Meta Says Nothing Sends Without Your Approval. Its Own Designers Worry You'll Approve Everything.
· CX Pulse
Meta's Muse for Small Business promises nothing sends without your approval. Its own design notes describe how approvals turn into a reflex.
Meta Says Nothing Sends Without Your Approval. Its Own Designers Worry You'll Approve Everything.
On Tuesday Meta launched Muse for Small Business, a version of its AI agent that connects to the tools a business already runs on. The list includes Shopify, Stripe, QuickBooks, Klaviyo and Slack, plus the company's Facebook and Instagram accounts. Meta says Muse arrives already understanding "what you sell, what your brand sounds like, and what customers keep asking you about," and that it'll flag emails that need a response and write the first drafts. It's available in the US and Canada, free for most use.
One sentence in the announcement carries the whole trust case: "You're in control: nothing publishes, sends, or spends without your approval."
That sentence sits next to links to two longer pages from the consumer launch earlier in September, one by a Meta engineering VP on the safety system and one by the product's designers. We read both, because they describe what an approval is in practice.
What "your approval" covers
The safety page describes a separate system Meta calls Sentinel. It decides whether each action is allowed, denied, or put to the owner as a question. When it asks, the owner gets a card with accept and reject on it. When a business connects email, the owner picks whether Muse can only read the mail or can also send.
An approval doesn't have to be given each time, though. The page lists the kinds Muse supports: "one-time, session-scoped, task-scoped, time-bounded, or perpetual permission." It's direct about the goal, too. "The point is not to ask the user about everything." Meta wants "friction where consent matters while keeping routine operations flowing freely," and then says, "This balance is something we expect to tune over time as we have more experience with real users." Purchases get the strictest treatment: checkout prompts an approval with the exact details, every time.
The design page comes at the same problem from the other side. Its authors write that they wanted to avoid "banner blindness," which they describe as "where people approve everything to make it go away." Their answer is a default that stops for anything hard to undo, plus controls to make Muse "more or less cautious."
None of that contradicts the promise. A perpetual yes given in week one is still the owner's approval. Meta deserves credit for publishing the trade-off, because it's the part of the product that decides what customers actually receive.
The approver is the busiest person in the building
Meta's example customer is Tom Mulholland, who owns a grocery in Malvern, Iowa. "I work about 65 hours a week, and there are so many things where I'm the only one who can do them," he says in the announcement.
An owner in that position is also the only one who can approve. Every customer reply the agent drafts, every post, every ad change waits on the same person who's behind the counter. That can go two ways. The card sits until there's a spare minute, and the customer's wait becomes the seconds the agent took to draft plus the hours the owner took to look. Or the owner taps yes between customers to clear the stack, and after a week of that grants the standing permission so the cards stop coming.
The promise holds in both cases, since nothing went out without an approval. What the customer got depended on whether that approval was a read or a reflex, and the activity log looks the same for both.
If you run CX for a large operation
You've had approval gates for years, and you know how they age. Agents accept the suggested reply. Supervisors approve the credit over the limit. Somebody signs off on the new macro. Agents built like Muse add more gates and push more volume through each one, so the questions are the ones you'd ask of any control.
How long does an approval take, by action type? That time is inside the customer's wait and usually outside the response-time report. What share of drafts get approved unchanged, and how long was the card open? A 200-word reply approved in two seconds wasn't read. How many standing permissions exist, who granted them, and when did anyone last review the list? They deserve the same review cycle as admin rights. When a customer disputes something an agent sent, the audit trail will show a name on the approval, so decide now whether that person owns the sentence. The log already says they do.
If you're a smaller operation
You're the approver, and probably the only one. The decision that matters comes before you connect anything: which actions get a standing yes and which never do. Reading the inbox, pulling a sales report and drafting a reply are easy to hand over. Anything a customer will read and anything that spends money is worth seeing every time, at least until you've watched enough drafts to know what it gets wrong.
Then keep an eye on the stack. If customer replies sit in it through the afternoon rush, the agent wrote faster and the customer waited just as long.
Meta's promise is that you approved it. Whether you read it first is the part no vendor can promise for you.