MCP for Insurance: Underwriting Data, Claims Estimates and the NAIC AI Bulletin

MCP lets an AI assistant pull loss cost data, property risk signals and restoration line items from licensed insurance data providers, inside one conversation. As of November 2026, Verisk publishes several MCP connectors and Cotality offers property and climate risk data over MCP. Insurance regulators have been clear that existing insurance law applies to AI systems, including ones supplied by vendors.

MCP (Model Context Protocol) is an open standard for connecting AI apps to other software through MCP servers. New to it? Read What Is MCP? first.

This guide covers what MCP can do for carriers, agencies and claims teams, which servers are documented, what the NAIC model bulletin expects and a starter stack. Facts were checked on 2 November 2026. It is not legal or regulatory advice: your compliance lead and counsel decide how state rules apply to you.

What MCP can do in insurance

Most early uses are research, preparation and drafting:

• Underwriting research. Look up loss cost activity or indications for a line of business and jurisdiction you are licensed to see.
• Property risk. Pull roof age, climate risk signals or hail history for a property under review.
• Restoration estimates. Build a first-draft estimate from a plain-language description of the damage, for an estimator to check.
• Policy questions. Answer staff questions from underwriting guidelines and procedures through a read-only server.
• Agency work. Summarize a client's file and draft renewal reminders in the agency's CRM.

Each of these saves look-up and data entry time. None should decide a policy or a claim on its own.

Which insurance MCP servers exist

Verisk

Verisk's 5 May 2026 announcement says its insurance analytics "are now available in Claude" through Verisk MCP connectors. It says access is "governed by Verisk's established data governance framework to support customers' controlled access and compliance needs."

Two connectors are listed in Claude's directory:

• Verisk Underwriting Intelligence. The directory listing shows five tools, including iso_losscost_activity_get, iso_losscost_experienceindex_get, iso_losscost_indications_get and iso_premium_audit_advisory_get. It is for "authorized customers" and requires sign-in.
• Verisk XactRestore. The directory listing shows 18 tools. Ten search or read data. Eight write, including CreateRemodRestorationProject, AddConstructionLineItemsToEstimate, UpdateConstructionEstimateLineItems and ArchiveConstructionEstimateLineItems.

XactRestore's write tools are the reason to set approvals. A line item added in error changes an estimate that a carrier or homeowner may rely on.

Cotality

Cotality's Claude connector, launched 12 October 2026, targets mortgage, real estate, insurance and capital markets users. The release lists property characteristics, market trends, climate risk signals (including roof age) and valuation data, subject to user licensing. It says the server "enforces strict authentication," so users reach only licensed data.

Core policy and claims systems

We did not find MCP documentation from the large policy administration or claims platform vendors that we could open and verify on 2 November 2026. Third-party marketplaces list wrappers around some of those systems. Treat them as untrusted until your security team reviews them, because they would sit next to policyholder data. Our server risk checklist helps structure that review.

CRM and documents

Agencies often run on general CRMs and file stores. Salesforce and Box both publish MCP servers, which we profile in our MCP Guide.

The NAIC model bulletin and state rules

The NAIC adopted its model bulletin, "Use of Artificial Intelligence Systems by Insurers," on 4 December 2023. Its implementation map, with status as of 1 April 2026, lists 24 states plus the District of Columbia as adopters (Nebraska through a guidance document). It separately lists insurance-specific AI rules or guidance in California, Colorado, New York and Texas.

The NAIC's March 2026 issue brief summarizes the position plainly:

• AI "does not alter insurers' legal obligations."
• "Existing state insurance laws apply regardless of whether decisions are made by humans, algorithms, or third-party vendors."
• The bulletin establishes "expectations for governance, documentation, testing, and third-party oversight."

For MCP, three practical points follow:

1. Vendors are in scope. A connector from a data vendor is a third-party AI input. Include it in your vendor oversight.
2. Document what the AI touches. Keep a list of connected servers, their tools and which decisions they feed.
3. Test before relying. Check AI-assisted outputs against your current process before using them in underwriting or claims.

Your state may have more specific rules. Colorado's regulation, for example, is listed separately on the NAIC map. Ask your compliance team which apply.

Security points for insurers

• Policyholder data. Claims files contain health, financial and contact details. Do not connect claims systems in phase one.
• Untrusted documents. Claim submissions, photos and contractor estimates come from outside. Hidden instructions in a document can try to steer an AI. Read Prompt Injection Through MCP Tool Results.
• Logging. Keep a record of which tools ran and on whose behalf. See MCP Audit Logging.

Example prompts

• "Show the latest loss cost activity for this line of business in this state, and note which filing it comes from."
• "Pull roof age, hail history and climate risk signals for this address for my underwriting notes."
• "Start a draft restoration estimate for a 12 by 14 foot bedroom with water damage to drywall and carpet. Do not save anything until I confirm."
• "Summarize our underwriting guideline for wood-burning stoves in homeowners policies."

The third prompt names the approval step out loud. Clients that show confirmation prompts will ask anyway, but stating it sets the expectation.

Questions to ask a data vendor

1. Does our current contract cover AI use of your data?
2. Which tools can write, and can we turn them off?
3. Is every tool call logged, and can we export the log for an examination?
4. How do you test that AI-surfaced results match your standard products?

These questions also feed the vendor oversight documentation the NAIC bulletin expects.

A starter stack

Use our tool permission planner to sort tools into allow, ask or block.

What to keep human

• Underwriting and pricing decisions.
• Coverage determinations and claim payments or denials.
• Rate and form filings.
• Fraud referrals.
• Communications that tell a policyholder what is or is not covered.
• Approval of every estimate change.

The MCP specification says applications "SHOULD" present confirmation prompts for operations so a human stays in the loop, per the 2026-07-28 tools specification.

Next step

Our MCP readiness audit reviews your data vendors, systems and governance documents before you connect an AI. If you need a custom read-only server over underwriting guidelines or internal procedures, see MCP development. For broader governance, read MCP for IT Leaders.

Frequently asked questions