How to set spending limits for an AI agent

A single number is not enough. A daily cap alone lets one payment use the whole day at once. A per-payment cap alone lets an agent stuck in a loop make the same small payment hundreds of times. And neither tells you which payments you want to see before they happen. The limits also have to live somewhere the model cannot edit, because an agent that can be told to raise its own limit does not have one. Finally, the limits need to be consistent with each other, or you end up with a configuration that looks safe and behaves strangely.

1. Start from the largest single payment you'd accept. Set the per-payment cap to the biggest amount you would ever want this agent to pay in one go, even with your approval.
2. Set a daily cap you could lose without trouble. The daily cap is the most the agent can spend in a day, counted across all its payments. Size it as a float you can afford to lose to a bug or a bad instruction, not as a credit line.
3. Choose an auto-approve limit for small, routine payments. Below this amount the payment can go through without you. Keep it at the size of the routine payments you identified, such as a few dollars of API credit.
4. Choose an ask limit for payments you want to see. Between the auto-approve limit and the ask limit, the payment is held until you approve it. Above the ask limit it is refused.
5. Keep the order consistent. Auto-approve should be no more than ask, ask no more than the per-payment cap, and the per-payment cap no more than the daily cap. A good policy service rejects anything else when you save it, rather than failing at payment time.
6. Tell the agent what to do when it hits a limit. Give it a way to read its remaining budget, and an instruction to stop and ask you rather than retrying or splitting a payment into smaller pieces.

• A limit does not stop an agent paying the wrong recipient. Pair amount limits with a recipient allowlist where you can.
• A daily cap does not reset when you choose. Check which time zone the service counts a day in; Veyra counts UTC days.
• Raising a limit in the prompt does nothing if the service enforces it, and lowering it in the prompt does nothing if the service doesn't.

• Veyra docs
• MCP specification: Tools (human in the loop, security considerations)
• Stripe docs: Issuing spending controls

• Auto-approve limitAn auto-approve limit is the largest single payment an agent may make without a person approving it, within any other caps that apply.
• Payment intentA payment intent is an object representing a payment that has been requested but not necessarily completed, which moves through defined states such as awaiting approval, submitted, confirmed or cancelled.

• Human-in-the-loop approval for AI agent purchasesHow to make an agent wait for a person before larger payments: where the request goes, what the approver checks, and what happens if nobody answers.
• How to restrict who an AI agent can payA recipient allowlist limits where an agent's money can go, which amount caps can't. How to build one, match addresses safely and handle new payees.
• How to stop an AI agent making duplicate paymentsAgents retry when unsure, and retries can pay twice. Idempotency keys, status checks and tool wording that keep one intended payment to one payment.
• How to let an AI agent make paymentsGive an agent a payment tool, not a key: limits enforced outside the model, a human for larger amounts, and a record of every decision.

All guides →