The Underwriting Surface

An insurance market for adaptive-AI errors is forming. The product needs an underwriting surface. The agent run does not present one today.

Carriers writing AI E&O are pricing without loss triangles. Reinsurers are quoting without a forensics path. Both are positions the market will not hold.

The substrate that produces the surface is the same substrate the regulator will ask for. The market and the examiner converge on the same artifact.

What the underwriter used to price.

Underwriting against a software product was a tractable problem. The product had a version. The version was stable for the policy period. Failure modes taxonomized. Logs survived. When a claim came in, the carrier reconstructed the failure, attributed it to a covered cause or an exclusion, and paid or denied.

Loss triangles for tech E&O took decades to assemble. The line matured because the underlying object held still long enough to be measured. The actuarial discipline that priced the line was built on the same assumption the model risk dossier was built on. The artifact does not move while you examine it.

An adaptive-AI deployment breaks that assumption inside the policy period. The system at month one is not the system at month nine. The version reference no longer fixes the behavior under examination. The actuarial input is missing.

What the adaptive run breaks.

An agent run is shaped at execution. The model weights may not have changed. The behavior did. In-context learning, retrieval-augmented context, tool-call permissions revised by an upstream change, agent-to-agent handoffs that did not exist last quarter. Any of these shifts the system without versioning it.

For the underwriter this is not a small problem. Underwriting requires that the insured risk be definable on the day the policy attaches. An adaptive system has no day-one definition that holds across the term. The carrier is writing against a moving target without instruments to measure the motion.

For the reinsurer the problem compounds. The cedant cannot consistently define the risk being transferred. Treaty wording cannot specify what a covered loss actually is, because the system that produces it is not the same system across the cession period. Aggregate covers are written against an exposure base the cedant cannot describe.

The forensics problem.

When a claim is submitted, the carrier needs to reconstruct what the deployed system did. Today the carrier gets production logs. Production logs were built for incident response and ops triage. They were not built for an external party to reconstruct a single agent run six months later for adjudication.

Reconstruction in this state takes months. The engineers who built the workflow are interviewed. Model weights at the moment of decision are often gone, overwritten by subsequent retraining. The tool-call permissions in force on the day of the incident have been revised twice since. The artifact the claim turns on cannot be retrieved.

Subrogation against the AI supplier is the second casualty. The carrier cannot pursue recovery from the system supplier without a record that shows where the failure originated inside the run. The current loss flows entirely to the carrier. The supply chain absorbs none of it.

What the trajectory delivers.

A structured record of the run. Signed at the moment of execution. Hash-chained so tampering is detectable. Replayable against frozen state captured alongside it. Five primitives. The input, the model state, the tool calls invoked, the branches considered and rejected, the output. The full decision graph between the question and the answer.

This is the underwriting surface. Pricing becomes empirical because loss can be attributed. Subrogation becomes possible because failure can be located inside the chain. Treaty wording becomes specifiable because the unit of coverage resolves to the record produced by the run, not to a version string that no longer pins the behavior.

The state regulators are already moving toward this framing from the other side of the market. The NAIC AI Systems Evaluation Tool entered a twelve-state pilot on March 2, 2026 and runs through September. Over half of states have adopted the NAIC Model Bulletin on the use of AI by insurers. The bulletin describes governance expectations for carriers deploying AI. It does not yet require a signed trajectory record for the AI they write coverage on. The actuarial pressure to require one is forming from the underwriting side, not from the regulator.

What changes when the surface exists.

The market grows from where it sits now. AI E&O can be written at scale because the underwriter has measurable input. Reinsurance treaties for AI-error risk can be ceded with wording that resolves under loss. Captives can self-insure adaptive-AI exposure with a defensible reserve calculation. The line graduates from speculation to actuarial.

For the insured firm the surface is leverage in the other direction. A firm deploying adaptive AI that produces a signed trajectory under every run can present a measurable risk profile to its carrier. The premium responds. The exclusions narrow. The recovery path under a claim is preserved.

The substrate that produces the surface also produces what SR 26-2 will eventually require when the interagency RFI closes, and what EU AI Act Article 9 will require when enforcement attaches in December 2027. One artifact, three audiences. The carrier reads loss. The reinsurer reads cession. The regulator reads governance. The substrate is the same.

What we are building.

Wayfinder Systems Group is building the substrate that produces the record the underwriter needs. Every decision signed onto a tamper-evident chain at the moment it happens. Every learning event signed alongside it. Every tool call written into a record that survives the run, the retrain, and the vendor change. The carrier reads loss. The reinsurer reads cession. The regulator reads governance. The substrate produces what each one consumes. We call her Velma.