Agent commerce: when the buyer is AI
The buyer stops being a person browsing and becomes an agent acting for them. Here is what that breaks in the online store as it exists today.
Someone tells their assistant: buy the 30 lb bag of brand X dog food, delivered by Friday, cheapest option with shipping included. The assistant searches, compares, chooses and completes. Nobody saw your homepage, nobody scrolled to the social proof, nobody read the coupon banner.
That is agent commerce: the purchase in which the buyer is not the person, but a program acting for them. The person still owns the money and the decision — what changes is that they delegate execution, and the delegation comes with explicit limits.
The temptation is to treat this as one more channel. It is not. You add a channel; a new kind of buyer redefines what the store has to be.
The agent is not a badly behaved visitor
The first reflex of any operation that sees traffic which does not look human is to block it. That makes sense: the internet is full of price scraping, inventory bots and card testing. But the buying agent is a different category — it has a human behind it who authorised it, a budget and an intent to pay.
Confusing the two is expensive in both directions. Blocking an authorised agent means refusing a sale. Treating every bot as a customer opens the door to fraud. The distinction is not solved with mouse heuristics; it is solved with identity and authorisation, which is why the emerging payment and credential standards matter more to retail than they look. Until they settle, the practical move is to make your fraud rules distinguish a declared, authenticated agent from an anonymous scraper instead of applying one blunt rule to both.
What the agent reads from your store, and it is not what you designed
A human buyer forms an opinion from the photo, the layout, the feeling of trust the site conveys. An agent forms an opinion from what it can extract and verify:
- what the product is, in attributes with units and types, not in prose;
- whether it is in stock now, not “we usually have it”;
- what it costs delivered to their postal code, shipping and delivery date included;
- what happens if things go wrong — returns, warranty, cancellation window;
- whether the transaction can be completed without a person clicking a button.
None of that is persuasion. The agent is not sensitive to fake urgency, countdown timers or the “last unit” that has been there for three months. Those devices do not merely lose their effect: they become noise that gets in the way of extraction. A product page written for machines is mostly a product page written honestly.
The three layers your store has to serve
What changes can be organised into three moments, and each one breaks differently.
Discovery. The agent has to find the product. Today that runs through search, marketplaces, feeds and, increasingly, direct queries against a structured catalogue. A store whose internal search only works with the exact words in the title is invisible to a natural language query. That is why catalogue and search stopped being a front end topic and became an infrastructure one.
Decision. The agent compares. And it compares against criteria the customer wrote, not the criteria the store would prefer. “Cheapest with shipping included, delivered by Friday” immediately eliminates anyone who cannot quote shipping through an API, anyone without a reliable delivery date and anyone hiding cost until the last step. A store that only reveals shipping after account creation drops out of the comparison without ever appearing in it.
Execution. The agent completes the purchase. This is where most of the friction lives, because checkout as it exists in most markets was designed to defeat a human’s distrust — and a good part of it exists only to prove there is a person on the other side. It is where the store discovers it is technically out of the game.
What breaks in retail as it is today
The persuasion funnel loses traction. Exit pop ups, social proof, manufactured scarcity: all calibrated for a human nervous system. The agent ignores them. What remains is what can be verified.
The page stops being the interface. If the purchase happens through APIs and structured data, the page becomes one of the outputs — important for people who shop by looking, irrelevant for the ones who shop by reading. A store that only exists as JavaScript rendered HTML, with no feed and no endpoint, is simply not machine purchasable.
Price gets genuinely exposed. Comparing stops costing effort. That does not mean the cheapest always wins: it means price is now checked together with delivery date, availability and return policy, under a single criterion. Whoever delivers in two days and accepts returns without drama competes with whoever charges less — as long as they can prove it in data.
Loyalty moves. It stops being brand recall and becomes behavioural history. An agent that hit a late delivery and a stuck return records that, and the record weighs on the next choice. Reputation becomes a decision variable, not a sentiment.
Post purchase becomes a buying criterion. If the agent can track the order, request the return and follow the refund on its own, the cost of buying from you drops. If every exception requires a human phone call, the agent learns to avoid the store. It is the logic of automated post purchase seen from the buyer’s side.
What does not change
This is worth insisting on, because the subject attracts too many promises.
Real inventory is still real inventory. Carrier lead times still depend on carriers. Margin is still margin. No agent fixes a bad operation — it only exposes it faster. The store that sells today despite wrong data, because the human customer shrugs and asks on chat, loses that safety net. The agent does not shrug: it discards and moves on.
That is a good side effect, by the way. Correct data was always an advantage; it was just an invisible one. With an automated buyer, it becomes a measurable revenue difference.
The other side: the store runs agents too
This text looks at the agent as a buyer. But the same technology sits on the inside, and that is how most operations enter the subject: an agent that answers on messaging, recovers carts, qualifies and resolves post purchase. That is the territory of agentic ecommerce, and on the sales side it becomes a pipeline run by agents. There is a useful symmetry here.
Everything you need to do for your agent to work well — a queryable catalogue, authoritative stock, policy written unambiguously, a tool that creates orders — is exactly what makes the store readable to the customer’s agent. These are not two projects. It is the same project, invoiced twice.
Where to start without betting on the future
The trap here is waiting for the standards to settle. While payment and agent identity protocols are still being disputed, there is a list of work that pays off regardless of who wins:
- Product data as data, not as text. Attributes with units, variants with identifiers, measurements outside the image.
- Availability and delivery dates through an API, with an authoritative answer rather than yesterday’s cache.
- Shipping quotes before account creation. If the total cost only shows up at the end, you are not in the comparison.
- A machine readable return policy, with window, condition and cost made explicit.
- Telling a hostile bot apart from an authorised agent in your fraud rules, instead of applying the same measure to both.
None of that is a bet. It is data hygiene that already pays for itself in human conversion — and that, when the buyer is a program, decides whether you exist on its list at all. The underlying discipline is the same one that makes a vertical agent work on your own side of the counter.
What to take from this
Agent commerce is not a technology a store adopts. It is a change in who is on the other side of the purchase. When the buyer is a program, persuasion counts for little and data counts for a lot; the storefront loses its monopoly on the interface; and checkout stops being a barrier against fraud and becomes a barrier against sales.
The practical question is not when this arrives. It is: if an agent tried to buy from my store right now, at which exact step would it get stuck? Anyone who can answer that already knows what to do on Monday.