The Agentic Commerce Checkout Gap Is an Infrastructure Problem

An agent can only buy what it can retrieve. That sounds obvious until the path is mapped end to end: a robots rule blocks the page, a login wall hides the price, a bot challenge rejects the session, checkout needs a human gesture, or payment authorization asks for a scope the agent cannot prove.

Each wall may be reasonable in isolation. Together they create a storefront that looks open to a person and closed to an agent. The result is a false diagnosis: teams blame the agent for abandoning the purchase when the infrastructure never exposed a safe path.

The fix is not to remove every control. It is to identify which controls distinguish abusive automation from legitimate delegated shopping and expose a supported path for the latter.

The market has already shown that an agent does not need to finish inside the assistant's interface to create value. OpenAI ended its Instant Checkout experiment in March 2026 after reporting roughly three-times-worse conversion in the in-chat flow, while Adobe reported a 42% conversion lift for AI-referred traffic in its measurement.

That changes the merchant mandate. Optimize the owned storefront for an agent visit: give it accurate facts, preserve user constraints, make the handoff explicit, and keep the final order and payment record under a system the merchant controls.

In-chat checkout is a channel tactic. A storefront that can survive delegated retrieval is infrastructure.

McKinsey's corrected estimate puts the global agentic-commerce opportunity at roughly $3–5 trillion by 2030, with up to $1 trillion in US B2C opportunity. The range is a scenario, not current revenue. It is still large enough to make the access problem a board-level capability rather than a speculative feature.

The mistake is to treat the market figure as permission to skip the mechanics. A bigger opportunity increases the value of a clean control surface; it does not make a broken checkout more usable.

Every merchant should ask what an agent can see without credentials, what it can verify with credentials, what it can propose, and what it can commit. Those are different product experiences and different liability surfaces.

Audit six checks in sequence: product feed completeness, crawl and retrieval policy, identity and bot controls, checkout interaction, payment authorization, and evidence. Record the agent's exact prompt, the source pages it saw, each policy decision, and the point of failure.

Use a test catalog with edge cases: unavailable inventory, bundles, returns, subscriptions, regulated goods, address changes, and a price that changes between retrieval and checkout. A system that works only on the happy path is not agent-ready; it is demo-ready.

Keep the outcome legible to a human. A user should see what the agent intends to buy, from which merchant, under which terms, and which final action still requires their approval.

Publish the facts an agent needs in structured and visible forms. Put constraints next to offers. Make policies explicit instead of relying on interaction design to reveal them. Keep machine-readable fields synchronized with the inventory and pricing systems that actually settle the order.

Separate browsing from authority. Let the agent read more than it can change, and ask for confirmation before a purchase, account mutation, or payment. Provide a receipt that records the delegated instruction, final product, price, policy, and approval.

Agentic commerce will feel seamless to the shopper only when the underlying systems are opinionated about boundaries. The smooth experience is the result of strict infrastructure, not the absence of it.