Meta released Muse on September 8, and OpenAI followed with Dots three weeks later. Each is an AI agent with its own cloud computer that keeps working on a task after you close the app, and each can connect to your email, calendar and other accounts. Muse has a free version with a usage cap and a $20 a month plan, and it is offered in the US and Canada. The cheapest way to get Dots is a $100 a month ChatGPT plan, and they are rolling out to Pro users outside the EEA, Switzerland and the UK. We covered how Muse shops for you and what Dots are when each launched; the more useful question now is where each one can actually work.
An agent needs the store's permission as well as yours
An agent that shops does what you would do. It opens a browser on its own computer, goes to the store's website and clicks through. The store sees a visitor it did not invite. Amazon's objection to Muse is that the agent does not announce itself, so Amazon sees an unknown program moving through customer accounts and order histories. Since September a popup tells people who try it that using an unauthorized agent breaks Amazon's conditions of use. Amazon had already moved against shopping agents from OpenAI and Google, so this is a pattern, not a quarrel with one company.
Ryanair shows how this usually ends
What the Ryanair deals changed was the route in. Collecting fares from a web page works until the page changes or the airline blocks it. A direct deal is a door that stays open, and the airline decides what goes through it. An agent clicking through a store's screen is in the same position, and Amazon has pointed to food delivery apps and travel agencies as the model: they sell for a business with that business's agreement.
I expect the large stores to split. Stores that want the order from anyone will build a proper route for agents, and stores that live on ads will keep fighting. If you sell online, this will become your decision too. A customer's agent never looks at your home page, so the shops it can buy from will be the ones whose prices, stock and delivery dates are easy for a program to read.
Start with jobs that only watch
The other permission is yours, and it is worth setting up carefully. Muse asks before it sends an email or makes a purchase, but the default permission setting asks about only some actions, and the setting that lets Meta use your conversations to improve its models is switched on until you turn it off. Dots use permissions and rules you set, and their proactive research is read-only until a later action passes the permission checks.
That points to where to begin. A purchasing manager at a small bakery could ask an agent to open four flour suppliers' price pages every morning and send a short list of any that changed. Doing that by hand takes about 10 minutes a day, which is close to four working hours a month. If the agent gets it wrong, you lose a minute reading a bad list. If it gets a purchase wrong, you lose money.
A watching job also tells you the thing you most need to know, which is whether your suppliers' sites open for the agent at all. You will find out in the first week, before you have trusted it with anything. If all you want is a watcher, the $20 Muse plan is the cheaper way to find out, as long as you are in the US or Canada. Give it a month of watching before you let it send or buy anything for you.