Agentic Commerce
Agentic commerce refers to purchases that a computer program carries out independently on behalf of a human: it searches for products, compares offers, and completes the purchase. The human only sets the goal and the boundaries, such as a budget.
Normally, online shopping works like this: you type something into a search box, click through offers, add something to your cart, and pay. In agentic commerce, a program takes over these steps for you. You just tell it what you want — for example, waterproof hiking boots in size 43 for no more than 120 euros. The program then searches shops, compares prices and delivery times, and can even submit the order itself. Such programs are called agents because they don’t just respond — they carry out actions independently. The human remains the one giving the instructions, but is no longer present for every click.
What changes for shops and customers
Classic online retail is built around humans looking at pages. That’s why there are product photos, discount banners, customer reviews, and elaborately designed homepages. An agent isn’t impressed by any of that. It extracts the raw data: price, availability, shipping costs, return policy. If agents handle a large share of purchases, advertising loses effectiveness — and price becomes more important.
For retailers, this is an uncomfortable prospect. They fear becoming interchangeable rows in a spreadsheet. At the same time, a new race is emerging: whoever makes their product data easy for agents to read will be found more often. Some providers are already setting up dedicated interfaces — technical access points designed specifically for programs rather than for humans.
For customers, the appeal lies in saving time, especially for tedious repeat purchases. Laundry detergent, printer cartridges, cat food — hardly anyone wants to spend twenty minutes comparing options for these. The catch is control: an agent can order the wrong thing, and then the money is gone. That’s why the real point of contention is how much decision-making freedom you grant it.
The path from instruction to order
It starts with an instruction given in plain language. A language model — a program that understands and generates text — breaks this request down into subtasks. “Cheap hiking boots” becomes search terms, filters, and a price ceiling. The agent works through these steps one after another, checking after each step whether it has moved closer to the goal.
There are two ways to access shops. One: the agent navigates a website like a human, clicking buttons and filling out forms. This works everywhere but is slow and breaks as soon as a shop changes its layout. The other way is interfaces through which a shop provides its data directly in machine-readable form. This is more reliable, but requires the retailer to cooperate.
The trickiest part is payment. An agent must not simply hand over credit card details. That’s why payment providers are working on methods in which the agent receives a limited authorization token: valid for one merchant, one amount, one time period. Think of it like a voucher that is valid for exactly one purpose only. If the agent falls for a fraudulent site, the damage is capped.
Where agentic shopping shows up today
The best-known examples are found in chat assistants. With some providers, you can now search for and buy products directly within the chat window without opening the shop. Major retail platforms are testing their own shopping assistants that answer questions like “Which printer fits my laptop?” and immediately present an offer. In most cases, the human still has to make the final click themselves.
In business news, the term mainly appears in connection with payment providers and search engines. Both fear that agents will insert themselves between them and the customer. Whoever controls the agent will, in the future, control access to the buyer — and with it, a great deal of money. Regulators are also taking a closer look, for instance at the question of who is liable for a bad purchase.
A common misconception is that agentic commerce is simply a better price comparison. The difference lies in the action: a comparison site displays results, an agent makes a selection and executes it. It is precisely this autonomy that makes the technology useful and risky at the same time. How much of it you hand over is best decided consciously — not as an afterthought.
