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 (Verisk newsroom, May 5, 2026). No P&C rate filing documentation standard governs LLM-mediated data access. The NAIC Third-Party Data and Models Working Group, advancing toward a Q3 2026 public comment period, explicitly flags the general-purpose LLM intermediary as an unresolved classification question in its draft framework (NAIC working group materials, March 2026).

Portal to Conversation: What the Architecture Shift Creates for a Rate Filing

When a pricing actuary pulls ISO Indications through Verisk’s portal today, the transaction produces a specific, versioned output: a dated edition of ISO loss cost data for a given state and line, with a defined effective date and an explicit record of when the pull occurred. The actuarial memorandum can document exactly what was retrieved, from which edition, and on which date. The documentation chain is deterministic by design, because the underlying database access is deterministic.

The MCP connector changes the interface, not the underlying data. ISO Indications still comes from Verisk’s statistical database; the connector is a structured conduit, using JSON-RPC 2.0 transport, between Claude and Verisk’s systems. Verisk CEO Lee Shavel framed this architecture at the announcement: “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 newsroom, May 5, 2026.)

Data integrity is partially protected by the architecture itself. Because MCP defines a structured client-server relationship, the data the Verisk connector returns comes from Verisk’s systems directly, not from Claude’s parametric memory. The LLM processes that data in constructing its response; it does not fabricate the loss cost figures. This distinction is real and materially different from a general-purpose LLM being asked to recall ISO loss costs from training data. The connector is retrieving; Claude is interpreting.

The interpretation layer is the problem. A pricing actuary asking Claude for the current ISO commercial auto bodily injury loss cost trend in Texas, and how it compares to the prior year, triggers two simultaneous processes. Verisk’s connector retrieves the deterministic underlying data. Claude synthesizes a comparison, formats a response, and returns an answer shaped by the specific prompt, the conversation context, and the model version in use. Run the same question in a different session, with a subtly different prompt or against a different Claude model version, and the synthesis will differ even when the Verisk data has not changed. For rate filings, where methodology must be documentable and reproducible, that non-determinism creates a documentation category that no current P&C filing instruction was written to capture.

Verisk estimates that integrating this workflow can save hundreds of hours per carrier per year (Verisk newsroom, May 5, 2026). Those savings are real. They do not resolve the documentation gap they also create.

What the Actuarial Memorandum Is Expected to Document Today

From tracking how P&C actuaries document third-party data sources across state rate filings, a standard actuarial memorandum addresses third-party data in a section that covers the source name, the edition or version of the data used, the date of retrieval, and the specific outputs incorporated into the analysis. For ISO Indications accessed through a portal, this documentation is straightforward: the actuary specifies which ISO Indications edition was current as of the filing date, documents the pull with a dated export, and records the specific loss cost factors incorporated into the indication calculation. The chain is transparent and independently reproducible.

LLM-mediated queries add two documentation inputs that have no home in current filing memorandum formats.

The first is the prompt record. The actuary’s question to Claude is a material input to the output. “What is the ISO loss cost indication for Texas commercial auto?” and “What is the ISO loss cost trend for Texas commercial auto bodily injury over the past three years, including the direction of the indication compared to the prior year?” produce different responses from the same underlying data, not because the ISO figures differ, but because the synthesis paths differ. Standard memorandum formats assume database queries with deterministic outputs. They have no field for prompt text as a methodology input, because prior to conversational AI access, prompt design was not part of an actuarial methodology.

The second is the model version record. Claude’s model version is a material variable in the output. A query executed on claude-sonnet-4-5 may produce a different synthesis than the identical query on claude-sonnet-4-6, even when the Verisk connector returns the same underlying ISO data. Rate filing documentation must account for which model version produced the interpretive output; no state filing instruction currently requires this. The data vintage, which actuaries have always documented, now has a companion variable: the model vintage that processed it.

The data vintage question is itself sharpened by conversational access. With portal-based retrieval, the actuary explicitly sees which edition of ISO Indications is current. With LLM-mediated access, the edition being queried through the connector may not surface prominently in the response unless the actuary specifically prompts for it. For a filing where the data vintage is a material disclosure, the conversational interface creates ambiguity that did not exist before.

Two Vendor Registration Questions the NAIC Framework Has Left Open

The NAIC Third-Party Data and Models (H) Working Group is advancing a registration framework covering third-party data and model vendors whose products are used in insurer functions with direct consumer impact, including pricing, underwriting, claims handling, utilization review, marketing, and fraud detection (NAIC Third-Party Data and Models Working Group, March 23, 2026). Registration would require vendors to file a model description and intended use, training data sources and date ranges, documented testing methodology including bias testing where applicable, known limitations, change-management practices, and a regulatory contact. The framework targets a Q3 2026 public comment period, with adoption consideration at the November 2026 Fall National Meeting and first state implementations expected late 2026 or early 2027 (NAIC third-party framework analysis, 2026).

The Working Group’s next scheduled open session is August 12, 2026, at the Summer National Meeting in Columbus, Ohio (NAIC Summer 2026 tentative agenda). That session is the last public review point before the comment period opens, and it is where the LLM intermediary question is most likely to receive direct attention.

A Verisk-in-Claude rate filing workflow raises two distinct registration questions that the current draft framework does not resolve.

The first is whether Verisk, delivering ISO Indications through an MCP connector rather than a portal, still triggers vendor registration. ISO Indications has always been a third-party data source used in pricing and underwriting decisions with direct consumer impact; the regulatory function of the data has not changed. Under the framework’s stated scope, the delivery mechanism shift from portal to connector should not alter the registration obligation. Verisk is still the vendor. ISO Indications is still the data. The carrier is still using it in rate making.

The second question is harder and explicitly unresolved. The NAIC framework’s working materials identify the “treatment of general-purpose foundation models repurposed for insurance, such as a general-purpose LLM accessed via API to summarize claim files” as an open classification problem (NAIC third-party framework materials, 2026). Claude is precisely that: a general-purpose foundation model accessed through an API-equivalent MCP connector, now used to interpret and synthesize insurance-specific data in workflows that feed rate filings. Under the framework’s logic, Anthropic could be a second registerable vendor in the same workflow. Not because Claude delivers the ISO data, which it does not, but because Claude’s interpretive layer produces the output the actuary uses in the methodology.

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

One workflow, two potential registration triggers, neither definitively answered as of July 2026. The framework’s note that registration “does not create a safe harbor” compounds the picture: even if both Verisk and Anthropic eventually register, carriers retain full accountability for AI-mediated insurance decisions under the 2023 NAIC AI Bulletin. Registration establishes regulatory visibility; it does not transfer responsibility to the vendor. The appointed actuary signing the rate filing remains responsible for the methodology, including the LLM-mediated data synthesis steps within it.

The Data Moat Question: Does LLM Delivery Make ISO Data More or Less Sticky?

Embedding ISO Indications in Claude creates a tension in Verisk’s competitive positioning worth watching over Verisk’s multi-year carrier contracts, which CEO Lee Shavel noted average “approximately between 4 and 5 years” on the Q1 2026 earnings call. On one dimension, MCP delivery deepens workflow integration. Filing actuaries who develop prompt libraries, build standard ISO query templates, and establish institutional knowledge around Verisk-in-Claude analysis will not readily switch data vendors. The switching cost is no longer just the cost of replacing the data source; it is the cost of replacing the entire documented workflow built around the conversational interface.

On another dimension, LLM-mediated access potentially commoditizes the access experience in a way portal access did not. When ISO Indications are accessible through a conversational Claude interface, a carrier evaluating an alternative loss cost source can run both through the same analytical environment and compare outputs without switching analytical tools. The competitive differentiation shifts from interface quality toward data quality and coverage. Verisk’s statistical database of 34.5 billion records spanning more than 50 years is genuinely difficult to replicate. That underlying data asset is the durable moat; the conversational interface is the access layer, which any vendor can now build.

The outcome depends on which barrier has historically mattered more in renewal negotiations. In markets where ISO’s coverage and statistical credibility are the carrier’s primary reason for subscribing, MCP delivery makes the data more accessible without weakening the moat. In markets where the interface and workflow integration have been the stickier factor, LLM-mediated access shifts the competition to the layer where Verisk’s advantages are actually weakest.

Documentation Steps for the Filing Actuary, Starting Now

The NAIC framework timeline, Q3 2026 comment with state implementation in late 2026 at earliest, means carriers already using Verisk-in-Claude for rate filing analysis are operating in an ungoverned period. The profession has not yet established documentation standards for LLM-mediated actuarial data access. That gap is not a reason to avoid the workflow; it is a reason to build a defensible documentation practice before regulation specifies one.

The filing actuary using Verisk’s ISO Indications through Claude should capture five categories of information for each query incorporated into a rate filing.

Documentation Log for Verisk ISO Indications Queries Through Claude

Documentation category	What to capture	Why it matters for filings
Data record	ISO Indications edition, state, line, coverage, query date	Establishes the deterministic data vintage, mirrors existing memorandum standards
Model record	Claude model version (e.g., claude-sonnet-4-6), available from API response headers	Identifies the interpretive layer; a different model version may produce different synthesis from identical data
Prompt record	Verbatim prompt text submitted to Claude in the session	The prompt is a material methodology input with no equivalent in portal-based documentation
Output record	Claude’s verbatim response, preserved before any editing	Creates an auditable record of what the LLM returned, distinct from the actuary’s independent judgment applied afterward
Verification record	Independent confirmation of key figures against Verisk’s portal or raw connector output	Separates deterministic data retrieval from LLM synthesis; establishes that filing figures trace to the authoritative source, not to Claude’s interpretation of it

The verification step is the most important. The MCP connector retrieves authoritative Verisk data; Claude synthesizes that data into a response. For numbers incorporated in a rate filing, the actuary should independently confirm that the figures match what Verisk’s systems actually returned, not solely what Claude’s synthesis produced. This distinction, between LLM-retrieved data and LLM-interpreted data, is where professional responsibility concentrates. The actuarial judgment in the filing belongs to the appointed actuary. The documentation must show it.

This is not a complex technical requirement. It is a documentation discipline that most actuarial teams have not yet developed because the workflow arrived in May 2026 and the profession has not had time to respond. The filing actuaries who build this practice now will have a defensible compliance posture before state regulators have issued formal guidance. Those who wait for the NAIC’s framework to resolve the question will be building documentation practices under regulatory pressure rather than ahead of it.

What This Means for Actuarial Practice

The governance pattern here extends beyond this one Verisk product. MCP connectors are standardizing how industry data flows into foundation models, and ISO Indications is among the earliest regulatory-grade actuarial datasets to make that transition. The documentation and compliance questions the Verisk-Claude workflow raises will recur with every actuarial data source that eventually flows through a conversational AI interface: catastrophe model outputs, experience data, credibility tables, trend selections.

Appointed actuaries signing P&C rate filings that incorporate LLM-mediated data access carry professional responsibility for the methodology underlying those filings. The non-determinism of LLM synthesis is not disqualifying; expert judgment has always been part of actuarial practice. But it is a new input category that requires documentation standards the profession has not established. The NAIC’s advancing vendor registration framework will resolve the vendor-side question eventually, probably by late 2027 at the pace of rulemaking. The actuary-side question, what the appointed actuary must document in the memorandum when the data analysis runs through an LLM, falls squarely within professional standards territory and is not something the NAIC vendor registry will answer. That answer will have to come from the CAS, AAA, or SOA, and it has not been requested yet.

Carriers that build documentation disciplines now, before the regulatory framework crystallizes, will be in a structurally better position when regulators begin reviewing rate filings that incorporate LLM-mediated analytics. The alternative is discovering the documentation gap during a rate filing review rather than before one.