Agentic commerce’s real battle is authorization

The central problem in agentic commerce is not whether software can browse, compare products, or complete a checkout. That capability has existed for some time. The harder question is whether a business can prove that an agent was authorized to spend, stayed within defined limits, and acted according to the consumer’s instructions.

That shift is moving the focus away from increasingly capable shopping assistants and toward payment networks, authorization protocols, and identity systems. The companies advancing the infrastructure are not only AI labs. Payment providers are building the controls that could determine how agent-led transactions work.

Visa, Mastercard, and Stripe build agent payment controls

At its annual conference, Stripe introduced wallets that agents can spend from and announced a partnership allowing businesses to sell directly inside AI search and chat environments. Its approach can issue single-use payment credentials for individual tasks, rather than exposing a customer’s card details to an agent.

Recommended reading

Tim Cook gives John Ternus a final Apple endorsement

Visa introduced an early protocol designed to help merchants distinguish a legitimate AI agent from a bot. Mastercard announced a framework for registering and authenticating agents before a transaction is permitted.

Although the implementations differ, the basic sequence is similar:

• Register or identify the agent.
• Verify the user’s intent and authorization.
• Release payment credentials only after trust requirements are met.
• Create evidence that can support a dispute if the transaction is challenged.

The missing capability was therefore not autonomy. It was a reliable way to establish who authorized a purchase, what limits applied, and whether the agent followed them.

AI shopping traffic is already changing

Adobe research from early this year found that AI traffic to US retail sites grew 393% year over year in the first quarter. Visitors referred by AI converted 42% better than non-AI traffic. One year earlier, AI-referred traffic converted 38% below standard channels.

Those figures help explain why payments companies are increasingly talking about authorization, intent, and trust rather than simply autonomous purchasing. The key questions are practical: how can a legitimate agent be distinguished from a malicious one? How can a business confirm that a consumer approved the purchase? How can it verify that the agent did what it was instructed to do?

These are identity and trust problems involving AI, not purely AI capability problems.

Payment protocols could set commerce rules

The infrastructure may look like back-office plumbing, but it will define the operating rules for agentic commerce. Each provider’s decisions about what qualifies as a verified agent, which information proves intent, and how disputes are handled can determine who participates and under what conditions.

An assistant may influence what a consumer wants to buy, while a payment network controls whether and how that purchase can be completed. Those are separate forms of control. The company that controls execution can capture the economics even when another system influenced the original decision.

For merchants, the immediate operational issue is evidence. If a transaction is disputed, can the business demonstrate what was authorized, by whom, and within which constraints? Visa, Mastercard, and Stripe are all working toward answers, and the first widely adopted approach could become the rulebook for other participants.

Agent-ready payments are only one layer

Routine reordering and B2B procurement are likely to adopt autonomous purchasing before more considered or emotionally driven purchases, where consumer trust may take longer to develop. But payment readiness alone will not make agentic commerce work.

Businesses also need a commerce stack that agents can access and interpret, including:

• Product data
• Pricing
• Inventory
• Checkout

If those systems are not structured and accessible, payment authorization will be only one of the company’s problems. The infrastructure now being built by payment providers will influence not just how agent-driven transactions are processed, but which businesses agents can find, which purchases receive trust by default, and which are diverted to friction or manual review.

The emerging contest is over ownership of the infrastructure of trust—not simply over whether agents can act autonomously.