ISO Indications Through Claude: The Rate Filing Documentation Gap the NAIC Framework Has Not Closed

Verisk's MCP connectors, launched May 5, 2026, route ISO Indications through Anthropic's Claude: advisory loss costs across 31 lines of business drawn from 34.5 billion statistical records. No property and casualty rate filing documentation standard governs LLM-mediated data access.

The NAIC's Third-Party Data and Models Working Group, heading for a Q3 2026 comment period, already lists the general-purpose LLM intermediary as an unresolved classification question in its draft framework.

Key Takeaways

• 31 lines of business and 34.5 billion statistical records move from a versioned portal pull to a conversational interface, with the connector using JSON-RPC 2.0 transport between Claude and Verisk's systems.
• The data stays deterministic; the synthesis does not. The connector retrieves from Verisk directly rather than from model memory, but the response is shaped by the prompt, the session context and the model version.
• Two new methodology inputs have no field in a filing memorandum: the prompt text and the model version that produced the interpretive output.
• Two possible registration triggers in one workflow. Verisk is plainly in scope; whether a general-purpose foundation model in the same chain is a second registerable vendor is explicitly open.
• Registration creates no safe harbor. Under the 2023 NAIC AI Bulletin the carrier and the appointed actuary keep full accountability whatever the vendors file.

From Portal to Conversation

A portal pull of ISO Indications produces a versioned artifact: a dated edition for a state and line, with an effective date and a record of when the pull happened. The memorandum can name the edition, the date and the factors used, and a reviewer can reproduce it.

The connector changes the interface rather than the data. ISO Indications still comes out of Verisk's statistical database, and the MCP client-server structure means the figures returned come from Verisk's systems rather than from Claude's parametric memory. That distinction is real and it matters: the connector retrieves, the model interprets.

Verisk CEO Lee Shavel framed the design intent at launch: "Trust is the foundation of insurance, and that doesn't change as new technologies emerge. Our role is to bring AI into insurance in a way that reflects the realities of the industry, where data must be authoritative, decisions must be explainable, and accountability remains with people" (Verisk, May 5, 2026).

Verisk estimates the workflow can save hundreds of hours per carrier per year. Those savings are real and they do not resolve the documentation category the same change creates.

The Interpretation Layer Is the Filing Problem

Ask Claude for the current ISO commercial auto bodily injury loss cost trend in Texas and how it compares with the prior year, and two things happen at once. The connector retrieves deterministic data. The model composes a comparison whose shape depends on the prompt, the conversation context and the model version in use.

Run the same question in a fresh session, with a slightly different prompt or against a different model version, and the synthesis differs even where the Verisk data has not moved. For a filing, where methodology has to be documentable and reproducible, that is a new input class rather than a new tool.

It shows up as two fields no memorandum format has. The prompt is a material input: "What is the ISO loss cost indication for Texas commercial auto?" and a question asking for a three-year trend with direction against the prior year draw on the same figures and produce different outputs. Standard formats assume deterministic queries, because before conversational access prompt design was not part of a methodology.

The model version is the second. A query run on one Claude version may synthesize differently from the identical query on the next, with the connector returning the same ISO data both times. The data vintage actuaries have always documented now has a companion variable, and no state filing instruction asks for it.

Conversational access also softens the vintage disclosure itself. With a portal the actuary sees which edition is current. Through the connector, the edition being queried may not surface in the response unless the actuary asks for it, which turns a material disclosure into something that depends on how the question was phrased.

Two Registration Triggers, Neither Resolved

The NAIC Third-Party Data and Models Working Group is advancing a registration framework for vendors whose data or models feed insurer functions with direct consumer impact. Registration would require a model description and intended use, training data sources and date ranges, testing methodology including bias testing where applicable, known limitations, change management practices and a regulatory contact, with a Q3 2026 comment period, adoption consideration at the November 2026 Fall National Meeting and first state implementations late 2026 or early 2027.

Two Vendor Registration Questions the NAIC Third-Party Framework Has Not Resolved

Workflow component	Registration question	Current framework status
Verisk ISO Indications (via MCP connector)	Does the MCP delivery mechanism change the registration obligation for a data source already used in rate making?	Probably no change: data function is unchanged regardless of interface; Verisk likely still registerable
Anthropic Claude (interpretive layer)	Does a general-purpose LLM that synthesizes third-party insurance data for filing-actuary use constitute a registerable model vendor?	Explicitly unresolved: the draft framework names this scenario as an open classification problem

Verisk is the easy half. ISO Indications has always been third-party data used in pricing with direct consumer impact, and moving delivery from portal to connector does not change the vendor, the data, or the use.

The second trigger is the open one. The working materials name the treatment of general-purpose foundation models repurposed for insurance as an unresolved classification problem, and that is exactly what sits in this workflow: a general-purpose model reached through a connector, interpreting insurance data that feeds a rate filing. Anthropic could be a second registerable vendor, not because it delivers the ISO data but because its interpretive layer produces what the actuary uses.

The August 12 session in Columbus is the last public review point before the comment period opens. Even a clean answer there leaves the accountability where it already sits: the framework states that registration does not create a safe harbor, and the 2023 AI Bulletin keeps the carrier responsible for AI-mediated decisions.

That leaves the harder question outside the NAIC's reach. Vendor registration will eventually settle who files what. What an appointed actuary must document in a memorandum when the analysis runs through a model is professional standards territory, and no such standard has been requested of the CAS, the Academy or the SOA. Carriers are using the workflow now, in the gap between the two.