// Line of business

The hard part is different in every line.

Insurance AI does not fail generically. It fails in ways specific to the line: the schedule that reads as authoritative when it was never right, the severity set in a courtroom, the form that is bespoke by design, the loss with no historical analog. Start where your documents are.

Each page below is written in the artifacts and the formats, not in use cases. If the detail is wrong, you will know inside a paragraph.

// The situation

The line decides the build. The state, what ships.

Roughly half the states have adopted the NAIC AI Model Bulletin, and New York’s Department of Financial Services states the position plainly: an insurer cannot rely solely on a third party’s claim of non-discrimination. Colorado’s expanded Regulation 10-1-1 compliance deadline passed on 1 July 2026, and the NAIC’s AI Systems Evaluation Tool is in pilot with adoption expected at the 2026 Fall National Meeting.

NAIC and state bulletins · current at September 2026.

Regulatory responsibility is non-delegable. Buying a product does not move the obligation, which means the governance file is part of the build rather than a document somebody writes afterward. Every page below names the surface it has to satisfy.

// We ran one of these

We spent a decade building and operating an AI-native carrier — underwriting, claims, servicing, distribution, the regulators and a P&L we had to answer for — which is why these pages are written in schedules, loss runs and filings rather than in capabilities.

The operating record →

// What's hard about this

Two things are true in all five.

The fact that decides the file is usually not in the document you were given.

Roof geometry is missing from the schedule. The reason the admitted market declined the risk was communicated by phone. The MFA answer is technically true and operationally meaningless. In each case the extraction is easy and the decision still is not, because the governing input was never written down. So the build reports its own confidence: every value arrives labeled extracted, inferred or defaulted, and the labels are data rather than a footnote. A system that normalizes a document without reporting how much of it was inferred has made the file harder to reason about, not easier.

Every model in these lines is trained on a series that has moved.

Auto severity, cyber aggregation and secondary-peril frequency are all structurally non-stationary, and a model fitted to the old series prices to a trend that keeps moving. The answer is not a better fit. It is to build where the historical series still holds — extraction, reconciliation, frequency, compliance assembly — and to instrument the moving part separately, so the thing that changed is visible as a measurement rather than absorbed silently into a score.

// What ships with it

The governance file is scoped to the line, not the vendor.

Every build leaves with a model inventory entry, data lineage, pre-deployment testing results, drift thresholds with remediation triggers, and a named human decision-maker specification. What varies is the surface it has to satisfy: outcomes testing in Colorado, an annual compliance certification attested by a named officer in Connecticut, wording that says eliminate rather than mitigate the risk in Virginia.

State AI bulletins and regulations · current at September 2026.

Bring us one workflow.

We will tell you where the work actually goes on that line, and which part is worth building first.