What agentic commerce actually needs before it can work

The idea is straightforward: instead of browsing, you tell something what you need and it deals with the shops. Comparison, questions, negotiation, purchase. The technology to have the conversation exists. The rest of it mostly does not, and the gap is not a technical detail.

Worth separating what is genuinely being built from what is being announced.

What is missing

Authorisation. What is this agent allowed to buy, up to what value, from whom, and how does a merchant verify that without trusting the agent’s own claim? A person can be identified. Something acting for a person needs a delegation model, and there is no settled one.

Payment. Card networks are built around a cardholder present or a merchant taking liability. An autonomous purchase fits neither cleanly. Several schemes for delegated payment credentials exist and are early; which survives is not yet decided.

Liability. The agent orders the wrong thing. Who bears it — the customer who delegated, the agent’s operator, or the merchant who filled the order? Consumer protection law was written for a human buyer, and it does not stop applying because a machine clicked.

Trust and verification. How does a merchant know an agent is acting for a real customer rather than scraping or gaming pricing? How does the agent know the product data is truthful? Both directions need something, and neither has it.

Returns. The mundane one that will decide everything. A machine that buys badly generates returns at machine speed, and returns are already the least automated part of retail.

What that means for what is being sold

A large share of what is currently called agentic commerce is a conversational interface in front of an ordinary checkout, with a person still approving the purchase. That is a real product and a useful one. It is not machines transacting with machines, and describing it that way makes the category harder to evaluate.

The useful question to ask any such product is simple: does anything complete a purchase without a person at the moment of purchase? If not — and today the honest answer is almost always not — then what you are buying is a better front end.

What is worth doing now

The data layer, because it is useful either way.

Machine-readable product knowledge. Structured attributes, relationships, limits and compatibility, available through a documented interface. Improves search and support immediately.

Honest stock and price semantics. What can actually be committed, at what price, with what delivery. An agent asking will not tolerate the approximation a human browser tolerates.

A clear delivery promise. Expressed as data rather than as text on a page.

Auditable answers. A record of what was asked and what was answered, because when the first dispute arrives it will need reconstructing.

None of that requires a standard to settle. All of it makes a retailer better at what they do now, which is the test any preparation for an uncertain future should pass.

What we will not claim

We are not doing agent-to-agent transactions, because nobody credibly is. We are building the catalogue and stock layer that any such thing would need, and we will implement the protocols when there are protocols to implement — and name them, with dates, rather than describing readiness in general terms.