Meta Muse costs nothing to start. What it takes in return is what you know about your clients and how you work, and Meta's own words say where that goes.
You're about to connect Box and Slack to Meta Muse. It's free, and the price is your data. Your work trains Meta's model by default, and the better agent that comes out of it sits on Meta's side.
Meta introduced Muse on 8 September 2026 as a personal AI agent, and put it in the US on iOS, Android and muse.ai. Free, for most of what people need.
On 29 September Meta added Muse for Small Business. CNBC reports it connects the agent to Asana, Zoom, Intuit, Box, Canva and Slack, and links it to Meta ad accounts and to professional Instagram and Facebook profiles. The pricing matches the main app. Free with usage limits, subscription beyond that.
That's the version of Meta Muse for small business I'm writing about. It's an AI agent with a desk of its own.
The desk is called Muse Secure VM, a dedicated cloud computer with its own browser and a file system. Meta says that's how it can browse, fill in forms and pay.
Meta's research post on how it built safety into Muse says something plain. Your conversations, and the tool calls and handoffs that follow them, train new versions of the core model. Meta calls these records trajectories. It calls using them the good default, because every user gets a better agent as people use the product.
There's a switch. You turn training off in Muse settings, and WIRED reports the same, that users can opt out.
A trajectory is the record of what Muse was doing for you. In a business, that record holds what you know about your clients and how you work.
I'm not saying Meta hides any of this. It published it, which is why I can quote it.
Access is the other half. Meta says it restricts access by its own people through operational policies. It also says that does not stop Meta from reaching your data when it needs to run the service. WIRED quotes Meta's David Singleton saying the company is barred by policy. He also says access would still be technically possible.
A Muse Confidential VM, where only you hold the key, is planned for later. Planned.
Neither line means Muse cannot learn. It learns two ways.
It keeps a memory about you inside your own virtual machine. Meta says you can read and edit those files directly.
That memory is yours. Separately, Meta's own post says your work trains the shared model by default.
An agent you own keeps its learning on your side. Muse splits it. The memory stays in your machine, where you can read it, and the training that makes the next version better goes to Meta.
I've written about the other arrangement, where an agent's memory is a file you hold. It's in giving an AI agent a soul file.
There's one more line in that post. When Muse browses the web, it appears as your activity, so what it does there can end up shaping the ads you see.
I sell AI visibility work, so weigh my view with that in mind.
If you try it, I'd hand it something from the should list. That's work you'd never hire a person to do but know you should be doing. No client's name in it.
I call an agent that does a defined job for you a digital employee. I mean it as added capacity for the people you already have. I cover the idea in AI staffing. If you're still choosing between tools, start with what each AI tool is for.
Whatever it does, you answer for it. Same as when someone on your staff gets it wrong.