Blog · 21 October 2026 · Espen Brathaug
Open to AI agents doesn't mean open to everything
In September Amazon started shutting Meta's new AI assistant Muse out of Amazon.com. Now Walmart says customers can connect their own agents to its online shop, with spending limits and approvals. Those are two very different answers. I have been thinking a lot about what this means for an ordinary business.
An AI agent is an assistant that also does things for you. It can open web pages, compare prices, fill in a form or put products in a cart. Greg Isenberg recently called it «a customer with a credit card and no eyes». It cares little about design, and a lot about price, delivery, opening hours and what is actually in stock.
Amazon has its reasons. According to Amazon, Muse doesn't say clearly that it is an agent, and it can see the customer's account. Those are real concerns. But for most businesses the risk is the opposite. They want to be found, and shut agents out without knowing it.
Many block by accident
Many websites have bot protection, for example from Cloudflare, that stops automated visits. That makes sense against spam and scraping. But the same protection can stop ChatGPT or Claude when they fetch your page because a customer asked. Then it doesn't help much that robots.txt lets them in.
So the AI check now tells you when a site looks like it has such protection. It says openly who it is and never pretends to be another bot.
Three steps that make it simpler
What helps me think about it is splitting what an agent does into three steps:
- Read. Products, prices, opening hours and terms. Almost everyone will allow this, since it is how you get found and cited.
- Prepare. Fill in a request form or put products in the cart. The customer sends or confirms it.
- Binding. Order, book or pay. Here the customer should always confirm, in your own checkout or booking.
You decide where the line goes, and you can say it in plain words. I made a template for rules for AI agents that you can put on your website and in llms.txt. The text alone stops nobody, so the guide also shows what actually enforces the line: your bot protection, the tools you give agents, and the checks on your own server.
Choices to make yourself
There is no single right answer. Some businesses gain from letting agents all the way into the cart. Others only want to be read and cited, and let the customer call or send an email. Both are fine, as long as it is a choice and not an accident.
One thing I am unsure about. If agents become the main way to reach customers, they may also start charging to recommend someone. That has happened with search and ads before. No business can control that alone, but you stand stronger when the data, the prices and the checkout are your own.
Test it yourself
The simplest advice I have heard is to ask the assistants yourself. Write ten typical customer questions, ask two or three assistants, and note where they stop. I wrote a recipe with a table you can fill in. It takes about an hour, and you learn a lot.
For online shops I also extended the agent checkout guide with what happens when something goes wrong, such as the wrong quantity, products selling out or an order going in twice.
VegvisAI is a small open source project I run in my spare time. Everything above is free, and nobody can pay for a better place in the guide. If you have tried this on your own website, or see something I have missed, I would love to hear from you at contact@vegvis.ai.
Sources: TechRepublic on Amazon and Muse, Forbes, 21 September 2026, on Walmart and customers' agents, and Greg Isenberg's video «Amazon and Delta Are Blocking AI Agents. How to Win Anyway.»