Agentic commerce: when AI buys on your behalf.

What is agentic commerce?

Agentic commerce refers to the set of commercial transactions in which an autonomous software agent, an AI assistant, acts on behalf of a human buyer: it searches for a product, compares offers, applies constraints (budget, deadline, conditions), then triggers the order and the payment. The human gives the intent and the limits; the agent carries the execution.

It should be distinguished from two neighboring notions. Predictive AI recommends (recommendation engines). Generative AI produces content (descriptions, answers). Agentic commerce adds a third capability: acting, that is, completing an operation with a third party. It is the shift from self-service commerce (a human browses) to delegated commerce (an agent transacts).

How does agentic commerce work?

An agentic transaction follows four steps: discovery (the agent finds the machine-readable offer), cart (it assembles and validates the contents), checkout (it structures the order with the merchant) and payment (it triggers an authorized settlement). Each step relies on a different technical layer, which the Universal Commerce Protocol framework breaks down into four layers: semantic, agent, transaction, governance.

A key point for merchants: in most current models, the merchant remains the seller of record. The agent does not become the merchant; it relays a purchase intent and a controlled payment token. Control of the catalog, the price and the customer relationship stays on the merchant side. Established

The protocols of agentic commerce

Several standards are emerging separately, each covering one layer. Here is the reading grid, with their backer and public announcement date.

These protocols are not competitors layer by layer: they stack. A merchant can be discoverable via MCP, expose a transaction via UCP or ACP, and accept a payment authorized by AP2. Only UCP and ACP genuinely contest the same transaction layer. Emerging

Concrete examples in 2026

• ChatGPT, Instant Checkout: OpenAI and Stripe enable direct purchase inside the conversation, via the ACP protocol. The merchant remains the seller of record. Established
• Google, AI Mode and the Gemini app: Google and Shopify are rolling out native checkout via the UCP protocol, announced at NRF in January 2026, with Walmart, Target, Etsy, Wayfair and 20+ partners. Established

Generalization (multi-country, multi-vertical, multi-agent) remains a trajectory in progress, not a given. Prospective

What agentic commerce changes for a merchant

To stay purchasable by an agent, the challenge is no longer only to rank well for a human, but to be readable, negotiable and actionable by a machine: structured product data, explicit policies (delivery, return, availability) and an agent-usable endpoint. The readiness audit methodology assesses this state layer by layer, and the use cases by business type illustrate the concrete trade-offs.

Stakes and limits

Three issues remain open. Fraud and authorization: proving that an agent acts on a human's mandate (the purpose of AP2's cryptographic mandates). Governance and neutrality: preventing a single platform from capturing the transaction layer. Traceability: keeping verifiable proof of purchase intent. See the market timeline analysis.

The agentic commerce timeline

The stack came together in under two years: MCP in late 2024 (tool access), A2A in spring 2025 (agents to each other), ACP and AP2 in autumn 2025 (transaction and payment), then UCP in January 2026 (transaction, on the Google and Shopify side). This acceleration is what makes the topic actionable now rather than in three years. Established

Frequently asked questions