Agentic commerce: when the shopper is an agent

Agentic commerce is shopping done by an AI agent on a person’s behalf, from finding a product to paying for it. For a UK ecommerce leader, the change is who reads your site. An agent ignores your photography. It reads your product data and tries your checkout, and when a price is ambiguous or a ‘prove you’re human’ check appears, it buys from someone else. So the work is in your data and checkout.

The best-known personal agents have not fully arrived in the UK: Meta’s Muse is currently available only in the United States, and OpenAI’s Dots are not available in the UK. Use the gap to prepare: UK shoppers already use AI to decide what to buy, and open-source agents run anywhere.

What is agentic commerce?

The key word is agent. An agent is software given a goal rather than a script. Ask one for a waterproof jacket within budget that arrives by Friday, and it decides its own steps: which shops to check, which sizes to compare, whose delivery promise to believe and, if you let it, where to pay. That freedom is why it behaves differently from a person on your site. It skips your home page, reads whatever it can parse and gives up on whatever it cannot.

It arrives in three stages, at different speeds:

1. AI-assisted research. A shopper asks an assistant what to buy, then visits and buys themselves. This is mainstream already: 37% of UK consumers have already used an AI assistant for online shopping.
2. Agent-referred visits. The assistant shortlists and sends the shopper straight to a product page, skipping your home page, category pages and most of your merchandising.
3. Agent-completed purchases. The agent picks the variant, fills the basket and pays, with the person approving one step at most. This is still small: in a Bain and ROI Rocket survey of US consumers, only 10% had ever used AI to purchase a product.

Most of the noise is about stage three; most of the money today is in stages one and two. The same work serves all three. As Stripe puts it: just as search engine optimisation made content legible to search engines, agentic commerce will require making your business legible to agents.

Who the new shopping agents are

Meta’s Muse

Muse is Meta’s assistant that can browse and buy for a person. Meta has announced partnerships with retailers, including Best Buy, Gap, Sephora, Walmart, and Wayfair, and Mark Zuckerberg has said Meta expects to profit over time by taking a small fee from transactions. A platform earning a fee on each sale will want merchants to integrate on its terms, and UK retailers are already considering the prospect of sales commissions or the monetisation of placement.

Meta admits that loading full pages, clicking through flows designed for human eyes, and parsing visual layouts is slower and more brittle than it needs to be, and wants businesses to point Muse at the right route: use this flow for bookings, use that connector for order tracking, check this structured source for real-time inventory. The vehicle is the Personal Agent Protocol, an open standard Meta and Sierra are developing with partners including Shopify, Stripe and Walmart.

OpenAI’s Dots

Dots are OpenAI’s personal agents: always-available assistants you can assign work to by chatting through text, Slack, or Microsoft Teams. When ChatGPT browses a site for someone, it uses a browser in OpenAI’s cloud, and OpenAI warns that a website may block ChatGPT from accessing the site or completing your task, even when the same site works normally in the person’s own browser. To you, that is a lost sale that never shows up as an abandoned basket.

Instinct

Instinct is an invite-only agent platform. Its founder says it sees more than 50% of transactions on the platform for travel, and that it is approaching a billion dollars in annual transactions, though TechCrunch notes he did not say how that was calculated. Travel leads because the task is tedious, rule-bound and comparable: dates, fares, cancellation terms. Agents do best where products have clear, comparable attributes, so write your catalogue that way.

OpenClaw

Personal agents were first popularised less than a year ago through OpenClaw, an open-source AI assistant that runs on your own computer and takes instructions through chat apps such as WhatsApp, Slack and iMessage. Nothing stops a UK shopper using it today. Because it runs on the shopper’s own machine, its visits are likely to come from an ordinary home connection rather than a known data centre: harder to tell from a person, and easy for anti-bot tools to mistake for a scraper.

Two ways an agent shops

• Through your pages. The agent loads your site in a browser and clicks through it as a person would. It works anywhere and breaks easily: cookie banners, pop-ups, size pickers drawn as images, delivery dates hidden until a postcode is typed, and anything that asks the visitor to prove it is human.
• Through a structured route. The agent asks a feed, an API or a protocol directly. A protocol is an agreed format for asking what you sell, at what price, whether it is in stock and how to buy it, and getting a reliable answer. The Agentic Commerce Protocol (ACP), Google’s Universal Commerce Protocol (UCP) and Meta’s Personal Agent Protocol are all versions of this.

Most agent visits today are the first kind; the platforms want the second. Make the first work now and keep your data ready for the second, without betting the budget on which protocol wins.

Which companies are using agentic commerce?

The furthest along are the large US retailers partnered with Meta, and they show what breaks. Walmart partnered with Meta to allow Muse to browse its site and make purchases, but technical issues hampered the rollout. As TechCrunch reported it, the website presents a button that has to be clicked to verify the visitor is human, and if that step is interrupted, the check can fail and the agent gets booted out.

Others draw the line more carefully. eBay says its approach is not to prohibit all third-party shopping agents from completing purchases, while restricting unauthorised agents and scraping. Zillow allows AI agents acting on behalf of consumers, but its anti-scraping measures occasionally block some. Most retailers will land here: agents welcome, scrapers not, and the hard part is telling them apart.

Behind the retailers, payment and platform companies are building the plumbing:

• Stripe and OpenAI developed the Agentic Commerce Protocol to define a common language for how agents and businesses transact, including checkout and payment credentials.
• Google describes UCP as a new open standard designed to unify digital commerce.
• Visa is running its Agentic Ready programme, launching first in Europe, including the UK. Its first phase is testing how agent-initiated payments work in real issuer environments, and early issuers include HSBC UK, Nationwide Building Society and Revolut.

Signifyd reports orders from AI agent referrals in Signifyd’s Commerce Network were up 1,247% from a year earlier. Growth that steep usually starts from a small base, so it shows the direction of travel and says little about share of revenue.

Does Amazon use agentic commerce?

Yes, its own, while keeping others out. Amazon runs Buy for Me, which identifies itself and lets brands opt out, and launched Alexa for Shopping, an AI agent that researches products and makes recommendations. Meanwhile Amazon has spent the past year trying to keep outside agents off its site, suing Perplexity over its Comet browser and moving to block shopping agents from Google and OpenAI, and it has now blocked Muse. Its stated principle is that agents buying from other businesses should operate openly and respect service provider decisions about whether or not to participate.

The courts have not simply backed that stance. Amazon won a preliminary injunction against Perplexity, then lost it when the Ninth Circuit ruled that the user, not the AI company, was the one accessing Amazon’s computers under US anti-hacking law. That ruling decides nothing in the UK, but it shows how unsettled the question of who is visiting your site has become.

The lesson for a UK retailer: Amazon can turn agents away because it is the destination. Most retailers are not. If an agent cannot buy from you, it buys from someone else and you never see the visit. The principle is still worth copying: an agent that says who it is, and a merchant who decides on purpose whether to let it in, by agent and by action, rather than leaving it to bot-protection defaults.

What to change when the visitor is an agent

IBM expects agentic commerce to shift competition towards machine-readable product data, availability and fulfilment reliability, rather than brand visibility alone. It is too early to justify a full investment in agents. What pays now is structuring two things so agents can read them: your product information, and the jobs a visitor can do on your site, such as checking stock, buying or starting a return, along with the steps each one takes.

Product data and feeds

Stripe says the agents arriving at merchants want clean and structured data, fast responses and zero guesswork. The reason is mechanical: an agent compares sellers attribute by attribute. Where your data is vague or contradicts itself, the agent guesses, which means wrong orders and returns, or moves on to a seller it can trust.

So the job is to keep pricing, sizing, materials, availability and shipping details consistent across your site, feeds and APIs, and to keep your product feeds, brand assets, return policies, and business contact information complete and current in Google’s merchant platform. AI platforms treat freshness as a trust signal, so a stale price costs visibility as well as orders.

A checklist for your next feed review:

• Every variant has its own identifier (a GTIN, the number behind the barcode, where you have one), URL, price and stock level.
• Attributes sit in fields, not marketing copy: measured sizes, materials, colours, compatibility, what is in the box.
• Delivery cost and dates appear in the feed and on the product page, not only at checkout.
• Returns window, warranty and contact details are published as text and as structured data: labels in your page code, usually in the schema.org format, that tell a machine which number is the price and which is the stock.
• Price, stock, delivery and returns read the same everywhere an agent or an AI assistant might find them: your site, feeds, marketplace listings and comparison sites.

A quick test: take your five best sellers and write down the price, stock in the most popular size, delivery date to a London postcode and returns window, exactly as each appears on the product page, in your Google feed, in the page’s structured data and at checkout. Every disagreement is a point where an agent gets it wrong or drops you from the shortlist.

How agents read and navigate your site

An agent working through your pages reads text and structure far more than visual design. Many sites, travel sites especially, are still designed only for humans. In practice:

• Put price, stock, sizes and delivery in page text, not images, tooltips or hover states.
• Use real buttons and labelled form fields. A size picker built from unlabelled boxes is invisible to an agent and to a screen reader, so accessibility work doubles as agent readiness.
• Send product information in the page itself, not only after a chain of scripts has run.
• Give every product and variant a stable URL, so an agent sent to a red medium lands on a red medium.
• Count what stands between arriving and paying: cookie banners, country selectors, pop-ups, forced account creation. Each is a step an agent can fail.

Then list the jobs an agent might do for a customer: check stock, check a delivery date, buy, track an order, start a return. For each, decide whether agents may do it and by which route. That is the list Meta is asking businesses for, and it is useful whichever platform wins.

Checkout and payments

In our experience this is where agents get stuck first. Captchas and other bot checks block agents, and a ‘prove you’re human’ step adds friction for every human customer as well.

Stripe sees companies designing checkout flows to run without a human in the loop and publishing agent-readable trust signal information like return policies, fulfilment guarantees and contact information. If you later take orders through a protocol, the duties stay yours: under Google’s UCP, as the Merchant of Record, it is your responsibility to perform real-time inventory checks. The merchant of record is the business legally making the sale, so refunds, tax and chargebacks remain yours whoever placed the order.

Payment will change more slowly than the headlines suggest, because shoppers are wary: in one survey, 72% said they would not share card details. Expect agents to pay with wallets and tokens issued by networks and banks rather than typed card numbers, which is what Visa’s testing is about. For now:

• Keep guest checkout and never force account creation before payment.
• List every human check across login, basket and checkout, and decide for each whether a verified agent should pass.

Trust and fraud signals

IBM states the root problem: existing fraud and payment systems were built for human verification, not machine intermediaries. Fraud rules treat speed, automation and data-centre traffic as warning signs, and a legitimate agent shows all three. Signifyd advises you to rework bot rules, velocity thresholds and risk logic so they can separate trusted agents from malicious automation.

Cloudflare treats an agent as verified when it declares who it is deterministically, through a cryptographic Web Bot Auth signature, a published IP list with a stable user-agent, or reverse DNS. A user-agent is the name a browser or bot announces with every request, and anyone can fake one. Web Bot Auth uses cryptographic signatures in HTTP messages to verify which automated bot a request really comes from.

Questions for your bot-protection and fraud providers:

• Can you recognise agents that sign their requests with Web Bot Auth, and treat them differently from anonymous automation?
• Can we set policy per agent and per action: browse, check out, or blocked?
• When a verified agent is challenged or blocked, does it show up in a report we will read?
• For orders taken off our site through a protocol, which of our fraud checks still run? Under the Agentic Commerce Protocol, businesses can choose to accept or decline transactions on a per agent, per transaction, or custom logic basis. Ask whether your provider can run a rule such as: accept orders from agents we have approved, up to an order value we set, and send everything else to manual review.

Being recommended in AI answers

Before an agent buys, an assistant decides whether you make the shortlist, and that traffic is worth having: Adobe found shoppers clicking through to UK retail sites from AI sources were more likely to complete a purchase than those arriving after using online search engines. Shoppers’ trust in the assistants is fragile, though: just 24% trusted recommendations and summaries produced by AI chatbots. Assistants have good reason to favour sellers whose facts they can check and whose facts agree from one source to the next.

Whether you can be bought through an assistant is a separate decision, and the platform makes it. Implementing a protocol does not guarantee a listing: under ACP, each AI platform will manage their own process for how businesses can participate. Before you build for a platform, ask it how merchants are chosen and what it needs from you.

Measuring agent traffic

Most analytics setups cannot tell you how many agents visit. Analytics tags run in the visitor’s browser and filter out known bots by default, so many agents either never run that code or are discarded. The evidence sits in your server and CDN logs (the CDN is the delivery network in front of your site) and your bot-protection dashboard. The signals that separate an agent from a person or a scraper:

• Declared identity. Many agents name themselves in their user-agent; the better ones prove it with a Web Bot Auth signature or a published IP list.
• AI referrals. Visits from assistants such as ChatGPT, Perplexity, Gemini or Copilot are usually people sent by an AI answer. Give them their own channel.
• Behaviour. Agents tend to land deep on a product page, go straight to the basket and fill forms almost instantly. Scrapers request huge numbers of pages and never touch the basket.
• Challenge outcomes. How many sessions hit a captcha or human check, and how many of those leave without buying.

A first quarter:

1. Month one: baseline. Turn on agent and bot logging at the CDN, set up an AI-referral channel in analytics, and record how often your human checks fire and fail.
2. Month two: try it yourself. Have someone attempt real purchases of your top products with the agents available in the UK, such as OpenClaw, and note exactly where each attempt stalls.
3. Month three: four numbers, weekly. AI-referred sessions and their conversion rate against search; identified agent sessions; agent checkout attempts against completions; and verified agents blocked by your own defences.

Watch the last one: it is revenue your own security is turning away.

Where not to spend yet

Adyen is right that agentic commerce is still in flux, with several models emerging at once and interfaces still changing. Where to hold back:

• Bespoke integrations for one agent. A dedicated connection for an agent your UK customers cannot use yet is a bet on one platform’s terms before you know them.
• A separate checkout for agents. Fix the friction in the checkout you have. Adyen’s advice to avoid making agent or platform-specific checkout logic is sound.
• Protocol work as a growth plan. Supporting a protocol is plumbing, and plumbing brings no customers by itself. It earns its keep when a platform you care about asks for it.
• Packaged AI-answer optimisation. If the offer is anything other than better, more consistent product data and clearer policies, ask for the evidence.

Fund the work your human customers benefit from too: cleaner product data, a calmer path from product page to payment, human checks that let verified agents through, and reporting that shows what agents do. If agents arrive slower than forecast, your customers still get accurate listings and an easier checkout, and you still get fewer returns.

Is agentic commerce the next big thing?

Big enough to prepare for now, not yet big enough to bet a budget on. Treat it as one phase of work with a clear end, then a pause until a specific trigger tells you to spend more.

End the first phase when three things are true: your five best sellers show the same price, stock, delivery date and returns window everywhere an agent reads them; no human check between product page and payment turns away a verified agent; and the weekly agent numbers reach whoever owns trading. Then hold the budget.

Reopen it for protocol work or platform integrations on either of two triggers: a platform your customers use launches in the UK and invites merchants in, or agent-completed orders in your own logs grow to the size at which you would staff any other channel.

Expect at least one of them to fire. An IBM Institute for Business Value study found that 45% of consumers already use AI for part of the buying journey, and Fortune reports Mastercard’s prediction that one in 10 shoppers will use a personal AI agent to purchase products or services. Your own logs will show when your customers get there.

If you want a second pair of eyes on where your site stands against that list, our senior builders work inside client teams on data and checkout problems like these: see what we build or start a conversation.