Stripe will let an AI agent pay your invoices. Start by letting it prepare them

What Stripe launched

On October 21, 2026, Stripe announced Treasury for agents. If your startup banks through Stripe Treasury, you can now let AI agents, including ChatGPT and Claude, check balances, pay invoices, transfer funds between accounts and convert currencies. Agents connect through Stripe's MCP server, the bridge that lets them read Stripe data and call Stripe's API.

Stripe says more is on the way, including letting agents create cards. It also says businesses will be able to set spending limits for their agents in the coming weeks.

That last detail is the one to plan around. Today you can decide what an agent sees and what it may do. The limit on how much it can spend isn't there yet.

The controls you do have

Stripe lists three choices you make for each agent:

• Which accounts and data it can access. An agent doesn't need to see every account to do its job.
• Whether it can read or act. Checking a balance and paying a bill are separate permissions.
• When money movement requires human approval. You can require a person to approve before funds move.

Those are good controls, and they're the right shape. The founder Stripe features in its announcement runs his business with AI agents. His agent checks balances, new and canceled subscriptions, failed payments and refunds every morning and sends him a report. He plans to let it pay contractors and move money between accounts, with every payment approved by him first.

That's the pattern we'd recommend for nearly everyone starting out.

Our rule: prepare freely, approve every move

Reading the balance is a report. Moving the money is a decision, and decisions should have a name next to them.

Here's how that splits in practice.

Let the agent read. Morning balance summaries, upcoming invoices due, failed payments, unusual refunds, currency exposure. This is where agents save real time, and the risk is low because nothing changes.

Let the agent prepare. Match an incoming invoice to a contract, draft the payment with the right amount and account, flag anything that doesn't match last month. A prepared payment waiting for approval is most of the work done.

Keep the approval with a person. Turn on human approval for every money movement, including transfers between your own accounts. It sounds excessive for an internal transfer, but that's exactly the kind of instruction an agent can get wrong, or be tricked into, through a malicious invoice or a crafted email in its inbox.

Narrow the accounts. Give the agent access only to the operating account it needs, not your reserves.

When Stripe's spending limits arrive, you can loosen this deliberately. Maybe small transfers between your own accounts stop needing approval, while payments to new vendors always do. But make that change because the agent earned it over weeks of good preparation, not because the setting appeared.

What a useful prepared payment looks like

The value of "prepare" depends on what lands in front of you. A bare request that says "pay $4,200 to Acme?" is barely better than doing it yourself, because you still have to go find out whether it's right. A good prepared payment answers your questions before you ask them.

Ask the agent to include, for every payment it drafts:

• What it's for. The invoice, the contract or order it matches, and the period it covers.
• What changed. Whether the amount differs from the last payment to the same vendor, and by how much.
• Whether the payee is new. A first payment to anyone deserves a slower look, and so does any change to a vendor's bank details.
• Which account it comes from. Especially if you hold balances in more than one currency.
• What happens if you wait. The due date, and whether paying late carries a fee.

That list turns approval from a guess into a check you can do in thirty seconds. It also gives you a record. When someone asks in March why a payment went out, the agent's note and your approval are sitting next to each other.

One more habit worth building early: have the agent tell you what it chose not to prepare. An invoice it couldn't match, a refund that looked odd, a subscription that failed twice. The things it skips are often more useful than the things it drafts.

What this doesn't solve

Approval only works if the person approving actually reads. If your agent prepares twenty payments and you approve them all with one tap every Friday, you've kept the button and lost the check. Keep batches small enough to review, and have the agent explain anything new: a vendor it hasn't paid before, an amount that changed, a currency it doesn't usually convert.

Also remember what you're connecting. Stripe's own fine print notes that Stripe is not a bank, and that checking accounts are provided by Fifth Third Bank. Your agent is acting on a real business account. Treat its access the way you'd treat giving a new bookkeeper online banking: limited at first, reviewed often, widened slowly.

And write down who is responsible. When an agent pays the wrong invoice, the question isn't what the agent was thinking. It's who approved it and why the check didn't catch it.

If you're setting this up

Start with read-only access for two weeks and see whether the reports are useful. Then let the agent prepare payments with approval on. Then decide, with real history in hand, whether any category of payment can go without approval once limits exist.

Keep the setup written down somewhere your team can find it: which agent has access, to which accounts, with which permissions, and who approves. Review it whenever Stripe adds a capability, like card creation, because each new ability is a new decision rather than an automatic upgrade.

If you'd like help designing the approvals, account permissions and checks around agents that touch your money or your books, that's what our business automation team builds. Talk to us.

Related work: Spin Market.