I've handed ad account access to a lot of people over the years. Freelancers, agencies, a nephew who "knows Facebook." The rule I landed on after a few expensive lessons is simple. Nobody touches the budget until they've shown me they can read the account.
Now there's a new applicant, and it works around the clock.
On September 29, Meta launched Muse for Small Business, a version of its Muse agent that connects to your Meta ad accounts, your Facebook Pages and your Instagram analytics, plus a stack of outside tools like Shopify, QuickBooks, Stripe, Klaviyo and HighLevel. Meta's pitch is that an agent with context on your whole operation beats a chatbot you have to brief from scratch every time.
It's free for most of what you'd do with it, and it lives inside the company that sells you the ads. Think about both of those before you click connect. Below is my setup and my test plan.
What it actually is
Muse launched as a personal agent on September 8. Each user gets an agent running on its own virtual machine in Meta's cloud, with a browser you can watch while it works. A permissions system Meta calls Sentinel decides which services it can reach and separates reading from writing.
The small business version adds skills and connectors on top of that same agent. Per TechCrunch and Engadget, here's what plugs in:
Meta's own stuff. Instagram professional account analytics, Facebook Pages and Meta ad accounts.
Fifteen named outside tools. Asana, Box, Canva, Dropbox, Figma, Granola, HighLevel, Intuit QuickBooks, Klaviyo, Lovable, Notion, Shopify, Slack, Stripe and Zoom.
Custom connectors for services Meta doesn't support yet. Muse can build these for you, but per eesel's breakdown of Meta's help pages, Meta doesn't review custom connectors or how they use your data. Remember that one. It comes up again below.
Meta's own examples are the kind of thing an owner would actually ask for. Analyze sales, campaigns and social activity and draft a growth plan. Look at what's performing and draft a week of posts plus a new campaign for review. Review monthly financials for odd expenses. Flag the emails that need a reply and draft them.
What's missing at launch matters too. Parix's review notes no HubSpot, Salesforce or Google Ads connectors. So if half your spend is on Google, Muse only sees half the picture, and it'll still talk like it sees all of it.
What it costs
The agent itself is cheap. Muse is free with a weekly usage limit, and paid plans run $20 or $100 a month for more usage. Meta's subscription page, as summarized by eesel and Parix, names the tiers Power at $20 and Maximum at $100, with weekly token allowances instead of task counts. Meta doesn't publish the free tier's limit.
A few things the price tag doesn't tell you:
It's one person, one account. There's no documented team plan, shared workspace or admin roles. If you have an assistant and a media buyer, each of them is on their own.
US and Canada only, ages 18 and up, for now.
Training is on by default. The "Help improve AI models" setting starts switched on and covers what flows through your connectors. Turn it off in Data Controls on day one. (Meta's coding product has the same default, which I griped about a few weeks back.)
You own what it does. Meta's terms put responsibility for the agent's actions on you. Same as an employee, minus the handbook.
The approval settings matter more than the model
Meta's headline promise is that nothing gets published, sent or spent without your approval. Out of the box, that's true. It's also easy to loosen without realizing it.
Per Meta's help center, as quoted in this breakdown of the permission defaults, approvals come in up to five flavors: Allow once, Allow for this task, Allow for this site, Always allow, and Deny. Each connector can sit on "Ask for some actions" or "Always ask."
You may remember the Marketplace seller in September whose agent handed his pickup address to a buyer after he tapped the wrong flavor of yes. Meta said no privacy controls were breached, and technically that's right. The controls did exactly what he told them to. He just tapped the wrong button.
Here's the setup I'd use for the first two weeks:
Meta ad accounts on "Always ask." Every action asks. It gets annoying, and I'd live with it.
Facebook Pages and Instagram on "Always ask." Nothing posts, replies or DMs without you.
Shopify, QuickBooks and Stripe on "Always ask," and connect them only when you get to the margin test below.
Never tap Always allow on anything that sends a message, publishes content or moves money. Allow once is the only button you need for those.
Leave Slack, Notion and your inbox disconnected until the ad tests are done. One variable at a time.
Turn off model training before you connect anything.
One more thing I couldn't confirm from any source I trust: exactly which Ads Manager actions Muse can take on its own versus only draft. Meta's material says it can analyze campaigns and draft new ones for review. Nobody I read documents whether it can edit a live budget, pause an ad set or duplicate a winner. Until someone documents it, assume it can do more than you think, and make it ask.
And set an account spending limit in Ads Manager before you connect. That limit holds no matter what any agent decides to do.
The five jobs I'd test first
All five are reading jobs, on purpose. If it can't read the account as well as a decent junior buyer, it doesn't get to make changes.
Job one: the account audit
This is the fastest way to find out whether it understands your account or just sounds like it does. Ask the questions you already know the answers to.
Look at my Meta ad account for the last 30 days. Don't change anything. 1. List every active campaign with spend, results, cost per result and frequency. 2. Tell me which three ad sets spent the most and what each one got for the money. 3. Flag any ad set where cost per result over the last 3 days is more than 30 percent worse than its 7 day average. 4. Flag any ad with frequency above 3. 5. For every number you give me, tell me the exact date range you pulled. If you can't see something, say so. Don't estimate. |
Then open Ads Manager and check three numbers by hand. If any of them are off, stop there and run it again tomorrow. If they're off twice, you have your verdict on reporting.
Job two: the creative fatigue check
This is tedious, pattern based work, which is exactly what I'd want an agent doing.
For each active ad in [CAMPAIGN NAME], compare the last 3 days, last 7 days and the full lifetime on CTR, cost per result and frequency. Sort them into three groups: still working, slipping, and tired. For each tired ad, tell me which metric moved first and on what date. Then draft three new hooks for the best performing ad in the same voice. Draft only. Don't create or publish anything. |
Notice the 1, 3 and 7 day windows. That's how I read an account, and it's how you should make it read yours. Budgets stay with a human. I don't let Meta's automated rules touch them and I'm not about to let Meta's agent do it either.
Job three: comment and DM triage
Your Page comments are half sales leads and half chaos. Have it sort, not answer.
Pull every comment and DM from the last 7 days on my Facebook Page and Instagram account. Sort them into: buying questions, complaints, spam, and everything else. For buying questions, draft a reply in my voice, under 40 words, ending with a question. For complaints, draft nothing. Just list them with a link so I can handle them myself. Show me the drafts. Send nothing. |
Quick note so you don't mix them up. Muse is your agent, not your customer service bot. Answering customers in WhatsApp, Messenger and Instagram Direct is a separate Meta product called Business Agent. Muse drafts. You send.
FROM THE AI NEWSROOM The Muse Test Drive Kit The exact permission settings I'd start with, twelve copy and paste prompts for your ad account and storefront, a seven day scoring sheet, the red flags that mean you pull access, and an exit checklist for when you're done. Reply MUSE and I'll send it over. |
Job four: the Monday report pull
If you're paying someone, or yourself, an hour every Monday to screenshot Ads Manager into a slide, this is the test that pays for the experiment.
Every Monday, build me a one page report for last week from my Meta ad account. Top line: total spend, total results, cost per result, and the change versus the prior week. Then: the best ad and the worst ad by cost per result, with one sentence on why you think so. Then: anything that spent more than [DOLLAR AMOUNT] with zero results. Plain sentences. No adjectives. If a number looks wrong to you, say so instead of smoothing it over. |
Compare its version to your own for two weeks. If they match, you've bought back an hour. If you'd rather have that report land in a sheet with no AI interpretation at all, a Make.com scenario can pull the same numbers on a schedule and drop them into a Google Sheet. It won't be clever about it, but you'll get the same thing every Monday.
Job five: the margin view
This is the one I'm most curious about, because it's the question Ads Manager can't answer. Connect Shopify (and Stripe, if that's where your money lands) only for this test.
Using my Shopify orders and my Meta ad account for the last 30 days: 1. Total revenue, total ad spend, and revenue divided by ad spend. 2. The top five products by revenue, and which campaigns most likely drove them based on timing and landing pages. 3. Any product getting ad spend whose sales didn't move. Mark every number as either pulled directly from a system or inferred by you. Don't blend the two. |
The last instruction matters most. Matching ad spend to store sales is guesswork for any tool or any person. A guess is fine with me as long as it's labeled as one.
The seven day test
Start to finish, it looks like this. One owner, one ad account, about 20 minutes a day.
Day one. Set up. Turn off training, set every connector to Always ask, set an account spending limit in Ads Manager, connect only the ad account and your Page.
Day two. Audit. Run job one. Check three numbers by hand.
Day three. Fatigue. Run job two. Compare its tired list to your own gut.
Day four. Triage. Run job three. Count how many drafts you'd send without editing.
Day five. Report. Run job four. Time how long your own version takes.
Day six. Margin. Connect Shopify. Run job five. Disconnect it if you don't like what you see.
Day seven. Score it. Each job gets a 0, 1 or 2. Zero means wrong or useless. One means useful after edits. Two means you'd trust it next week.
Seven or more out of ten and it earns a second week with the same permissions. Less than that, disconnect and check back in a quarter. These things get better fast, but you don't owe it a retainer while it learns.
What it doesn't earn in week two, under any score: budget changes, publishing without approval, or Always allow on anything that talks to customers.
FROM THE AI NEWSROOM The AI Business Accelerator For operators who would rather build it with someone than build it alone. The full stack, installed alongside you, with the decisions made in the room instead of guessed at. Ninety seven dollars. Join the Accelerator for $97 |
Muse versus the tools you already have
Against Go High Level. If your leads already live in Go High Level, Muse can connect to it, which is the interesting part. HighLevel holds the pipeline, Meta holds the ads, and Muse is the first thing I've seen that reads both in one conversation without you building anything. For running conversations with leads, I'd still keep that inside HighLevel, where your workflows and history live.
Against Make.com. Make.com does the boring version of half these jobs with no AI judgment at all: pull numbers, drop them in a sheet, ping Slack when a cost per result crosses a line. It's slower to build and much easier to trust. My split: Make for anything that has to run the same way every time, Muse for questions you'd ask a smart person once.
Against doing it yourself. The honest benchmark is you, in Ads Manager, on a Monday morning with coffee. If Muse isn't faster than you and just as accurate, I wouldn't bother with it.
Whatever you test, record the calls where you review numbers with your team or clients in Fathom. When the agent's report and the human conversation disagree, you'll want both on file.
Where it falls short
It grades the house's homework. Meta built the agent, sells the ads and keeps the data. Meta says Muse conversations and the agent's machine aren't shared with its ad systems. I'll take them at their word and still assume no advice from Muse will ever tell you to spend less on Meta.
The ad permissions aren't documented. I couldn't find a clear list of what it can change in a live account versus only draft. Until that list exists, Always ask stays on.
One seat, no admin. Agencies and anyone with a team will hit this wall immediately.
It only sees Meta's half of your ads. No Google Ads connector at launch, so cross channel questions come back half answered.
Custom connectors are on you. Meta doesn't review them. Don't build one into anything holding customer data.
The web isn't always open to it. Amazon blocked Muse from its store on September 21, saying the agent didn't identify itself. If your workflow depends on an agent visiting someone else's site, that site gets a vote.
My verdict: worth a week, on read only, starting today. The account audit and the creative fatigue check are where I'd expect real value, and I'd start on the free tier and only pay if the weekly limit actually gets in the way. Anything with a send button stays behind Always ask, and you keep your hands on the budget.
This week
Turn off model training in Muse before you connect a single account.
Set an account spending limit in Ads Manager, then connect only your ad account and your Page, both on Always ask.
Run the account audit prompt and check three numbers against Ads Manager by hand.
Run the creative fatigue check on your biggest campaign and compare its tired list to yours.
Score each job 0 to 2 on day seven and decide whether it earns a second week.
Jordan
The AI Newsroom | Practical AI for people with a business to run.

