Meta’s Muse for Small Business has been live since 29 September, and most of the coverage stopped at the connector list: fifteen third-party tools, plus Facebook Pages, Instagram professional analytics and Meta ad accounts, with a hard rule that nothing publishes, sends or spends without the owner’s approval.
The more useful signal came later. Meta sent follow-up use-case material describing what business owners actually type into the agent — the top prompts, in the company’s own words, and the workflows they turn into. Vendors usually publish capabilities; this time Meta published demand. For anyone running client accounts, the pattern shows which deliverables a client can now produce for free, and which ones they still cannot.
The three patterns Meta described
According to Social Media Today’s account of Meta’s use-case notes, the requests cluster into three buckets:
- Assessing ad performance. The single most common request. Meta’s own examples of what owners ask: “Analyze my Meta ads. What are my highest performing ads and why?” and “What are top recommendations to improve my Meta ads performance?”
- Building campaign structure conversationally. The most common prompt here is: “Taking into consideration the attached creative brief, how do you suggest I set up a Meta ads campaign?”
- Chaining connectors into one flow. Meta’s example chains the Ads connector with Shopify, Canva and HighLevel: read the product catalogue, settle the promotional angle, brief the creative, then build the campaign.
Meta also said businesses use Muse for setup chores — catalogue configuration, pixel installation. Keep that detail in mind.
Pattern 1: diagnosis comes before generation
Nobody’s first prompt is “write me five captions.” The first prompt is “tell me what’s happening and why.” That is a meaningful ordering. Generation is now cheap and universal; interpretation is where the anxiety sits, because a wrong read of the numbers costs real budget while a mediocre caption costs nothing.
So if your core deliverable is content production, you are competing with a free tab in the client’s own app. If your core deliverable is a defensible read of the account and the decision that follows it, the same agent becomes free research that arrives at your desk pre-summarised.
There is a catch, and it is the most actionable thing in the story: the agent’s answer is only as good as the measurement layer underneath it. If the pixel double-counts, if naming conventions are inconsistent, if the catalogue is half-synced, Muse will quote the wrong numbers with total confidence — and you will be arguing with a spreadsheet the client got for free. Hours spent fixing tracking are now leverage rather than overhead, because someone else’s software is going to read your work out loud.
Pattern 2: the brief becomes the deliverable
The top structural prompt attaches a creative brief. That is the whole tell. The object the client is being asked to produce is not the campaign; it is the brief that constrains the campaign.
An agency that produces a tight brief — offer, audience, proof, constraints, budget band, what the client must not say — hands over something a free agent can execute against without wrecking brand consistency. That is a clean split: strategy and specification stay billable, execution keeps sliding toward commodity, and the client gets speed without losing the guardrails. Template the brief once, make it a fixed input to every campaign, and you have turned a competitor into a delivery channel.
Pattern 3: chained connectors are a data-model test
Ads plus Shopify plus Canva plus HighLevel only produces a coherent answer when product IDs, naming conventions, conversion events and cost data line up across all four systems. In most small-business accounts, they do not. The vendor demo works because the demo account was built for the demo.
That makes “make this account legible to an agent” a real engagement: SKU naming that survives a join, catalogue sync that does not duplicate variants, deduplicated conversion events, one agreed source of truth for cost of goods. Fixed scope, durable result, and plumbing most owners will never do themselves. The warning cuts both ways: chaining tools multiplies the places a wrong number can enter, and a chain has no error surface — it just returns a confident answer built on a bad input.
Three questions Meta has not answered
The launch post is explicit about approval gates and silent on the parts that matter to agencies. PPC Land’s teardown of the announcement lists the gaps: whether Muse can change budgets, bids, targeting or creative in an ad account, or only read performance; whether an agency or partner with access to a client’s ad account may use Muse at all, and under whose permissions; and what the paid tiers cost.
There is also an incentive question that no vendor can settle for you. When an agent can see both the ad account and the store, its easiest path to “find new customers” is more Meta ads, because that lever is entirely inside the company that built the agent. That is not a scandal — it is just the shape of the tool. Treat the agent as a competent analyst with a house bias, and grade its output against blended acquisition cost rather than in-platform return.
The mitigation is unglamorous and effective: keep approval gates on, set hard campaign-level caps inside Ads Manager instead of trusting the conversational gate, and write down which actions may run unattended. We covered the protocol for that in setting guardrails before agents spend, and the same logic applies when the agent belongs to the client.
What to do with this week
- Write the diagnostic prompt library. Six or seven questions your team asks every account, each with a definition: what “top performing” means here (CAC payback, not click-through), which date window, which attribution setting. It works whether a human or the client’s agent runs it.
- Audit the measurement layer for legibility. Pixel, conversion events, catalogue, campaign naming, UTM structure. This is now the foundation of every number anyone repeats.
- Publish the permission map. What may act unattended, what queues for approval, what never leaves the queue — and apply it to your own internal agents too, not only the client’s.
- Rewrite one scope line. Move from “we run your ads” to “we own the standard your numbers are measured against.” Only the second one survives a free agent in the client’s app.
The uncomfortable read of Meta’s data is that the first job clients hand an agent is the job most agencies describe as reporting. That is not a threat if reporting was already the loss leader; it is the whole business if it was not. The compensating move is to own the definitions, the data model and the approval boundary, because those are the parts an agent cannot supply for itself — the argument in client reporting as the place agency AI earns its keep, and it sharpens the question from when clients get their own agent: not whether the client will have one, but what standard it is held to.
If the plumbing is the part you would rather not touch, have us audit the tracking, naming and catalogue setup your client’s agent is already reading from — because every number that agent reports back will be judged as yours.