Yes, AI agents can make purchases today. Claude, ChatGPT, and coding agents like Claude Code can find a product, fill a cart, and reach the checkout page on their own. What they cannot do by themselves is pay, because a model has no card. Give one a payment method and the purchase completes. The interesting questions are what people actually use this for, and how they hand over a card without regretting it.
Can AI agents make purchases today?
They can, and the limiting factor is not capability. An agent that can read a web page and fill in a form can complete most online checkouts. Airline sites, grocery delivery, SaaS billing pages, and office supply stores all look the same from the agent's side. It navigates, it enters details, it submits.
The gap has been payment. Agents do not come with money. Until recently the only way to let one buy something was to paste your card number into a prompt or a config file, and most people were right to hesitate. That is changing. The card networks have launched programs for agent-initiated payments, Visa Intelligent Commerce and Mastercard Agent Pay, and Authoryze uses them to hand an agent a single-use virtual card for each approved purchase instead of your real one. Because Authoryze is an MCP server, the agent asks for a card the same way it calls any other tool.
So the honest answer is that agents make purchases today, in a narrower set of situations than the demos suggest, and mostly where the person behind the agent has set up a payment layer they trust.
What people are buying with agents
The pattern so far is that agents buy things that are repetitive, well specified, or tedious to research. A few concrete examples.
Flights and travel. You give an agent dates, a budget, and preferences, and it searches, compares, and books. Travel is where the approval step earns its keep. A $60 ferry ticket can auto-approve. An $800 flight should land in your inbox first, and it does when your threshold is set that way.
Groceries and household restocks. Recurring orders fit well because the list rarely changes and the merchant is always the same. People point an agent at one allowed grocery domain with a weekly cap and let it reorder what is running low. That merchant rule matters. An agent restricted to one grocer cannot be talked into buying somewhere else.
SaaS and developer tools. This is the most common use so far, and it started with developers. A coding agent that already works inside your infrastructure is the natural thing to top up API credits, renew a domain, buy a plugin, or pay for compute when it runs out. These purchases are small, frequent, and easy to bound with a monthly cap and a short list of allowed vendors.
Office supplies. Toner, paper, cables, and the other things nobody wants to spend an afternoon reordering. An agent with a per-transaction limit and an allowlist of two or three suppliers handles it, and the audit trail gives you a clean record for expenses.
What these have in common is that the human still sets the policy. The agent picks the timing and does the work.
What it takes to do it safely
Letting an agent spend money is fine. Letting it spend money without limits is not. Four things make the difference.
The agent should never hold your real card. Cards pasted into prompts end up in context windows, logs, and transcripts. A prompt injection on a web page the agent reads can send them anywhere. Use a single-use virtual card issued per purchase, so what the agent holds is worth one transaction and nothing after. We wrote a full post on how agents pay without seeing your card number.
Rules should be enforced somewhere the agent cannot reach. Telling an agent in its system prompt to stay under $100 is a suggestion. A per-transaction limit, daily, weekly, and monthly caps, and a merchant allowlist checked on a server before any card is issued is a rule. Authoryze evaluates every request server-side, and a request that breaks a rule is denied before a card exists.
You should approve anything above a threshold you choose. Small purchases auto-approve. Larger ones wait for a human. The threshold is yours to set, and it can differ per agent. Your grocery agent and your travel agent should not share a limit.
Everything should be logged. Every request, the rule check, the decision, and the card issuance, so you can always answer what was bought and why.
If you want to see this end to end, the step-by-step guide walks through giving an agent a card in seven steps. The docs cover every option, and the comparison pages show how Authoryze differs from other agent payment tools.
Common questions
Can ChatGPT make purchases?
Yes, once it has a way to request a payment method. With Authoryze connected as an MCP server, ChatGPT calls request_purchase, receives a single-use virtual card for approved purchases, and asks you to confirm each request before it goes out. It needs a Plus, Pro, Business, Enterprise, or Edu plan with Developer mode turned on.
Can Claude make purchases?
Yes. Claude.ai, Claude Desktop, and Claude Code connect to Authoryze over OAuth or an API key. Claude requests the purchase, Authoryze checks your rules, and Claude retrieves the card once the request is approved.
Do I have to approve every purchase?
No. Set an auto-approve threshold and everything under it goes through on its own, as long as it passes your other rules. You can also authorize an automatic spending envelope on an agent's page. Everything over the threshold waits for you.
What stops an agent from buying the same thing twice?
An idempotency key. The agent passes a stable key with each request, and a retry with the same key returns the original result instead of a second card. How to stop duplicate purchases covers this in detail.
Ready to let your agent buy things with limits you control? Create an Authoryze account and connect your first agent today.