For two years AI agents have been able to plan a purchase — find the flight, build the basket, compare the options — then hand you back to a checkout page to actually pay. Visa and OpenAI just removed that last step.

The deal

Announced today at the Visa Payments Forum in San Francisco, Visa and OpenAI are integrating Visa's payment network directly into OpenAI's products. The result: an AI agent in ChatGPT can complete a real purchase on your behalf, paying with tokenized Visa credentials — not your raw card number, but a secure, scoped token.

For merchants and developers, Visa is providing a standard way to accept agent-initiated payments, which is the unglamorous-but-essential plumbing that makes any of this work at scale.

The guardrails are the point

The obvious worry — an AI spending my money? — is exactly what the design targets. Every transaction runs inside user-defined controls:

• Spending limits — caps on how much an agent can spend.
• Merchant categories — where it's allowed to spend.
• Required approvals — actions that need your explicit sign-off.

On top of that sits Visa's existing real-time authorization and fraud monitoring. In other words, the interesting engineering here isn't "can an AI pay" — it's "can an AI pay safely, within bounds you trust."

The sleeper detail: agents that buy their own fuel

One line points at where this goes. OpenAI's Codex coding agents could eventually use authenticated credentials to buy their own inference, APIs and other services autonomously — within your limits. An agent that can purchase the compute and tools it needs to finish a job is a meaningfully more capable (and more autonomous) thing than one that stops to ask.

Why it matters

This is a real inflection in agentic commerce: AI moving from answering to acting — and acting with money is the highest-stakes action there is. Whoever owns the trusted rails for agent payments owns a huge piece of the agent economy, which is why the biggest payment network on earth is moving now.

The takeaway for everyone else: the question stops being whether agents can transact, and becomes how much you trust the limits around them. Get those right and a lot of friction disappears. Get them wrong and it's a new attack surface. Either way, your AI assistant just got a wallet.