• Every model runs inside the insurer's security boundary on 24 GB GPUs, on premise or in a German cloud.
• No external fallback, outbound traffic blocked, tool access follows your permissions.
• Human approval gates on every decision. Knowledge enters EigenBrain only after approval.

Bring us one claims workflow

We are seeking an insurance pilot partner to evaluate one bounded workflow on historical cases, inside your infrastructure: decision quality, evidence quality, human review effort, completion reliability and cost per usable proposal. Our target is that a 20 minute manual claim becomes a 5 minute review. Results go to you first.