For the last two years the expensive part of putting AI into a small business was never the model. It was getting the model to see the business: orders in one system, payments in another, books in a third, ad performance in a fourth. Joining those up was integration work, and integration work is paid for in weeks.
Meta has just given that away.
01 — What happened
On 29 September 2026, Meta announced Muse for Small Business, an extension of the Muse agent it launched on 8 September. The announcement is titled “The Future Is for Everyone”, and the substance of it is a connector list.
Muse now connects to fifteen named third-party services: Asana, Box, Canva, Dropbox, Figma, Granola, HighLevel, Intuit QuickBooks, Klaviyo, Lovable, Notion, Shopify, Slack, Stripe and Zoom. Owners can also connect Facebook Pages, Instagram professional account analytics and Meta ad accounts. Where a service is unsupported, Meta documents how to build your own connector, and says more will follow.
The stated skills are operational rather than creative: analyse sales data and build a growth plan, triage a workload by priority, review ad analytics and draft campaigns, assess financial metrics and flag suspicious expenses, surface the emails needing a reply and draft them.
Three details matter more than the feature list. Meta says Muse “is free for most of what people need”, with subscription plans for higher usage and no separate business tier. Nothing “publishes, sends, or spends without your approval.” And the launch covers the United States and Canada only.
Meta’s framing of why is unusually plain: small businesses “told us they’re short on hours, not ideas.”
One more event belongs here. On 21 September, eight days before the business launch, Amazon blocked Muse from shopping on its store, saying Meta had neither told it the agent would access the store nor sought authorisation, that the agent did not identify itself while browsing, and that it appeared to capture and store customer credentials. Meta denies the last point, stating Muse has no visibility into passwords or payment methods.
02 — Why it matters
Nothing in this announcement makes a model cleverer. What changed is the cost of a single capability: one assistant that can look at your orders, your payments, your books and your advertising at the same time, for nothing.
That capability used to be the business case for a connected system. An owner asking “which of my products actually makes money once I count ad spend and refunds?” was asking a question that needed three systems joined together. The joining was the project. Now a free agent will attempt it on Tuesday afternoon.
This is a genuine shift, and it cuts in an uncomfortable direction. While reading across your business was expensive, the quality of your data stayed hidden — nobody asked the awkward question, so nobody found out that product cost lives in a spreadsheet on someone’s laptop and refunds are reconciled by hand at month end. A free reader exposes all of it, not by complaining but by answering confidently from whatever it can reach.
It also introduces an unfamiliar dependency. This is not an agent you commissioned, running on access you granted, inside a system you own — the distinction we drew when access became the real bottleneck in AI automation. It is a platform agent, reading through connectors maintained by two parties, either of which can withdraw. Amazon demonstrated that eight days early.
03 — The economics
Start with what actually went to zero, because it is narrower than the headlines suggest. The free thing is the read: pulling numbers out of several systems and reasoning across them. The write — publishing, sending, spending — still needs human approval by design. The record underneath has not improved at all.
Take an illustrative example. A retailer does 900 orders a month: storefront on Shopify, payments through Stripe, books in QuickBooks, email through Klaviyo, advertising on Meta. Connecting all five takes an afternoon and costs nothing. The owner asks the question they have wanted answered for a year: which products are profitable after advertising?
To answer properly the agent needs landed cost per product, ad spend attributable to product, and refunds netted off revenue. In most businesses of that size the first lives in a spreadsheet no system reads, the second does not exist at product level, and the third reaches the books a month late. The agent will still answer. It will be specific, well-formatted and wrong in a direction nobody can see.
That is the real economics of this release: an asymmetry rather than a saving. Reading is free. Being confidently wrong costs what it always did, except that it now happens faster and with more authority behind it. An owner who reprices a range on a bad margin figure has done more damage than one who never asked.
There is a second asymmetry. One of the connectors is your Meta ad account, and Meta’s revenue comes from advertising. The agent reviewing your ad performance and drafting your next campaign is supplied free by the company selling the inventory. That is not a scandal, it is an alignment — the normal condition of free tools. It means advertising is the one recommendation to check against your own numbers rather than accept, which is harder than it sounds when those numbers are the thing you lack.
Finally, the switching cost. A free connector carries no licence to cancel, which makes it feel costless to depend on. It is not. If a weekly process works only because an agent reads Stripe for you, and that connector changes, you are not re-buying software — you are rebuilding a process under time pressure, with the people it freed up already redeployed. Hours saved is the wrong way to value an automation, and that holds for hours saved free of charge.
04 — What businesses should do
Use it as a diagnostic before you use it as a workflow. Spend a week asking the five questions you most want answered — margin by product, where enquiries are lost, which customers are quietly leaving, what next month’s cash looks like, which marketing paid for itself. Note which it answers and which it cannot. The failures are the valuable output: they name, for free, which parts of your record are not fit to be read.
If you are small and your systems are tidy, stop there. For a business handling, say, under 200 orders or enquiries a month across two or three mainstream systems, a free read-only agent is plausibly the whole answer. Do not commission integration work to improve on something you have not outgrown.
Reconcile by hand any figure you would act on, at least once. You are not testing the model’s intelligence; you are testing whether your systems contain what the answer requires. Those are different failures, and only one is fixable by a vendor.
Keep approval switched on for anything that spends, and grant read-only where a connector allows it. Meta’s default already requires approval to publish, send or spend. Convenience will push you to loosen it on the ad account first, which is the wrong place to start.
Write down the manual fallback for any process that comes to depend on a connector. One paragraph each, naming who does it by hand and how long it takes. Amazon’s block is a reminder that access between two large companies is a commercial arrangement, not a feature.
If no agent like this has reached your market yet, use the time. The question it asks of a business — is your data worth reading? — is the same whoever ships first, and the answer takes far longer to fix than a connector takes to appear.
05 — Where technology fits
For a large number of businesses, honestly: nothing needs building here. If your orders, payments and books already sit in mainstream systems and your record-keeping is reasonable, a free agent plus an hour of connecting is the right amount of technology, and anyone proposing a project on top of it should be asked what it adds.
The case for building arrives in one situation: the diagnostic shows the agent cannot answer your most valuable questions because no system holds the data. That is not an AI problem. It is the cost of disconnected business systems becoming visible, and the fix is a connected record rather than a cleverer model — the work connected business systems exist to do.
One boundary is worth stating plainly. A free, human-approved agent is a good reader and a poor operator. If a process must run unattended, on a schedule, with an audit trail and a named owner — invoicing, fulfilment, anything a customer will dispute — that is automation you own, not a chat window you trust. One is for asking questions; the other is for running the business.
06 — RevenueStack’s take
Our view: the significant thing Meta shipped is not an agent. It is the collapse in the price of reading across a business, from a project to nothing, in a single afternoon.
That collapse moves the competitive question. For a decade the advantage went to businesses that could afford to see themselves clearly. Now everyone can — and what they will see is the state of their own record. Two businesses will connect the same fifteen services this month. The one whose customer and cost data was already in order will get decisions out of it. The other will get fluent answers built on a spreadsheet nobody has updated since March.
Free intelligence does not level the field. It moves the advantage from who can afford to look to who has something worth looking at.
Sources
- The Future Is for Everyone: Muse for Small Business — Meta, 2026-09-29
- Meta is expanding its AI agent Muse to small businesses — TechCrunch, 2026-09-29
- Meta launches Muse for small businesses, linking its AI agent to Shopify — The Next Web, 2026-09-29
- Meta gives small businesses an AI agent that knows their work — Help Net Security, 2026-09-29
- Amazon shows Meta's Muse AI shopping agent the door — The Register, 2026-09-21
