Agentic commerce
The customer sent an agent.
AI agents are starting to shop for people. The brands they choose will be the ones a machine can read and a person can trust.
For twenty years, the person we designed a shop for was the person in the shop. They arrived, looked around, compared, hesitated and paid, and every screen in between was ours to shape. That assumption is starting to break. More of the looking, comparing and paying is about to be done by an AI agent working for that person: in a chat window, in a browser, or with no interface at all. It will meet the brand somewhere the brand did not design.
That is agentic commerce: software that discovers, evaluates and buys on a customer’s behalf. It is early. Most people are not ready to hand an agent their card and walk away, and they are right not to be. But the conditions for it are already in place, and the gap between what agents can do and what people will let them do is a design problem.
This is how we approach it, as a studio that designs and builds the thing: from the trust underneath it to the first experiment worth running.
Trust is built into the product.
When someone sends an agent to buy for them, they are delegating judgement, and delegation has to be earned. Brand tone will not earn it. Trust in an agent comes from behaviour: it does what it said it would, explains itself at the moments that matter, and makes it easy to give control and to take it back.
So permissions, limits, undo and receipts deserve the design attention the checkout flow used to get. Stripe’s Shared Payment Tokens, part of the Agentic Commerce Protocol it built with OpenAI, are a good model. Each token is scoped to one merchant, capped at an amount, set to expire, and revocable at any moment. The permission is something a person can see and withdraw. Most design for agents looks like this: less like decorating a moment, more like writing a contract people can actually read.
If an agent can’t read it, it isn’t on the shelf.
Digital marketing has spent a decade chasing people around the internet in the hope of catching them in a buying mood. Agents reverse the direction. The customer, or their agent, comes looking, and what matters is whether the brand is ready to be useful when it arrives.
Ready means legible to a machine: complete, structured product data, prices and stock that are true right now, and proof in a form software can check, from specifications and sizing to returns and reviews. The plumbing arrived quickly. OpenAI’s Agentic Commerce Protocol feeds merchant catalogues into ChatGPT. Google’s Universal Commerce Protocol went live in January, with Shopify, Target, Etsy, Wayfair and Walmart among its launch partners. Perplexity runs a merchant programme of its own. An agent can only recommend what it can parse, and a product with gaps in its data quietly drops out of the comparison.
Fig. 02What an agent can put on the shortlist
Shortlist3 of 6
None of this is new to anyone who has done technical SEO. It is the same discipline with higher stakes, and we wrote about the reading side of it in Most readers never run your JavaScript. The difference is that this reader doesn’t only send traffic. It decides.
Let people choose how much to hand over.
The test for any agentic experience is simple to state: does the person come away feeling more capable, or more managed? Most of the difference lies in who decided how much the agent does.
We design participation as a range, not a switch. At one end the agent assists: it finds, compares and explains. In the middle it is handed a task with limits. At the far end it acts alone. People should be able to move along that range from one purchase to the next, and within a single one. Google’s Agent Payments Protocol writes the range into the transaction itself. An Intent Mandate lets an agent search and negotiate; a separate Cart Mandate authorises the purchase; and only once trust exists does a customer let the agent buy with nobody present. The person picks the posture, not the platform.
Fig. 01A range, not a switch
Assists
Finds, compares and explains. The person buys.
e.g. a gift, manual on purpose
Delegated
Handed a task with limits: a budget, a merchant, a date.
e.g. new running shoes
Alone
Acts with nobody present, inside the mandate it holds.
e.g. the weekly shop
Intent mandatemay search and negotiate
Cart mandatemay buy
Some moments should stay manual on purpose. Choosing a gift, or the sofa you will sit on for ten years, is not friction to be removed. It is the part people came for.
A chat window is a thin shop.
Too many conversational shopping experiences feel like a chatbot bolted onto a storefront, and customers can tell. Text is a poor way to judge a colour, a fit or a texture. Seeing and hearing something builds more confidence than reading about it.
The better pattern is for the conversation to direct the interface. As intent sharpens, the experience brings forward the right comparison, photograph, video or swatch, beside what the person was already looking at, and changes in place when they ask a follow-up rather than starting over. That needs a design system that can recompose itself in real time and still look like the brand. It is the same problem as building components with agents: the rules have to be written down well enough for software to follow them with no designer in the room.
Measure the month, not the moment.
Conversion measures the seconds after a click. An agent that buys for someone every week acts over months, and over months it shapes things a conversion rate never sees: a household’s budget, how much a person has to think about, how confident they feel in what they bought, and what all of it costs the environment.
A badly designed agent adds to decision fatigue. A good one remembers that someone prefers things that last, or the lower-carbon option, and is willing to suggest a smaller purchase or none at all. That sounds like bad business until you remember the agent is shopping again next week, and the customer decides whether to let it.
Underneath sits a question every brand in this space will be asked eventually: who does the agent work for, the customer, the merchant or the platform? Once agents control checkout, defaults and rankings compound into real money saved or spent. I would answer that question in the customer’s favour before anyone has to ask it.
Start narrow, on purpose.
No brand gets to agentic commerce in one move. The way through is to agree a direction and a few rules, then run small experiments: one interaction, one level of autonomy, one multimodal pattern, with the evidence deciding what comes next. The rules become a filter for what to test, what to shelve and what to refuse outright.
Amazon’s Buy for Me is a good example of restraint. It launched as a clearly labelled button, with a limited set of brands and the customer confirming every order. Narrow enough to learn from, and narrow enough to trust.
This is where our habit of putting a working prototype on the table early earns its keep. An agent’s behaviour is almost impossible to judge from a slide. Hand someone a prototype that takes an instruction, asks before it spends and shows its working, and the conversation about trust stops being abstract.
Where we would begin.
All of this will look different for every brand, and how fast any of this arrives depends on how comfortable customers get, how mature the category is and how ready the brand is. The first step is the same everywhere. Pick the area you are furthest from, run one focused experiment against it, and let what you learn choose the next one. If you would like help deciding which, tell us where you are.
The customer is still the one deciding. Increasingly, they decide by sending someone.
Asked often.
What is agentic commerce?
Shopping in which AI agents discover, compare and buy products on a person’s behalf, within limits the person sets, through protocols such as OpenAI’s Agentic Commerce Protocol and Google’s Universal Commerce Protocol and Agent Payments Protocol.
How do brands get ready for AI shopping agents?
By making their products legible to machines: complete structured product data, prices and stock that are always current, and proof software can check, such as specifications, sizing, returns and reviews. Then by joining the commerce protocols the agents use.
How do you design trust into an AI shopping agent?
Treat permissions, spending limits, reversibility and receipts as core product design. Let people choose how much autonomy to give for each purchase, have the agent explain itself at key moments, and make taking control back easy.