The most important insurance-data launch last week didn’t come with a new portal. It came with a chat box.
On October 28, Verisk said that eligible customers can now get ISO Loss Costs and Premium Audit Advisory Service (PAAS) content inside Microsoft Copilot and Amazon Quick. Inside Claude, Verisk’s Underwriting Intelligence offering now has PAAS alongside the ISO Loss Costs it already carried. Verisk also added a new Benchmark connector that brings in hail and wind history for individual properties. Verisk says customers’ existing entitlements and governance controls still apply.
That can read like a routine integration announcement. It’s actually a shift in where underwriting work gets done, and most insurers haven’t connected it to their own AI governance yet.
What a “connector” actually is
For years, licensed insurance data lived behind the provider’s own portal. An underwriter or premium auditor logged in, searched, copied what they needed, and went back to their workflow.
A connector turns that around. The AI assistant your staff already uses can query the licensed source at the moment someone asks a question. The assistant fetches the relevant content, checks it against that user’s entitlements, and builds it into the answer. The data comes to the work, and nobody has to go to the data.
What changes for underwriters and premium auditors
- Less context switching. A classification question, a loss-cost lookup or a check on a property’s hail history can happen in the same window as the email, submission or spreadsheet being worked on.
- Faster answers to routine questions. PAAS-type questions, such as how to classify an operation or what counts as payroll for a given exposure, are well suited to a conversational interface.
- New failure modes. An assistant can summarize, paraphrase or blend licensed content with other sources. The more fluent the answer sounds, the harder it is to tell which parts came from authoritative data and which parts the model inferred.
The governance question most teams haven’t asked
Most states have adopted some version of the NAIC’s model bulletin on insurers’ use of AI. The bulletins expect insurers to govern AI systems that support decisions affecting consumers, including AI built by third parties. A general-purpose assistant that answers rating and classification questions is now part of that picture, whether or not anyone formally put it there.
Four practical questions:
- Who can see what? Vendor entitlements carry over, but are your own assistant permissions aligned with them? Does a broad Copilot rollout give licensed content to people outside underwriting?
- Can you reconstruct an answer? If a classification decision is later challenged, can you show what the assistant retrieved, which version of the content it used, and what the person did with it?
- Where does the output go? Answers pasted into rating worksheets or audit files become decision records. Treat them that way.
- Who owns the risk? Your vendor contracts and your AI inventory should both state whether these connectors count as third-party AI systems under your governance program.
Data where the work happens, or agents inside the core?
The timing is interesting. In the same week, Guidewire announced a $1.7B deal for the pricing platform hyperexponential and promoted agents that run inside its core platform. These are two competing ideas about where insurance AI will live. One says it belongs inside the system of record. The other says it belongs in the general-purpose assistant people already open every morning.
Most carriers will end up with both. The data and governance layer has to work across both, which is why getting it right now matters.
Questions to ask your data vendors this quarter
- Which of our licensed products are available through AI assistants, and which assistants?
- How do entitlements map to assistant users, and how are changes synced?
- What usage logging do you provide, and can we export it?
- How do you separate your licensed content from content the model generated in a response?
- Do our license terms permit this use, and does using the assistant count toward our usage limits?
The chat box is now part of the underwriting desk, and governance needs to cover it.