Say your agent's code expects a payment call to answer in two seconds. Approvals work on a different clock. The call comes back pending, the person who has to say yes is in a meeting, and your agent has 40 minutes to fill.

An agent that goes silent for those 40 minutes looks broken, and one that keeps hammering the payment call looks worse.

What should an AI agent do while it waits for a purchase approval?

An AI agent waiting for a purchase approval should treat pending as a state with its own rules. Tell the user what's being decided and by whom, keep the quote current, and stop at a deadline set in advance. Accept a decline as a normal result, and stay on the one payment path it was given, whatever the task says about finishing.

Start with the message to the user, because it costs one line of code and removes most of the confusion. A good waiting message names the purchase, the amount, who's deciding and when the agent will give up. For example: "Waiting for Dana to approve 240 euros at the hotel. I'll hold until 15:10, then release the room and tell you." A spinner tells the user none of that.

Store the pending purchase somewhere that survives a restart. A wait measured in minutes will outlive a deploy or a context window, and an agent that forgets it asked will ask again.

How do you keep a quote fresh while the approval is pending?

Prices and availability move while a person decides, so an approval for 240 euros is good only while the purchase still costs 240 euros. Re-check the quote on a schedule during the wait, and once more right before the charge. If the price has moved, stop and ask again with the new amount. Charging 265 euros on a 240-euro approval spends money nobody agreed to.

Find out whether the seller will hold the price for you. Where a checkout keeps a basket or a fare for a fixed time, the end of that hold is the natural deadline for your wait.

How long should an agent wait before it times out?

Set the timeout from the purchase, and set it before you send the request. The right deadline is whichever comes first: the moment the quote expires, or the moment the result stops being useful to the user. A hotel room for tonight might get 50 minutes. A software renewal due next week can wait until tomorrow morning. Put the deadline in the request so the approver sees it.

When the timeout fires, cancel the pending request so nobody can approve it an hour later against a stale quote. Then release anything you were holding and tell the user what happened.

A hotel booking that waits 40 minutes

Say your agent is booking a hotel room for a user's trip tonight. The room is 240 euros and the cap per purchase is 150, so the purchase goes to the user's manager for approval.

At 15:01 the user has a booking reference.

Now change one line. At 14:50 the room goes to 265 euros. The agent cancels the pending request and sends a new one for 265, so the manager's answer applies to the current price. Change another: nobody answers by 15:10. The agent cancels the request, tells the user the room wasn't booked, and offers the two rooms under 150 euros that it could buy on its own. AI travel booking agents and payments has more on bookings.

What should an agent do when a purchase is declined?

A declined purchase is a normal outcome, and the agent should handle it the way it handles "out of stock." Report what was declined and by whom, then offer the user a next step such as a cheaper option. The one thing the agent must never do is try the same purchase again with a different card or another payment method.

The retry is tempting because of how tasks are written. "Book the room" reads as a goal, and a model given a goal and an obstacle may look for another route. If a second card sits anywhere in its reach, in an environment variable or an old tool config, the agent has just overruled the person who said no. Give the agent one payment path.

Where Shatale fits

With Shatale your customer sets a monthly budget and a cap per purchase in the Shatale app. Inside those, your agent pays on its own, unless the service asks for their approval or the price moved before the charge. Over them, the purchase is blocked or sent to your customer to approve or decline, with the merchant, the amount and the agent in front of them.

Your agent gets the details of the Shatale card for that one purchase after it clears, and it never gets your customer's own card. You integrate through an API or the MCP server. The waiting behaviour in this article is still yours to build.

FAQ

How long should an AI agent wait for a human to approve a purchase?

An agent should wait until the quote expires or the result stops being useful to the user, whichever comes first. Choose that deadline before sending the request and show it to the approver. A same-day booking might wait 50 minutes, and a renewal due next week can wait overnight. At the deadline, cancel the request and tell the user.

Should an AI agent retry a declined purchase?

An agent can retry a network error on the same payment path with the same idempotency key. A purchase that a person declined, or that a spending limit stopped, is finished, and a new attempt needs a new approval request. Retrying with a different card overrides the decision the approval step exists to protect.

What should an AI agent tell the user while a purchase is pending?

The agent should name the purchase, the amount, the person deciding and the time it'll stop waiting. For example: "Waiting for your manager to approve 240 euros at the hotel, holding until 15:10." Send an update if the price changes or the deadline passes, and confirm with a booking or order reference once the purchase goes through.


When to require a human approval covers which purchases should wait for a person in the first place. You can simulate a purchase in guest mode, which needs no key and makes no payment: npx shatale-mcp-server.

Shatale is the control layer for AI-agent payments. Its authorization architecture is the subject of European patent application EP26194994.5 (filed; priority 28 July 2026).