Visa and OpenAI Created an AI-Agent Shopping System: How It Works and Whether to Trust It

Visa and OpenAI announced a strategic partnership that will allow AI agents to independently conduct financial transactions — without human involvement at each moment of purchase. This is the first case in which a world-scale payment infrastructure is officially integrated with a commercial AI-agent system.

How the System Works

Instead of an AI agent using the actual user's card number, Visa issues it a virtual token — a special digital identifier with clear limitations. The agent can spend only a strictly defined amount, exclusively with authorized merchants and only on allowed product categories. Technically, this is a continuation of Visa's years-long tokenization strategy: real card data is never transmitted to third parties. The principle of operation is similar to Apple Pay or Google Pay — except instead of tapping a smartphone to a terminal, an AI makes the decision based on pre-set instructions. For the merchant, the transaction looks ordinary — they don't know that an agent initiated it.

An important detail: the user themselves sets the agent's "mandate" — the maximum amount, list of allowed categories and specific stores. The agent cannot exceed these boundaries even in theory: the token simply won't work.

What the Shopping Agent Can Do

The new system was created for so-called "agentic commerce" — when AI independently performs a chain of tasks without constant human supervision. Here are examples of what the agent will be able to do within its granted authority:

• order groceries from a dwindling list
• book flights or hotels based on specified price parameters
• purchase and renew service subscriptions
• compare prices on marketplaces and select the best option
• pay bills on schedule without reminders

The user sets the rules once in advance, and the agent acts strictly within their bounds. According to the companies, OpenAI is integrating Visa API into its AI agents — first and foremost into ChatGPT Operator functionality, and the ability to make purchases will be rolled out to users gradually.

Security and Trust Problems

Cybersecurity experts point to several fundamental vulnerabilities. The first is prompt injection: if the agent visits a malicious website, attackers can "embed" a hidden instruction to make an unwanted purchase. The second is permission management complexity: most users aren't accustomed to thinking about what exact authority they're granting the system, and risk giving the agent too broad a mandate.

"The question is not whether AI can pay.

The question is whether the user understands well enough what exactly they've allowed the agent to do," warn financial security specialists.

Visa states that in case of an unauthorized transaction, the same buyer protection mechanisms apply as in standard credit card fraud. However, the responsibility for correctly setting limits lies entirely with the user. It remains unclear exactly how the dispute process will work if the agent technically acted "according to the rules," but the outcome didn't match the human's intention.

What This Means

The Visa and OpenAI partnership is not an exotic experiment, but the first standardized step toward embedding AI agents into global financial infrastructure. Similar agreements are already being discussed by other payment operators and AI companies. For businesses, this is a signal: a significant share of transactions will soon be initiated not by people, but by programs. For ordinary users — a new type of convenience with a new type of responsibility: you need to learn to properly set boundaries for digital assistants to whom you're entrusting your money.