Agentic commerce is already happening. Here’s how your store should respond

11 Sep 2026 • Technology & Digital Development

Agentic commerce gets talked about as a new sales channel you can just add to your store. It isn’t. It means an AI agent does the shopping instead of a person, and that changes how your store gets found and judged. This article explains what that actually looks like, and what it requires from you right now.

What is agentic commerce?

The same purchase looks different depending on which side you’re on.

As a buyer, you’d tell an AI assistant what you need, say, “order more of the coffee I usually buy, as long as it’s under 15 euros.” The assistant lives in something you already use: a chat app, a browser extension, a voice assistant. It finds the product wherever it’s sold and completes the purchase without you ever visiting a website.

Now flip it around to the seller’s side, which is what this piece is about. That same assistant shows up at your store: not a person browsing, but a program reading your product data, checking price and stock, and trying to complete a purchase, all without a human loading your page. You can’t control it, because it belongs to the customer. Your job as the seller is to make sure that agent can read your product data correctly, and actually complete a purchase, without a person ever being involved.

There’s also a third option: instead of just making your store readable to outside agents, you can build your own assistant into your own site. Picture a furniture retailer whose website lets you ask for a three-seat sofa under 800 euros, then searches its own catalog and places the order for you. Here, the seller builds and controls the agent, not the buyer.

That leaves two separate questions. Can outside agents, the ones your customers already bring with them, find and buy from your store? That part isn’t optional, and it’s already happening. Should you also build your own shopping assistant? That’s a separate, optional project, and it only works once your product data is already solid.

In the latest Logica Research, surveying 3,000 online shoppers, commissioned by PayPal and Commerce, the result is a clear picture of what’s driving agentic AI adoption, what barriers remain, and what shoppers actually need to trust it. It surfaces that 67% of online shoppers are interested in trying an agentic AI shopping tool and that 83% see clear benefits, led by saving time and finding better prices.

What changes when the buyer is an AI agent?

Everything changes once the buyer is an agent. Your customer is now a piece of software. Nobody visits your site to decide. And the purchase itself happens machine to machine, with no person watching either side. For twenty-odd years, online stores have been built for human eyes: good photos, an appealing layout, design that builds trust and nudges people toward the cart. An agent doesn’t see any of that.

An agent can’t be persuaded. Nice banners, mood, and trust-building design have no effect on it. What it responds to is plain, machine-readable fact: your product data, your APIs, and information with no room for guessing, whether the agent pulls that straight from your data feed or just scrapes your page. If your data is unclear or your systems are slow to respond, you’re out of the running before a single human ever looks at your store. And when two computer programs need to agree on price, stock, and delivery, that’s just a systems-integration problem, and those already have standard solutions.

Is this something you need to worry about now, or in a few years? The answer is now. The number of purchases made this way is still small today, but the shift ahead looks dramatic. Gartner predicts that by 2028, AI agents will handle 90% of B2B buying, moving more than $15 trillion through machine-to-machine deals (Gartner 2026). Whatever the exact number turns out to be, the direction is not in question.

Here’s what this shift looks like in practice. OpenAI tested letting people buy things directly inside ChatGPT’s chat window, then changed course: now it’s letting individual online stores build their own buying experience inside ChatGPT instead. Either way, the payments run on shared technical infrastructure OpenAI built together with the payments company Stripe. Google has built something similar: a shared set of rules that let an AI agent prove it has a person’s permission before it spends their money. Over 60 payment and technology companies have already signed on to use it. Visa and Mastercard are each building their own version of the same idea. None of this is finished or fully settled yet, but it’s real and running today, not just a plan on paper.

Sharon Gee, Senior Vice President of Product for AI at Commerce, put the structural shift bluntly: “The traditional funnel had four steps: awareness, consideration, evaluation, purchase. And, increasingly, agentic commerce has two: intent and transaction.”

That’s why being found is more important than ever, and that starts with your product data.

Why product data quality decides AI discoverability

AI shopping assistants make decisions based on the signals available to them. Every attribute you provide is a signal that helps an agent match your product to a shopper’s intent, and every gap is a reason for the agent to choose something else.

Consider what an agent needs to confidently recommend a product:

• Complete attributes so the agent can filter and match against specific shopper criteria such as size, compatibility, material, or use case.
• Clear, descriptive content that explains what the product does and who it’s for in language an LLM can parse and reason about.
• Consistent taxonomy and categorisation so the product lands in the right consideration set rather than being miscategorised or missed.
• Structured, machine-readable formatting that lets agents extract meaning reliably instead of guessing from free text.

When these elements are in place, your products become candidates the agent can confidently surface, compare, and recommend. When they’re missing, even a great product becomes invisible at the exact moment a shopper is ready to buy.

Nothing here needs to be invented from scratch

Most people talk about agentic commerce as if it’s something completely new, but we see it differently: none of this needs to be invented from scratch. A shopping agent is an AI agent like any other, and there’s already a documented, well-known set of patterns for how to build one. Agentic commerce is what you get when you apply those same patterns to buying and selling.

Take one pattern and see what it actually does. Once an agent has filled a cart, it doesn’t just hit buy. First it checks the order against the rules its owner set: is this the right product, does the price stay within budget, is this the seller the customer approved. If any check fails, or the price goes over an agreed limit, the agent stops and asks the person for a decision. That one moment is really three building blocks at once: the agent double-checks its own work, it stays inside set limits, and it hands the decision back to a human when it needs to. The same applies elsewhere in the process. Pulling product data and placing the order through an API are both agents talking to systems and to each other. And fixing an order that failed partway through is just standard error-handling and recovery.

But knowing the building blocks doesn’t do the work for you. They give you a shared vocabulary and a starting point, nothing more. The real work is applying them to the specifics of buying and selling: who’s responsible when something goes wrong, how you stop your product data from being manipulated, how an order gets fixed when it breaks. These are familiar problems. They’re just not solved yet, which is exactly why the questions below matter.

Three questions worth settling early

The hard questions start the moment an agent actually places an order and money changes hands: who’s responsible if the agent makes a wrong purchase, how you protect your store from agents being manipulated, and what happens if a purchase fails partway through, for example the payment goes through but the order never confirms. These are implementation details that depend heavily on your own setup, not something a platform can fully answer for you out of the box. But getting them right is exactly what decides whether people trust agent-driven buying at all.

1. Who’s liable when an agent buys wrong?

Say a customer lets an agent buy things for them within a budget, and the agent orders the wrong product or gets a rule wrong. Who’s on the hook: the customer, whoever built the agent, or you, the merchant? Right now, in practice, that responsibility usually lands on the merchant, unless the customer’s permission to the agent was written down very precisely. That’s why you need a clear, written record of what the agent is allowed to do, and a clear point where a human has to say yes.

One thing you can set up today: a fixed spending limit, with a person required to approve anything above it. Shared rules for how agents are allowed to pay are starting to appear, like the payment infrastructure OpenAI built with Stripe, or Google’s AP2 standard, but they’re still early. Companies are testing them, not yet running them at full scale, and none of them fully settle the liability question on their own. You still need to work out who’s responsible in your own setup, case by case.

Your product content can now be used against you

An AI agent doesn’t just read your product descriptions, reviews, and other text as information. It can also read them as instructions. That opens up a new risk called prompt injection: someone hides a command inside product text or a review, and it tries to trick the buying agent into doing the wrong thing. Basic safeguards for this, like checking incoming data, keeping trusted sources separate from ones anyone can write to, and limiting what the agent is allowed to do on its own, aren’t a nice extra. They’re a basic requirement. If customers can leave reviews or other content on your site, this affects you directly.

2. What happens when something breaks?

A payment can fail partway through. The agent’s connection can drop, it can get confirmed twice by mistake, or it can end up waiting for a response that never comes. Your systems need to handle a repeated request without charging twice, give clear error messages, and be able to cleanly undo an order that didn’t finish. None of this is new; it’s ordinary, well-understood system design. But it’s not optional, because an agent isn’t going to call your customer service line to sort out a mess.

3. How do you track sales that happen without a website visit?

The old way of tracking a customer’s journey through your site stops working. You can’t fine-tune a page-by-page path when software is doing the walking, and your usual analytics tools won’t even show you that an agent was there. You need different things to measure: can an agent actually work with your product data, does the purchase go through from start to finish, and does your store get mentioned when AI tools answer shopping questions.

Two of those you can start measuring right away. Build a test agent of your own and send it to buy from your store, then see whether it finds the right price, stock, and product variant, and where it gets stuck. You can also regularly ask the popular AI assistants the same kind of shopping questions a customer might ask, and check whether your store comes up, what it says about you, and who else it mentions. That’s a simple, repeatable check you can start this quarter without building anything new.

Where you should start as a store owner

Don’t start by bolting a flashy shopping assistant onto your homepage. That’s starting in the wrong place. The work that actually matters doesn’t show on the page at all: making your product data something a computer can understand without having to guess. That same work is also what makes agents able to find you in the first place, whether they get your data by reading your page directly, through a technical connection like MCP (Model Context Protocol), or from a data feed you provide.

Your practical advice for anyone starting their LLM visibility journey:

1. Audit your catalog. Look at what data is going in and how you’re speaking to each channel. Remember, it’s no longer just structured data. It’s unstructured, intent-rich, conversational content too.
2. Build the business case, consumer-first. When leadership asks, “What are we doing with AI?”, lead with a consumer-facing, revenue-driving test before tackling internal change management. Show impact where it hits the top line.
3. Take one small step. Understand your current visibility, pick a bite-size test, prove it out, and expand to the full catalog, then to other brands and markets.

This is also where your platform can carry some of the load for you. For example, BigCommerce exposes catalog, price, and stock through its APIs, so an agent can read what it needs without loading the full page. 

Start by checking your product data against these questions:

• Can a computer get your price and stock level directly, without needing the full page to load first?
• Is your product information, like size, color, and product codes, clear and readable by a machine, with nothing left open to interpretation?
• Are delivery time, shipping cost, and return terms available as clean, structured data, not just buried somewhere in a paragraph of text?
• Do your product details follow a standard format (like schema.org Product, or a structured product feed), so a machine doesn’t have to guess what they mean?
• Have you actually tested this by running an agent against your store to see if it finds the right information?

Once your data is solid, decide what the agent is allowed to do on its own, and at what point a human needs to approve it. Set up a way to measure this from day one, because you can’t improve something you can’t see.

Which platform you use matters less here than it used to, and the agent is the reason. A human would overlook a slow page or a confusing detail. A machine won’t. It judges you fresh every single time, with no memory of your brand’s reputation. So what matters most is the quality of your data and your technical connections, whether you’re on BigCommerce, Shopify, WooCommerce, or something custom-built.

How we see agentic commerce

A lot of the market treats agentic commerce as a passing trend. We don’t. We see it as more of the same work we’ve been doing since 2000: building custom web services for our clients, from the first idea, through building it, to keeping it running for years afterwards. An AI agent is just a new kind of user. What makes a system good hasn’t changed: clear data, reliable technical connections, deliberate limits, and results you can actually measure. We’re an official BigCommerce and Shopify Plus partner, and we also build fully custom solutions. That means we look at commerce across every platform, not tied to one. Once your buyer is a machine, that broad view is exactly what decides who gets chosen to help.

Before you think about agents or platforms at all, there’s one thing worth doing this week: check whether a computer can get your product’s price and stock information without having to guess. Everything else starts from there.

FAQ

What does agentic commerce mean?

Agentic commerce means an AI agent makes a purchase on a person’s behalf: it searches for the product, compares the options, and places the order, inside limits the person set in advance. That limit is often a budget, but it can also be a delivery deadline, a preferred brand, or an ethical rule. The person sets the goal and the boundaries; the agent handles the rest.

How does agentic commerce differ from a chatbot or a recommendation engine?

A chatbot answers questions, and a recommendation engine suggests products, but in both cases a person still makes the final decision and clicks buy. In agentic commerce, the agent also makes the decision and completes the purchase itself, within the authority it’s been given, for example, reordering a supply that’s run out without asking every single time. The difference is who acts: the agent decides on its own.

Should an online store build its own AI agent?

Usually not as a first step. What matters more is making your product data machine-readable, your systems reliable, and your rules clear, so that the agents your customers already use can find your store and complete a purchase there. Building your own agent can come later, once that foundation is solid.

Is agentic commerce safe?

It’s exactly as safe as you design it to be. The biggest risks are unclear liability, manipulation hidden in product data, and handling things when something breaks. All three are manageable with known methods: clear, machine-readable authorisation with a defined approval point, safeguards against manipulation, and systems that can recover cleanly from a failed transaction.

Which platform suits agentic commerce best?

The platform matters less than the quality of the implementation. What matters is machine-readable product data, reliable technical connections, and clear limits on what the agent is allowed to do. You can build all of that well on BigCommerce, Shopify, WooCommerce, or a custom build. The choice comes down to what’s right for your business.

Ready to make your store agent-ready?

An AI agent judges your store on your product data alone — and it does that work whether you’re ready or not. If you want help getting that foundation right, from auditing your catalog to setting the limits an agent is allowed to act within, get in touch. We’ll take it from where you are.