The blog

ChatGPT Is Shopping for Insurance. Put Your Agency on the Shelf.

Robert Badoux · Founder, Insurance Agent · October 14, 2026 · 4 min read

Quick question. When your client asks ChatGPT to find them cheaper insurance, who does it call?

Not you. Cool. Love that for us.

So this happened on October 1st. A personal auto agency in Texas opened its quoting up to AI assistants. A driver opens ChatGPT, types "help me find affordable car insurance," and the assistant asks the questions, pulls prices from several carriers, lays them out side by side and walks the driver all the way through buying the policy without ever leaving the chat. That same month, some of the big comparison sites went the other direction and blocked the assistants completely, because they say the robots strip out the limits, the deductibles and the state disclosures. And today a startup launched with ten million dollars to do the agency side of it, reading submissions straight out of the inbox and working the carrier portals for the broker.

So the shopping cart moved again, and this time it moved into a chat window that your client already has open on their phone. Right?

The plumbing behind all of this has a name. It's called MCP, the Model Context Protocol, and it showed up at the end of 2024. Think of it as a universal plug. You put a small program on the internet that says "this is what I can do," and any assistant that speaks MCP, which today includes Claude and ChatGPT, can walk up to it and use it. Fifteen years ago the question was whether your agency showed up on Google. The question now is whether your agency is on the shelf when a robot does the shopping for your client.

In the 1600s a mathematician named Blaise Pascal made a famous bet. If you believe and you're wrong, you lose very little. If you don't believe and you're wrong, you lose everything. I'm not here to settle the theology, but the math fits this perfectly. If AI assistants never become how people buy insurance, you lost one afternoon and a hosting bill of a few bucks a month. If they do, and you're the one who blocked them, you're the agency that refused to be in the phone book. And who wants to be that agency?

Doux It Yourself: Level 1 takes the request, Level 2 gives the price. Under four minutes.

So how do you build it? Honestly, easy peasy. It's about thirty lines, and you don't need a developer to read them, because most of it is your own appetite written in plain English.

  1. Write down what you write. Your niches, your states, what you won't touch. If you can explain your appetite to a new hire, you can explain it to this.
  2. Give it two jobs and only two. Answer "do you insure this kind of business," and take a quote request. No prices, no binding, no promises. Every request lands in front of a licensed human who calls the client back. That answers the exact thing the comparison sites are worried about, and it's what keeps you out of trouble.
  3. Paste this. It's Python, using the free MCP library (pip install mcp). Swap in your own appetite.
from mcp.server.fastmcp import FastMCP
import json, datetime

mcp = FastMCP("Your Agency", host="0.0.0.0", port=8000)

APPETITE = {
    "contractors": "General liability, workers comp, commercial auto and bonds. California only.",
    "restaurants": "Package policies, liquor liability and workers comp.",
}

@mcp.tool()
def what_we_write(industry: str) -> str:
    """What this agency insures for a given industry."""
    return APPETITE.get(industry.lower(), "Not our niche. Ask us who to call.")

@mcp.tool()
def request_quote(business_name: str, industry: str, state: str,
                  annual_revenue: int, email: str, phone: str) -> str:
    """Send a quote request to a licensed agent. A person reviews every request."""
    req = dict(business_name=business_name, industry=industry, state=state,
               annual_revenue=annual_revenue, email=email, phone=phone,
               received=datetime.datetime.now().isoformat())
    with open("quote_requests.jsonl", "a") as f:
        f.write(json.dumps(req) + "\n")
    return "Got it. A licensed agent will call you within one business day."

if __name__ == "__main__":
    mcp.run(transport="streamable-http")
  1. Put it online. Any small host works. You end up with a web address that ends in /mcp, and that address is your shelf.
  2. Plug it in and test it. In Claude, go to Settings, then Connectors, and add a custom connector with your address. ChatGPT takes the same address as a custom connector once you turn on developer mode. Then ask it "do you insure roofers in California?" and watch it answer as your agency.

A couple of guardrails, because I'm still an insurance agent. Number one, this collects names, emails and phone numbers, so treat that file exactly like your website's contact form. Number two, don't wire it into your management system on day one. Let it take requests for a month and see what comes in. Number three, never let it quote a price you can't stand behind, because the comparison sites are right about one thing - a price with no limits and no deductible attached is a lawsuit with extra steps.

I'm not telling you to bet the agency on this. I'm telling you to spend one afternoon on it, because the cost of being early is tiny and the cost of being late is the whole channel. The kid in my office who built our first piece of AI did it with a free account and a bad afternoon. You've got the same tools.

Build it this week, ask ChatGPT about your own agency, and then call me and tell me what it said. We'll grab a coffee. The six-dollar kind.

And if you'd rather somebody build it with you, that's what we do at DouxProcess.

Sources: the Texas agency's launch · the broker startup · apps in ChatGPT · MCP

← All posts