Why Your AI Agent Shouldn't Just Google Its Insurance

Can an AI agent buy insurance using web search? It can find information, but it can’t reliably buy the right policy that way. Search results and training data miss live prices, insurer appetite and current policy terms, and a web checkout was never built for an agent. Agents need a structured, trusted connection to real insurers instead.

Ask a capable AI agent to “sort out insurance for my business” and it will make a reasonable start. It will explain what cover you probably need, suggest a few well-known insurers, and maybe even estimate a price. It sounds convincing. That’s the problem.

Insurance is a regulated product where the details decide whether a claim is paid. The ways a general-purpose agent goes wrong here are predictable, and they’re worth spelling out.

1. Knowledge goes stale

A model’s training data has a cutoff date, and its weights aren’t updated continuously. Insurance moves faster than that. Prices change, insurers widen or narrow which risks they’ll accept, products are launched and withdrawn, and policy wordings are updated.

An agent answering from memory is describing the market as it was, not as it is. Even web search mostly surfaces marketing pages, not the live price and eligibility for your risk.

2. Confident answers aren’t always correct ones

Language models can produce fluent, plausible advice built on generic information. In most areas, a slightly off answer is harmless. In insurance, advice that misses an exclusion or misjudges eligibility can leave someone uninsured without knowing it, and they only find out when they claim.

3. The web wasn’t built for agents

Insurer websites are designed for people: multi-page forms, cookie banners, “call us for a quote” buttons and PDF policy documents. An agent can sometimes push through, but it’s slow and brittle. A layout change breaks it. A CAPTCHA stops it. And a purchase completed by guessing at form fields is not something anyone should trust.

4. It wastes time and tokens

Without a better option, an agent repeats the same work for every user and every insurer: searching, reading, filling in, re-checking. That burns tokens and time on work that could be done once, properly, and reused.

5. It can’t see what really matters

The hardest parts of buying insurance aren’t finding a name and a price. They are:

• knowing which insurers will actually accept this risk
• gathering the right information in the format each insurer needs
• comparing cover limits, excess and exclusions like-for-like
• completing a regulated purchase with the customer’s consent
• helping later, at renewal and when something goes wrong

None of these are solved by better search.

What agents need instead

Agents need a dedicated, trusted insurance layer. Something that works like a careful human broker, but speaks the agent’s language:

• Live quotes from real insurers, based on current underwriting, not memory.
• Structured tools for each step: start an application, find out what’s required, request quotes, compare them, select one and submit it.
• Consistent data, so the agent can compare policies properly instead of parsing a dozen formats.
• Customer control built in, so nothing is bought without the human’s explicit go-ahead.
• Humans on call for complex cases and exceptions.
• Help after the sale, so the same connection handles renewals and claims.

This is what the Model Context Protocol (MCP) makes possible: a standard way for any agent, whether Claude, ChatGPT or a custom build, to call tools like these directly.

What is an insurance MCP? An insurance MCP is a service that exposes insurance tasks (quoting, comparing, buying, servicing) as tools an AI agent can call directly. Instead of scraping websites, the agent connects once and gets structured, up-to-date answers from real insurers.

ZapCover is building this layer, starting with the Customer MCP: structured tools to quote, compare and apply for cover. See the tools or watch an agent use them.