Merchant playbooks

Rules for AI agent orders: spending limits, holds and human review

Listen · 9 min

Narrated by an AI voice. Code examples are on the page.

Short answer

Give AI agent orders their own rules: a maximum amount per order, a daily limit per customer, a list of agents you accept, and a hold for anything unusual so a person can approve or decline it. Agent platforms and payment providers already add some limits, such as Stripe's Shared Payment Token caps and AP2 mandates, but none of them know your store's risks. Default a held order to decline if nobody answers in time.

Key takeaways

  • Set a per-order cap and a daily cap per customer.
  • Hold refunds, cancellations and unusual orders for a person.
  • If nobody reviews a held order in time, decline it.

Why do AI agent orders need their own rules?

Because the person paying is not the one clicking. When an AI agent buys from your store, a shopper gave it a task, and the agent filled the cart and sent the order. Your normal checkout assumes a human looked at the page and pressed "buy".

Google put the problem plainly when it announced AP2, its Agent Payments Protocol: "today's payment systems generally assume a human is directly clicking 'buy' on a trusted surface." Agents break that assumption.

Most of the time, AI agent orders will look like any other order. The rules are for the rest: an agent that orders 40 units instead of 4, a cart padded with extras, a refund request nobody asked for, or a request that claims to come from a known agent but doesn't. Rules let you decide in advance what happens, so you are not deciding at 2 a.m. after the goods have shipped.

What limits do agents and payment providers already set?

Several layers already cap what an agent can spend. They help, but they are set by the shopper or the agent, not by you.

  • Stripe Shared Payment Tokens. An agent paying through Stripe hands you a shared payment token, not a card number. Stripe says the agent grants each token "with usage and expiration limits", and that "The agent sets the maximum amount to match the total amount of the transaction." Stripe sends you an event when a token is deactivated, which it says covers "consumed, resolved, expired, or revoked".
  • ACP delegated payments. The Agentic Commerce Protocol's payment spec says a delegated token "MUST ONLY be usable within the provided Allowance (reason, max_amount, currency, expiry)." The allowance reason must be one_time.
  • AP2 mandates. In AP2, a shopper who delegates a task signs an intent mandate that, in Google's words, "specifies the rules of engagement", such as price limits, timing and other conditions. The approved cart becomes a signed cart mandate.
  • The agent platform. OpenAI says that "across all our users, we require human review before any consequential action." For ChatGPT Wallet, it says "The customer consents only to the specific transaction."

Card networks add their own controls through agent-specific tokens. See how card networks handle AI agents for that side.

What none of these layers know is your business. A token capped at the cart total doesn't stop an agent buying your whole stock of a limited item. That is what your own rules are for.

Which rules should you set for AI agent orders?

Start with a few simple limits, then add more once you see real traffic.

Rule What it stops A sensible starting point
Per-order spending limit One unusually large order Your typical order value, times a margin you're comfortable with
Daily limit per customer Many small orders that add up A few times your average daily spend per customer
Accepted agents Orders from agents you haven't decided to trust Only agents that identify themselves and that you've chosen to accept
Country rules Orders shipped where you don't sell Match the countries your checkout already allows
Currency match A cart priced in a different currency from the payment Decline on mismatch
Refunds and cancellations Money going out without a person seeing it Always hold for review
Subscriptions A one-off purchase quietly turning into a recurring charge Hold or decline unless the shopper clearly agreed

Keep each rule easy to explain. If a customer or a payment provider asks why an order was declined, you want to point to one clear rule.

Also watch for text that tries to steer the agent, such as instructions hidden in a product review or an order note. Our guide to prompt injection and shopping agents covers that risk.

When should an order be held for a person to review?

Hold an order when a wrong decision would be expensive or hard to undo. A held order waits for a person on your team to approve or decline it.

Good candidates for a hold:

  • Refunds, cancellations and returns.
  • Orders above your per-order limit but not obviously wrong.
  • The first order from an agent you haven't seen before.
  • Orders that change something recurring, such as a subscription.

Give every hold a deadline, and decide now what happens when it passes. For most stores the safer default is to decline. Stripe takes the same approach with its own approval step: "If your hook doesn't respond within this time, Stripe declines the payment."

Make sure someone actually sees the hold. A queue nobody checks just turns into slow declines, which annoys real customers.

How do you approve or decline agent orders on Stripe?

If you sell through Stripe's Agentic Commerce Suite, Stripe runs checkout for agents by default. Stripe says it checks inventory against your catalog and "might run fraud checks depending on your Radar setup."

To add your own decision, Stripe offers an order approval hook. Before it completes a checkout, Stripe sends your endpoint a request and, in Stripe's words, "You must approve or decline the request." Three details matter:

  1. It is fast. Stripe says it "enforces a four-second timeout for your hook." That is enough for automatic rules, not for a person to look. For human review, combine it with the next option.
  2. Retries happen. Stripe warns that "Some agents might retry these API requests, so you must make sure that your endpoint is idempotent." Your rules must give the same answer to the same order twice.
  3. Availability varies. Stripe lists the hook's availability as "Frontier", so check your own account's settings.

To give a person time, Stripe lets you switch agent payments to manual capture in the Agentic Commerce settings. By default, it says, "Stripe captures payments immediately after purchase." With manual capture on, you capture each payment yourself through Stripe's API, so a person can review the order first.

Why keep a record of every decision?

Because agent orders raise a new question when something goes wrong: who authorised this? A record of what was ordered, which rule applied and who decided gives you an answer.

AP2 was built around this idea. Google describes the chain "from intent, to cart, to payment" as creating "a non-repudiable audit trail that answers the critical questions of authorization and authenticity." Your side of that trail is your own decision record: the order, the rule that fired, and whether a person approved it.

Keep it somewhere it can't be quietly edited, and keep it for as long as you might face a dispute. For more on how signed mandates work, see AP2 and the FIDO Alliance.

How Sill helps

When an AI agent places an order, Sill applies your rules to approve it, decline it, or hold it for a person to review. Rules include a per-transaction spend cap, a daily spend cap per customer, and human review for refunds and cancellations. Held orders notify your team by email, and by Slack if you connect it, and are declined automatically if nobody decides by the deadline. Every decision is signed and recorded. Sill also accepts AP2 mandates and checks them against your rules. Sill is not a payment processor: Stripe moves the money. You can check how your site looks to agents with a free readiness scan.

Questions people ask

Can I make ChatGPT ask the shopper before it buys from my site?

Not as a site owner, at the time of writing. OpenAI says it doesn't yet have features that let specific site owners control what actions users can take, but that across all users it requires human review before any consequential action.

What happens if my Stripe order approval hook is slow?

Stripe says it enforces a four-second timeout. If your hook doesn't respond in that time, Stripe declines the payment.

Can an agent reuse a Shared Payment Token for another purchase?

Only within the limits the agent set. Stripe says each token carries usage and expiration limits, and a token that is used, expired or revoked is deactivated. A recurring charge needs a token with a recurring interval set.

Do order rules replace fraud checks?

No. They sit alongside them. Stripe says it might run fraud checks on agent purchases depending on your Radar setup. Your rules cover business limits that a fraud model doesn't know about.

Sources

  1. Stripe Docs: Shared payment tokens (sellers) · checked 17 Nov 2026
  2. Stripe Docs: Customize with hooks · checked 17 Nov 2026
  3. Stripe Docs: Handle orders · checked 17 Nov 2026
  4. Stripe Docs: How Agentic Commerce Suite works · checked 17 Nov 2026
  5. Google Cloud: Announcing Agent Payments Protocol (AP2) · checked 17 Nov 2026
  6. Agentic Commerce Protocol: Delegate payment RFC · checked 17 Nov 2026
  7. OpenAI for Developers: Help ensure cloud browser access · checked 17 Nov 2026
  8. OpenAI for Developers: Accept payments · checked 17 Nov 2026
  9. OpenAI for Developers: Agent access · checked 17 Nov 2026

Written by The Sill Signals Team. Last reviewed 17 Nov 2026.

Share this articleLinkedInXBlueskyEmail

Keep reading

See which AI agents visit your site, and decide what they may do.

Scan your site free