A card number inside an agent's context is worth whatever that card can spend. Put your customer's company card with a 5,000-euro limit into a prompt or a config file, and you've put 5,000 euros of their money there too, for anyone who can read it.

Before a demo with a pasted-in card becomes a product your customers use, settle four questions and write the answers down. They're the questions your customers' finance people will ask you.

What is the card number worth if it leaks?

A leaked card number is worth the card's remaining limit for as long as the card stays open. Work that figure out for the card you plan to use. A customer's shared company card with a 5,000-euro monthly limit puts 5,000 euros of theirs at risk every month. A card capped at 30 euros for one purchase puts 30 euros at risk, once.

Then assume it leaks. Your agent reads web pages and tool output written by strangers, and any of it can carry instructions asking the agent to repeat what it knows. The prompt injection threat model for payments walks through the routes.

You can make the figure smaller by lowering what the card can spend or by shortening the time the agent holds it. The strongest answer does both: the agent holds card details for one purchase at a time, on a card capped at that purchase.

Who set the spending limit, and where is it enforced?

A spending limit counts when your customer chose it and the card system enforces it at the moment of the charge. A limit written into the agent's prompt is a request to the model, and text the agent reads later can talk the model out of it. Write down who chose the figure, where they set it, and who's allowed to change it.

Ask the same about the limit's size. A limit inherited from the card was set for a different job. Say your agent renews domains and tops up API credits for a customer, and its largest normal purchase is 40 euros. A cap per purchase of 50 euros and a monthly budget of 300 fit that job. The 5,000-euro limit fits the people who used to share the card.

Check who can raise it, too. If an engineer on your team debugging a failed checkout at six in the evening can lift your customer's cap in a config file, the limit is as strong as that engineer's patience. Why a spending limit has to be enforced at authorization has the detail.

What happens to a purchase over the limit?

An over-limit purchase can end two ways: a flat decline, or a request sent to your customer, who can approve or decline it. Choose before the first one happens. A flat decline with no path forward stalls the task, and after a few stalled tasks, someone is tempted to raise the limit, and a raised limit stops less.

Settle the details of the approval path now. Ask your customer to name the person who gets the request, and a second person for when the first is on a plane. Make sure that person sees, at minimum, the merchant, the amount and which agent is asking. Then decide how long a request stays open, because a request nobody answers is a decline that takes longer.

Say your agent tries to buy a 199-euro annual plan against a 50-euro cap. With an approval request, your customer sees the merchant and the 199 euros on their phone and decides in ten seconds. When to require a human approval covers where to draw that line.

What does your agent do while it waits for a person?

An agent waiting on a purchase approval should tell its user what it's waiting for and give up cleanly after a set time. Approvals take as long as the person takes, and a payment call in your code may assume seconds. Test the wait before launch: send an approval request, leave it unanswered for an hour, and read what the agent did.

Watch for two failures in that test. The first is the retry. An agent that treats a pending approval as an error may attempt the purchase again, and you get a second request for the same thing. The second is the detour. An agent told to "complete the booking" will look for another way to pay, and if any other card is within its reach, it'll use it.

The fix for the detour is one payment path, with no other card details anywhere the agent can read.

Decide what a decline means to the agent as well. A person saying no to an approval request is a normal result, and the agent should report it to its user and carry on.

How Shatale answers them

Every purchase runs on a Shatale card assigned to that purchase alone and capped at its amount. Your agent gets the details of that card for the one purchase after it clears your customer's policy or approval, and it never gets your customer's own card. When the purchase ends, the card is released and isn't handed to another agent until it's cleared to go back in the pool.

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. Each decision lands in a tamper-evident audit log, per agent.

The fourth question stays with you on any provider. What your agent does while it waits is your code.

FAQ

Is it safe to give an AI agent your customer's company credit card?

A customer's shared company card gives the agent the card's full limit for as long as it holds the number, and anything that can read the agent's context can read the card. A safer setup gives the agent a card capped at one purchase's amount, with limits your customer chooses and the card system enforces.

Can you put a spending limit in an AI agent's prompt?

You can write one there. A prompt limit is an instruction to the model, so text the agent reads later can argue it away, and nothing stops the charge when that happens. Set the limit where the card is authorized, and treat the copy in the prompt as a courtesy that helps the agent plan.

Who should approve an AI agent's over-limit purchases?

Your customer, as the person who delegated the spending, with a named backup for when they're unreachable. An approval request should show the merchant, the amount and which agent is asking, so the approver can answer in seconds. Agree in advance how long a request stays open, and have the agent report an unanswered request to your customer as a decline.


When to require a human approval goes further into the third question. You can run a simulated 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).