Verisk launched Model Context Protocol connectors on May 5, 2026 that put ISO Indications and XactRestore analytics inside Anthropic's Claude. It is the first time a major insurance data vendor has pushed regulatory-grade content out of its own platform and into a foundation model.
Every comparable vendor move over the past two years wrapped an LLM around the vendor's own interface. This one runs the other way, and the interesting question is what it does to a rate filing.
Key Takeaways
- Two connectors shipped: Verisk Underwriting Intelligence, carrying ISO Indications advisory loss costs across 31 lines of business, and XactRestore for restoration estimating.
- The underlying asset is 34.5 billion statistical records, 8.2 billion commercial lines and 21.5 billion personal lines, accumulated over more than 50 years, which no individual carrier can rebuild.
- Verisk was the only insurance-specific connector in the batch Anthropic launched the same day alongside 10 finance agent templates and eight other financial services connectors.
- MCP returns data from Verisk's systems rather than generating it, so provenance survives the model. Whether a state regulator accepts that distinction in a filing is untested.
- OpenAI appears in roughly 9 of 10 carrier stacks per the IA Capital Group survey, so this gives Anthropic a capability in insurance that market position alone had not bought.
What Actually Shipped
Two connectors, aimed at different users:
- Verisk Underwriting Intelligence. Brings ISO Indications into Claude's conversational interface. ISO Indications is the industry source for advisory loss costs, experience data and regulatory filing signals across 31 lines of business, drawing on a statistical database of 34.5 billion records, 8.2 billion commercial and 21.5 billion personal, built over more than 50 years. Verisk estimates the connector saves "hundreds of hours per carrier per year."
- XactRestore. Converts natural-language descriptions of damage into structured estimating actions priced off Verisk's databases. Verisk puts the saving at 30 minutes to two hours per restoration estimate; at five to 10 estimates a week per estimator, that is 2.5 to 20 hours recovered weekly, with the range driven by job complexity.
The timing places it inside a larger move. Anthropic launched 10 finance agent templates and eight additional MCP data connectors for financial services the same day, including Dun & Bradstreet, Moody's and SS&C Intralinks. Verisk was the only insurance-specific one.
MCP itself is now infrastructure rather than a vendor feature. Anthropic developed it in November 2024 on JSON-RPC 2.0, defining hosts, clients and servers, and donated it to the Linux Foundation's Agentic AI Foundation in December 2025. The ecosystem runs over 10,000 active public servers with SDK downloads above 97 million a month.
Provenance Is the Feature, Not the Time Saving
The hours saved are real and unevenly distributed. Experienced actuaries working efficiently in the Verisk portal gain modestly; the larger gains go to less experienced staff and to cross-functional requests where an underwriter needs actuarial data they would otherwise ask the pricing team for.
The consequential change is what the data is being used for. ISO Indications is not general-purpose reference content. It is regulatory-grade actuarial data that appears in rate filings submitted to state insurance departments, and a pricing actuary citing it has to demonstrate it came from an authoritative source unaltered.
MCP's architecture is what makes that defensible. Because the protocol defines a structured client-server relationship, the data returned by the connector is sourced directly from Verisk's systems rather than produced by the model. The LLM composes a response around the data; it does not author the numbers. Provenance survives the round trip, which is the difference between a citable figure and a plausible one.
That distinction has not been tested where it matters. A state regulator reviewing a commercial auto filing has no established position on whether an LLM-mediated retrieval of an advisory loss cost carries the same evidentiary weight as a portal download, and the NAIC's 12-state AI Evaluation Tool pilot is the first framework that will ask carriers to document how AI-mediated analytics are validated.
Competitively it is the capability Anthropic did not previously have. The IA Capital Group survey of 36 senior carrier technology leaders puts OpenAI in roughly 9 of 10 stacks with production AI up from 37% to 61% in a year. Anthropic's insurance footprint is narrower but placed: Allianz made Claude available to 156,000 employees across 70 countries in January 2026, and AIG runs Claude through Palantir Foundry. Enterprise-wide the picture differs again, with Menlo Ventures putting Anthropic at 32% of enterprise LLM spend against OpenAI's 25%, having moved from 12% and 50% respectively since 2023.
The Buyer Cannot Govern What the Connector Delivers
The constraint is on the carrier side, and Verisk is already meeting it in contract negotiation.
CEO Lee Shavel told the Q1 2026 call the company is "having to spend some more time working our way through these issues," referring to AI governance and compliance requirements extending procurement timelines. That friction is a symptom of the buyer's position rather than the seller's product.
Grant Thornton's 2026 survey found 44% of insurance leaders citing governance and compliance as a primary cause of AI project failure, only 24% very confident of passing an independent governance review within 90 days, and 76% acknowledging they could not evidence their controls on demand. A connector that lowers integration cost does not lower that bar; it moves the work from engineering to governance, which is where the carrier is weakest.
The spending data says the capacity is not there either. Per the Datos Insights forum survey, 70% of carriers spent under $500,000 on AI in the past year and only 8% believe they lead peers. At that level the integration layer is affordable and the validation framework around it is not.
There is a symmetric problem on the vendor's side. Verisk's business model has depended on being the destination insurance professionals visit. Distributing analytics through someone else's interface keeps control of the data and the connector while giving up the workflow context in which the data is consumed, and Shavel has noted that Verisk's contracts run approximately between 4 and 5 years. The revenue base supports the experiment, at $783 million in Q1 revenue up 4% with subscription at 84% of the total and adjusted EBITDA margin of 55.9%. What it does to pricing leverage at the next renewal is the part neither company has priced.
Further Reading on actuary.info
- Verisk Q1 2026: Seven New AI Modules and a Growing Carrier Pipeline - The Q1 module output that set the stage for the MCP connector launch, including augmented underwriting pipeline and aerial imagery revenue growth.
- OpenAI Sits in 90% of Carrier AI Stacks: The Vendor Concentration Risk - IA Capital survey quantifying the OpenAI dominance that Verisk’s Anthropic partnership is positioned to challenge.
- Insurance AI Pivots From Claims to Underwriting: ILTF 2026 Takeaways - Datos Insights survey data on carrier AI production deployments and the build-versus-buy context for vendor analytics.
- Carrier AI Projects Fail at the Audit Layer, Not the Tech - Grant Thornton governance data that frames the compliance challenge when regulatory-grade analytics flow through third-party LLMs.
- Verisk Q1 2026: AI Governance Friction Extends Sales Cycles - How governance contracting requirements are creating systemic deployment bottlenecks across the insurer AI vendor ecosystem.
- ZestyAI Recruits Stephenson to Challenge Verisk in Property Risk AI - The executive who built Verisk into an S&P 500 analytics powerhouse joins an AI-native competitor, signaling a shift in the insurance data market.
- Orchestrated Agent Fleets Become the Carrier AI Standard - How Verisk’s MCP connectors fit into the broader multi-agent orchestration architecture emerging across AIG, Gen Re, and the Linux Foundation protocol stack.
- Big Four Consulting Firms Reshape the Carrier AI Delivery Model - The PwC-Anthropic 30,000-person certification program and $1.6B in consulting-channel capital investment that turn MCP-connected data sources into production underwriting tools at carrier scale.
- How Verisk’s Connector Model Compares to Guidewire and Duck Creek AI - Comparative analysis of three competing P&C embedded AI architectures with monetization, switching cost, and regulatory accountability implications.
- Verisk Adds McKinsey AI Strategist to Board - Pradip Patiath’s board election as a governance signal of Verisk’s competitive positioning against Guidewire, CCC, and EXL in the AI analytics layer.
- ISO Indications Through Claude: The Rate Filing Documentation Gap - Two unresolved NAIC vendor registration questions for a Verisk-in-Claude rate filing workflow, the prompt and model-version documentation gaps in actuarial memoranda, and a five-category compliance log for appointed actuaries.
- Akur8 Discover Turns Competitor Rate Filings Into a Live Pricing Signal - A second vendor pipeline routing competitor SERFF filings into the pricing workflow, and the actuarial certification question it raises alongside Verisk’s.
- XactAI at 7,000 licensees and the token-cost fine print – how the connector strategy monetizes.