Can AI agents buy insurance today? We tried 12 sites

AI assistants can now book a table, compare flights and fill a cart. So we asked a simple question: can an AI agent get a car insurance quote for a person in the United States?

In October 2026 we pointed a browser-based AI agent at 12 major US insurer and comparison websites and asked it for one quote for an ordinary driver. It got zero quotes. Not one site returned a price.

This article explains what went wrong, why it is not really the agent's fault, and what a working setup looks like.

What the agent ran into

Every site failed in one of a few predictable ways:

• Bot walls. Several sites blocked the agent outright before it reached a form, or showed a CAPTCHA built to stop automated visitors.
• Long, multi-page forms. Others accepted the visit but asked for many fields across several pages, often with steps that change based on earlier answers. Agents can fill forms, but these flows are brittle and break on small changes.
• "Call us" and contact capture. Some flows asked for a phone number or email before showing any price, then routed the request to a callback or a lead form.
• Identity details up front. A few asked for details that a careful agent should not hand over without a person's say-so.

You can watch the full run in our research video, which also explains the method.

It is not that agents are bad at forms

It is tempting to say "the AI isn't good enough yet". That misses the point. These websites were designed for a person in a browser, and many are deliberately designed to keep automated traffic out, often for good reasons: scraping, fraud and fake leads are real problems.

So we have two goals pulling against each other:

1. People increasingly want an assistant to shop for them.
2. Businesses want to stop bots that scrape, spam or buy things nobody asked for.

Making screen-scraping agents cleverer at getting past bot walls is not the answer. That just escalates the arms race, and it leaves the most important question unanswered: did a real person agree to this purchase?

What a working setup looks like

The fix is to give agents a front door built for them, with a person firmly in charge of the decision. In practice that means three things.

1. A structured way to ask for quotes. Instead of clicking through web pages, the agent sends a request over a standard interface such as the Model Context Protocol (MCP), Agent-to-Agent (A2A) or a plain API. It gets back structured quotes it can compare and explain to the person.

2. A person approves every purchase. Getting a quote is one thing. Buying is another. The approval should happen on a page the agent cannot fake or control, with the person confirming with something only they have, such as a passkey on their own phone or laptop. The agent can ask; only the person can say yes.

3. Proof of what was agreed. After the person approves, everyone should be able to check exactly what was approved: which quote, at what price, on what terms, and when. A signed receipt that can't be quietly edited later protects the person, the business and the agent.

This is not just our idea. Large payment and AI companies have spent the last year publishing standards for agent payments that rest on the same principle of explicit, verifiable user consent, including Google's Agent Payments Protocol, the Agentic Commerce Protocol from Stripe and OpenAI, and agent payment programs from Visa and Mastercard. Insurance and other regulated services need the same thing, with extra care about what the person saw before they agreed.

Why insurance is a good test case

Insurance is a demanding place to get this right:

• Many details feed the price. A quote depends on the driver, the vehicle, where it is kept and more. The person needs to see which details the price was based on, and who supplied each one.
• Terms matter as much as price. Deductibles, exclusions and cooling-off periods all affect what the person is actually agreeing to.
• Consent has legal weight. If something goes wrong, someone will ask what the person was shown and whether they really agreed.

If agent shopping works for insurance, it will work for most everyday services.

What we built

Moltbound is a consent and checkout rail for AI agents. Agents get quotes across 20 US services, including 11 kinds of insurance, over MCP, A2A or an API. Nothing is bought until a person approves on Moltbound's own page with their passkey, and every yes gets a signed receipt that records the exact approval screen they saw.

It is a sandbox today: the partners are fictional, the quotes are simulated and nothing is bought for real. But the approval flow, the receipts and the rules are real, and every rule is tested and published.

Try it

If you have an AI assistant that can use tools, point it at moltbound.dev/llms.txt and ask for a quote. Or follow the prompts on the try page. If you build agents, the agent quick start has everything you need.

We would love to hear what your agent runs into.