6 min read
Ten minutes. According to Wix’s announcement, that is how long a small business owner has to approve a purchase by WhatsApp or SMS before an AI agent can spend their money. For most people, the interesting news was that agents can now pay at all. I think the opposite is true. The approval step is the most valuable part of this launch, and a solo owner should be in no hurry to give it up.
Symphony, Wix’s team of AI agents for small businesses, now connects to Link, Stripe’s digital wallet, for customers with US accounts paying in dollars. Wix’s own examples include reordering stock, paying invoices received by email, paying registration fees and sending a gift to an unhappy customer. The technology is well designed. The risk is in how a tired owner will use it.
Credit where it is due: the guardrails are real
A skeptic should start by saying what works. Per Wix’s explanation, the design is careful in several ways:
- The agent never sees your card. It receives a one-time card or token, capped at the approved amount.
- You see the merchant, the item and the exact total before you say yes.
- Any change means a new approval. A different total or a different shop starts the process again.
- Duplicate payments are checked. Symphony asks Link before retrying, so a glitch should not charge you twice.
Stripe’s documentation describes the same pattern: a spend request for a specific purchase, approval by the customer, then a one-time credential. Compare that with handing a bot your saved card details, and this is a large step forward. It also matches the principle we set out in our five rules for giving an agent an account: give an agent the least access that gets the job done.
Where a good design meets a real week
So what is the problem? Three things worry me, and none of them is a flaw in Stripe’s engineering.
Approval fatigue is a human bug
A ten-minute window is short by design. It also means the request lands while you are with a client, driving or packing orders. The first dozen approvals get careful attention. By the fiftieth, you are tapping yes to make your phone stop buzzing. Every safety system built on human review shares this weakness, and it gets worse as the agent earns your trust.
“Pays invoices received by email” is the riskiest sentence here
Fake invoices are one of the oldest business scams, and they work because they look routine. The FBI’s 2024 Internet Crime Report put reported losses from business email compromise at about $2.8 billion. That is the category of fraud where a criminal impersonates someone you do business with to redirect a payment.
An agent reading your inbox can be fooled by the same tricks that fool people, and possibly by new ones. A message may even contain hidden instructions written for the AI. The approval screen will show you the merchant and the total, but it cannot tell you that the supplier’s “new bank details” came from a thief. If the amount looks like your usual invoice, a quick yes is exactly what the scammer is counting on.
Wix’s post is silent on when things go wrong
The explanation I read does not cover refunds or disputes for agent purchases. Nor does it state fees for agent payments. Those may well be handled by your normal card rules, but “may well be” is not a plan. Before you approve anything meaningful, find out what happens when an agent buys the wrong thing from a seller who does not take returns.
The next step is the one to resist
Wix describes per-purchase approval as deliberate, and says Stripe’s roadmap points toward budgets: you approve a monthly restocking budget once, and the agent spends inside it, with the cap enforced by Link rather than by the AI model. Wix has not promised this for Symphony, but the direction is clear.
Budgets sound like freedom. For a solo owner, I think they are premature. A cap limits how much can go wrong, but it says nothing about whether each purchase was a good one. If your agent reorders the wrong scent of candle within budget, the cap worked and you still own 200 candles you cannot sell. Larger firms have a finance person who reviews the statement. You are the finance person, and you are also busy.
This is where the AI autonomy ladder earns its keep. An agent should climb a rung only after it has a track record you can measure, and spending money is one of the highest rungs there is.
My rules for a solo owner trying agent payments
- Start with repeat purchases only. A monthly reorder from a known supplier at a known price is a fair test. Do not start with anything new.
- Keep invoices out for now. Pay emailed invoices yourself until you have months of clean history. Confirm any change of bank details by calling a number you already have, never one from the email.
- Fund it from a low-limit card. Set a limit you could lose without losing sleep, and keep it separate from your main account.
- Ignore requests you cannot look at properly. Letting a request expire costs you almost nothing, so let it. A missed reorder can be redone at your desk.
- Review the log weekly. Spend five minutes comparing every agent purchase with your orders. Our earlier piece on where to stop when AI moves into the money goes further on this.
- Write down your budget test. Say in advance what would earn the agent a budget, such as 90 days and 30 purchases with no errors. Without that, the decision will get made by convenience.
The honest case against me
You could argue that I am overrating human review, and that an agent checking every invoice against a purchase order would catch more mistakes than a distracted owner does at midnight. For a business with high order volume, that may prove right, and the math on time saved could eventually justify budgets.
But that argument needs evidence, and very little exists yet for businesses this small. A solo owner has thin margins and no one to catch an error. Letting bigger companies run this experiment first costs you a few minutes per purchase. That is a price worth paying.
Would you let an AI agent spend your money without asking first? Tell me what number on the budget would make you comfortable, or whether any number would.
