What Is Agentic Commerce? AI Agents That Shop and Pay (2026)

If you've noticed AI assistants increasingly able to "just buy it for you" instead of only recommending what to buy, that's agentic commerce. The rails were laid in 2025 — four frameworks shipped between April and October of that year, from different layers of the stack: the card networks (Visa, Mastercard), the open-protocol layer (Google), and the processor and developer-tooling layer (Stripe). 2026 is the year that infrastructure started carrying real consumer volume rather than the year it was built.

How Agentic Commerce Actually Works

The core technical challenge agentic commerce had to solve wasn't AI capability — models could already describe what to buy — it was trust and authorization: how does a merchant or bank know a transaction initiated by an AI agent is actually authorized by the real cardholder, and not a compromised or malicious agent acting without permission?

Both major card networks solved this with a similar underlying approach, scoped, tokenized credentials issued specifically to AI agents, rather than handing an agent a consumer's actual card number:

Visa Intelligent Commerce combines scoped, tokenized credentials issued to AI agents with behavioral and issuer-side authentication built specifically for machine-initiated payments, plus direct integrations with major LLM platforms (Anthropic, OpenAI, Microsoft).

Mastercard Agent Pay uses Agentic Tokens, an extension of Mastercard's existing Digital Enablement Service (MDES), letting verified AI agents transact on a consumer's behalf. As of this writing, it's live in Malaysia and Singapore, with expansion across Asia planned.

In both models, the agent never holds a consumer's raw card number — it holds a scoped, revocable token tied to specific permissions (spending limits, merchant categories, time windows), so a compromised or misbehaving agent has a bounded blast radius rather than full account access.

The Adoption Timeline

Visa's own timeline, per its APAC product leadership, put commercial deployment of personalized, secure agent-driven payments as achievable starting in the first quarter of 2026, with Visa projecting millions of consumers completing AI-agent purchases by the 2026 holiday shopping season specifically. That's an aggressive timeline for a genuinely new payment paradigm, reflecting how much groundwork the card networks had already done on the authentication and tokenization side before the public-facing rollout.

Why This Matters Beyond Retail

Agentic commerce is a useful lens for a broader pattern in agentic AI generally: the technology to have an agent decide what to do has consistently outpaced the infrastructure to let it act safely and accountably. The same trust problem — how do you let an autonomous agent take a consequential, hard-to-reverse action on your behalf without creating unacceptable risk — shows up in agentic commerce, in enterprise AI agent guardrails, and in the broader agent governance and Guardian Agent trend covered elsewhere on this blog. Payments happened to be one of the first domains where a mature, purpose-built solution (scoped tokenization) reached production at scale, and it's a reasonable template for how other high-stakes agent-action domains will likely evolve.

What This Means If You're Building Agentic AI Systems

Even if you're not building a shopping agent specifically, the agentic commerce pattern is worth studying as an architecture reference: scoped, revocable, purpose-specific credentials instead of broad standing access; explicit authentication built for machine-initiated actions rather than retrofitted human-authentication flows; and clear accountability boundaries (the token defines exactly what the agent was authorized to do) so a failure or compromise has a contained, auditable blast radius. These are the same principles any production agent handling consequential actions — not just payments — should be designed around.

Frequently Asked Questions

Is agentic commerce safe? The major card networks have specifically built agentic commerce around scoped, tokenized credentials rather than giving AI agents direct access to a consumer's actual card details — meaning a compromised or misbehaving agent has bounded permissions (spending limits, merchant categories, time windows) rather than full account access. As with any new payment technology, real-world safety will be proven out over time as adoption scales, not guaranteed by the design alone.

What's the difference between Visa Intelligent Commerce and Mastercard Agent Pay? Both solve the same core problem — authorizing AI agents to transact on a consumer's behalf — using broadly similar tokenization approaches. Visa Intelligent Commerce integrates directly with major LLM platforms (Anthropic, OpenAI, Microsoft) and combines tokenized credentials with behavioral authentication; Mastercard Agent Pay uses Agentic Tokens built on its existing Digital Enablement Service infrastructure and, as of this writing, is live in specific Asian markets with broader rollout planned.

Can AI agents already make purchases for consumers in 2026? Yes, in a growing but still-expanding capacity — Visa has projected millions of consumers completing AI-agent-initiated purchases by the 2026 holiday season, and roughly half of U.S. shoppers report already using AI in some part of their shopping process, though fully autonomous agent-completed purchases are still an emerging rather than default behavior.

How is agentic commerce different from a shopping chatbot? A shopping chatbot recommends products or answers questions; agentic commerce specifically involves the AI agent completing the transaction itself — finding the product, confirming details, and executing payment — using purpose-built, authenticated payment credentials rather than a human clicking through checkout after getting a recommendation.