AI tools are moving from answering questions to carrying out tasks.
Shopping makes that easy to see.
You can already ask an AI to compare products. That part is familiar.
But what happens when it can also open the stores, work through the options, fill out the checkout, and come back for your approval before it buys?
That is the idea behind Muse, the personal AI agent Meta introduced on September 8.
It has its own apps and website, separate from the Meta AI app many people already know.
Meta says Muse works across the apps you use every day, from a dedicated virtual machine.
It keeps working after you close the app.
It comes back when something changes or when it needs your approval.
Shopping is one of its clearest examples.
What can it actually do?
According to Meta, Muse can open a browser, fill out forms, and negotiate on your behalf.
At checkout, it can pay with Link by Stripe.
Meta says it asks before sensitive actions, and asks for approval on every purchase.
Meta also says it shows a complete audit trail of what it has done and plans to do.
Those are Meta's claims.
I have not tested a purchase with Muse.
"It can shop" tells you what kind of work the product is built to attempt.
It does not tell you how well it handles your budget, your requirements, or your exceptions.
Which stores and payments work today?
That is changing fast, so read the dates.
At launch on September 8, Meta said Shop Pay and 1Password support were coming soon.
At its Connect event, in a September 24 recap, Meta said Muse launched with access to the entire Shopify catalog.
It announced more retailers, including Walmart, Best Buy, Gap, Sephora, and Wayfair.
It also named Shop Pay and PayPal for payments.
I read that as an expansion announcement.
I would not assume every store and payment option is live in every account today.
Who can use it?
Meta says Muse is rolling out in the US on iOS, Android, and muse.ai.
A Mac app followed on September 18.
Meta says it is free for most of what people need, with subscription plans for people who want to do more.
What are early users saying?
The public accounts I found were mixed.
One user posted on September 27 that Muse set hourly restock alerts for a camera and searched local store inventory and prices.
The same person said some tasks took longer with Muse than doing them personally.
That post also includes stock-trading commentary, so I read it as one person's experience, not a neutral review.
Another user posted on September 25 that a daily shopping-scan report they had requested was not arriving.
That post does not explain the setup, so it cannot tell us why.
I found both through a search of public posts on X.
They suggest things to try and things to check.
They do not add up to a verdict on reliability.
I would want to see Muse finish a modest task with my own requirements before trusting it with a recurring one.
What should you actually hand it?
Say you need five headsets for facilitators who run remote workshops.
You have a budget.
They must work with the laptops your team actually uses.
The microphones need to handle noisy rooms.
Replacement ear pads matter, because the headsets will be shared.
"Find me five good headsets" leaves too much to guess.
I would give an agent a clearer job:
"Find three options that meet these requirements. Show the price for five, the warranty, the source for each compatibility claim, and anything you could not verify. Do not place the order. Bring me the comparison first."
Then I would check the shortlist.
Is the price current?
Is it in stock?
Does "compatible" mean tested with our setup, or just a vague product-page promise?
Now the arithmetic.
A model that is $20 cheaper saves $100 across five headsets.
If its microphone fails in a noisy room, you saved $100 on five headsets nobody can use.
The quality of an agent's work depends on the task you give it and the decisions you keep for yourself.
Why should instructional designers care?
I do not see this as only a shopping story.
It is a small, concrete example of a skill people will need at work.
Telling an AI agent what matters.
Checking the evidence it brings back.
Deciding what it is allowed to do next.
If I were designing practice for that skill, I would give someone a realistic request with a hidden problem.
Maybe the cheapest headset is missing one required connector.
The learner would review the agent's recommendation, spot the missing requirement, and decide: approve, revise, or stop.
That gives them a reason to use judgment at the exact moment the software is ready to act.
Meta has published a security explanation of Muse's app permissions and approval prompts.
Read it before you connect work accounts or payment details.
An approval button helps most when you can tell what the agent found, what it skipped, and what happens next.
Write the shopping list before you hand it over.
Sources
Checked September 27, 2026.
- Meta, Introducing Muse, September 8, 2026, for what Muse does, checkout with Link by Stripe, approval, the audit trail, the coming-soon list, the US rollout, and pricing tiers: about.fb.com
- Meta, How We Built Safety Into Muse, September 8, 2026, for permissions and approval on every purchase: research.meta.ai
- Meta, The Biggest News From Connect 2026, September 24, 2026, for the Shopify catalog, the new retailers, and Shop Pay and PayPal: about.fb.com
- TechCrunch, September 18, 2026, for the Mac app: techcrunch.com
- Engadget, September 12, 2026, for Muse having its own apps and website separate from the Meta AI app: engadget.com
- Post on X, September 27, 2026, the restock alerts and slower tasks report: x.com
- Post on X, September 25, 2026, the missing daily report: x.com
Want This Capability on Your L&D Team?
I can build your game from your source material, run a Build Day with your team, or you can start free with my tools.
See how we'd work together Try the tools free