Meta Muse for Small Business: A 6-Day-Old Free AI Agent, Costed Out — What It Saves, What It Can't Touch, and the Approval Ga
SMALL BUSINESS TOOL LAB · COST AND WORKFLOW REVIEW
A free AI agent can still cost an owner hours of supervision. In our hypothetical five-hour-a-week workflow, 25 one-minute reviews per workday consume 2 hours and 5 minutes a week. Add corrections and maintenance, and the recoverable time is much smaller than the advertised task time.
Meta announced its small-business expansion on September 29, 2026. That makes the announcement six days old on this review date—not the entire Muse product six days old. The personal agent was introduced earlier in September. A durable buying decision needs the launch date, a clear cost model, and the current settings, rather than a headline that becomes stale tomorrow.
Start with a narrow question: can the tool finish a specific recurring job with less total human effort than your current process? The answer depends on reviewing time, correction time, connected-app costs, and the authority you grant. “Free” describes one subscription option. It does not describe the total cost of operating the workflow.
What Meta has actually published about pricing
The official Muse subscription help page lists a free option with a usage limit, Power at $20 per month with 500 million Muse tokens per week, and Maximum at $100 per month with 3 billion per week. These paid prices are official published figures, not merely unconfirmed press estimates.
| Plan | Published monthly price | Published allowance | Budget implication |
|---|---|---|---|
| Free | $0 | Usage limited; numeric free allowance not stated on the help page reviewed | Test real workload before relying on uninterrupted capacity |
| Power | $20 | 500M Muse tokens/week | Add $240/year before any applicable checkout charges |
| Maximum | $100 | 3B Muse tokens/week | Add $1,200/year before any applicable checkout charges |
Meta says availability and benefits can vary by region and account, and that onboarding shows the available benefits. When the free limit is reached, users can upgrade or wait for it to refresh. Confirm the account’s actual offer and final payment amount. The published monthly figures do not establish a universal tax-inclusive checkout price.
A token allowance is not a promise of a particular number of completed campaigns or customer replies. The official page does not provide a task conversion. A long research job, repeated revisions, and a short summary need not consume the same amount. Track completed useful work alongside usage instead of inventing a tasks-per-token exchange rate.
Existing services remain separate expenses. A connection to a paid accounting or marketing product does not make that product free. Include new connector-related subscriptions, API charges, or service upgrades only if your intended workflow requires them. Leave existing subscriptions out of the incremental cost calculation when they stay unchanged, while keeping them in the overall business software budget.
Fifteen named integrations are not the agent’s outer boundary
Meta’s small-business announcement names Asana, Box, Canva, Dropbox, Figma, Granola, HighLevel, Intuit QuickBooks, Klaviyo, Lovable, Notion, Shopify, Slack, Stripe, and Zoom. It also describes connections to Facebook Pages, Instagram professional analytics, and Meta ad accounts. Suggested uses include sales analysis, campaign drafts, and unusual-expense review.
Those are announced capabilities and examples, not independent proof of accurate outcomes. Ask what data the tool used, which period it analyzed, and what it excluded. A growth plan based on gross sales alone might overlook returns, discounts, and advertising costs. A flagged expense still needs a person to decide whether it is actually wrong.
The list is not a hard ceiling. Meta’s connector documentation supports custom connectors, while its technical overview describes a cloud computer with a browser. An app missing from the launch list is therefore a compatibility question, not proof that Muse can never interact with it.
Custom access adds implementation work. An API may lack the particular operation you need, a provider may restrict automation, or your account may not have the required permissions. Meta states that it does not review custom connectors or how they use information. A possible integration is not the same as a supported, tested production workflow.
Build your own task map with three columns: required source data, required final action, and the available access route. “Analyze advertising” is too broad. “Compare last month’s campaign spend with completed orders, excluding refunded orders” is specific enough to test. Keep a human fallback for any job whose access or reliability remains uncertain.
The operating-cost model: five hours saved is only the starting point
Consider a two-person online shop whose owner currently spends five hours each week assembling routine drafts and reviewing business reports. We assign the owner’s time a planning value of $35 per hour. Both numbers are assumptions. They are not measured Muse performance or a claim about typical small businesses.
To make the test concrete, imagine that Muse removes those five hours of original task work but creates 25 approval reviews per workday. Use five workdays and one minute per review. That gives 125 minutes, or 2.0833 hours, each week. The original draft rounded this to two hours and understated the review burden.
We also assume 30 minutes of weekly correction work and 15 minutes of routine maintenance. “Correction” means fixing drafts or checking questionable classifications. “Maintenance” means keeping the workflow working. Do not include the same minute in both categories.
| Weekly time ledger | Minutes | Calculation |
|---|---|---|
| Original task work removed | 300 | 5 hours × 60 |
| Approval reviews added | −125 | 25 × 5 days × 1 minute |
| Corrections added | −30 | Assumed weekly total |
| Maintenance added | −15 | Assumed weekly total |
| Net time recovered | 130 | 2 hours 10 minutes/week |
At $35 per hour, 130 minutes has an assumed value of $75.83 per week. Using 52 weeks divided by 12 months gives $328.61 per month, or $3,943.33 per year, before incremental cash costs. The original gross calculation—five hours times $35 times 52—is $9,100, but it omits all the new work in this example.
Now suppose the shop also incurs $15 a month in incremental connected-service costs. This is a hypothetical input, not a Muse fee. On Free, the model’s monthly net economic value is $313.61. On Power it is $293.61, and on Maximum it is $213.61, before setup, tax, and other changes. An account that needs no additional service costs should replace $15 with zero.
Setup matters in month one. If onboarding, testing, and cleanup take four hours, the assumed setup time is worth $140. Month-one net economic value becomes $173.61 on Free, $153.61 on Power, or $73.61 on Maximum. The model assumes the weekly performance is already achieved during that month; a slow rollout could produce less benefit.
Recovered time is not automatically money in the bank
The $35 hourly input values owner time. It does not mean a bank deposit rises by $35 whenever an hour is saved. If the owner uses the recovered time to rest, the benefit may be worthwhile without creating revenue. If it allows more client work, the incremental profit depends on the work actually sold and delivered.
Cash savings require a cash expense to disappear: an avoided contractor invoice, fewer paid overtime hours, or a cancelled service. Do not claim lower payroll just because a salaried employee has more capacity. And do not count both the employee’s recovered time and an avoided contractor bill for the same completed work unless they are distinct benefits.
Keep two totals in your pilot record. One tracks money paid or avoided. The other tracks net time recovered. Their relationship helps you make a decision, but merging them into one “savings” headline can hide whether the tool actually reduces spending.
The approval gap has two parts
The first gap is operational: the agent can generate work faster than the owner can inspect it. Twenty-five items arriving while you serve customers are not equivalent to 25 items reviewed in one quiet block. Count interruption and switching time where it occurs; a one-minute assumption may be too optimistic.
The second gap is permission scope. Approval does not always mean a new click for every action. Meta’s guidance and approval help page describes one-time, task, site, and standing connector permissions. A standing grant can authorize future actions of the covered type without another request. Read the actual scope rather than interpreting a marketing assurance as a universal per-action prompt.
The documented settings include connector and web-access controls, with an “Always ask” option. The activity log records actions and permissions. Meta also warns that some actions, such as sending an email, cannot be undone. These controls help you supervise; they do not make every business consequence reversible.
For a pilot, write down which actions may proceed and which require a separate decision. A draft can be useful without being sent. A suggested advertising change can be reviewed without changing a live campaign. A refund recommendation can remain a recommendation until the authorized person approves it through the business’s normal process.
A text instruction alone is not evidence that a hard spending cap exists. If your test requires a budget ceiling, verify where that ceiling is enforced—within the product, the connected service, or your manual process. Reject a setup that depends on a feature you cannot find or test.
Our sensitivity test: when supervision erases the benefit
Keep the same five gross hours, 30 correction minutes, and 15 maintenance minutes. Change only approval volume or time per item. This isolates the bottleneck instead of changing every assumption at once.
| Approval pattern; five workdays | Approval time/week | Net recovered time/week |
|---|---|---|
| 10 items/day × 1 minute | 50 minutes | 205 minutes |
| 25 items/day × 1 minute | 125 minutes | 130 minutes |
| 25 items/day × 2 minutes | 250 minutes | 5 minutes |
| 30 items/day × 2 minutes | 300 minutes | −45 minutes |
The two-minute review case leaves five minutes a week, worth about $12.64 a month at our assumed hourly value. Even the Free scenario with $15 of incremental service costs then has negative modeled economic value. A task may technically work and still fail the operating-cost test.
Use this equation with matching monthly units: net value = net recovered hours × hourly value − subscription − incremental service costs − other incremental costs. Divide one-time setup costs across your chosen evaluation horizon when comparing longer-term options, or charge them entirely to month one when examining cash and time requirements for launch.
In the Power example, recurring assumed cash costs are $35 a month: $20 plus $15. At $35 per hour, the workflow must recover more than one hour per month to clear those recurring costs. That threshold excludes setup and applies only to the assumed value of time. A cash-payback test needs actual cash savings instead.
Which work should remain outside this pilot?
Do not confuse an operating boundary you choose with a permanent technical limitation. This pilot leaves bank transfers, tax filing, contract acceptance, and sensitive customer communications with authorized people. That is our proposed operating scope, not a claim that no AI system can technically perform those actions.
Physical work still needs someone on site. A campaign draft cannot inspect damaged inventory. An expense flag cannot confirm what a supplier actually delivered. The agent can organize supporting information, but the business still needs evidence and accountable decisions.
Read access also deserves thought. A report can reveal customer identities or commercial details even when the agent cannot send a message. Meta’s connector documentation notes that disconnecting stops further exchange but may leave previously used information in memories or conversation history. Revoking a connection is not a promise that every earlier copy vanishes.
A five-workday pilot you can actually score
Choose one recurring task with enough repetitions to measure. Record the current process on comparable items before adding the agent. Use anonymized or synthetic test data first where it serves the task, and confirm the required access before connecting live business systems.
For each item, record the manual baseline, approval minutes, correction minutes, final acceptance, and any fallback work. An attractive draft that cannot be used does not count as a completed outcome. Include time spent finishing the item manually after an unsuccessful attempt.
At the end of each workday, reconcile the task list with the activity log and check usage. Keep implementation details in your working record, including the granted scope and any unexpected behavior. Pause the test when the observed actions no longer match the pilot’s intended boundaries.
After five workdays, compare net time and actual incremental spend. If corrections dominate, narrow the task or improve its inputs. If approvals dominate, reduce generated volume or choose a job where review can be faster without sacrificing quality. Do not remove necessary checks merely to make the savings figure look better.
Choose the workflow before choosing the tool
A rules-based automation may fit a fixed trigger-and-action process, such as routing a completed order into a reporting queue. An AI agent may fit a job requiring context, interpretation, and drafts. A manual template may be the cheapest option when only a handful of items arrive each week.
Compare all options on the same accepted outcomes, not connector counts or subscription prices alone. Request each vendor’s current quote and include required base subscriptions and maintenance. This review does not rely on outdated Zapier plan names or treat a Microsoft add-on as a complete standalone software stack.
Muse is worth a contained trial when its access fits a real task and your measured review burden is manageable. Expand only after the evidence supports expansion. The decision is not whether a free agent sounds impressive; it is whether one well-defined job becomes less costly to complete, with authority you understand.