Over the last year, every major AI platform has shipped the ability for agents to buy things on behalf of users. The experience is still clunky, most consumers don't trust it yet, and the error rates are high. But the infrastructure is being built anyway, and the investment behind it is enormous. As I mentioned in my article last week, I think the age of AI Agents doing much more of our online browsing and buying is nearly upon us.

The consumer experience is still rough. There's still a majority of people that would rather shop themselves than have a robot do it. I'm not saying that every transaction will be done by a robot by next year. But things are changing fast, and these AIs are being employed by more and more people to complete the mundane purchases that we previously spent hours clicking around.

Who Has Released What? And When?

Amazon's "Buy for Me" launched with Alexa+ nearly a year ago, letting agents purchase from third-party sites. OpenAI's Instant Checkout followed in September — enabling users buy directly inside ChatGPT, powered by Stripe, with over a million Shopify merchants coming online. In January, Microsoft launched Copilot Checkout, where shoppers find products inside Copilot and buy without leaving the conversation, with every Shopify merchant automatically enrolled. And Google's "Buy for Me" is now live — a shopper finds a product, reviews the details, and confirms, then Google's agent navigates to the merchant's site and completes checkout via Google Pay. The shopper approves the purchase but never touches the retailer's website directly.

These are all US-first rollouts for now, but the protocols behind them are open standards and the infrastructure is being built globally. Real money is already changing hands.

OpenAI's own testing of Operator found a 13% consequential error rate — wrong items ordered, wrong dates set. Reviewers found themselves intervening six or more times to complete basic tasks and asking at what point it's easier to just do it yourself. Google's Buy for Me requires Google Pay, which cuts out most large merchants, and early deployments have hit issues with variant selectors not stabilising before checkout — meaning the wrong SKU gets purchased. Alexa+ has been described as a "buggy mess" by early users. And only 14% of consumers currently trust AI to place orders on their behalf.

So why does this matter now if the agents are still clunky? Because the infrastructure is being laid regardless, and agents will improve quickly. The sites that are easiest for agents to navigate today will be the ones that benefit first as the experience gets better. And Microsoft's early data already suggests that where it does work, shopping journeys involving Copilot are 33% shorter than traditional search paths, with a 53% increase in purchases within 30 minutes. If those numbers hold up, these are high-intent, fast-converting interactions — but only if the agent can actually complete the transaction on your site.

Two days ago, WebMCP landed in Chrome 146. It's a W3C standard co-authored by Google and Microsoft that lets any website expose structured tools directly to AI agents through the browser. One early benchmark reported a 67% reduction in computational overhead compared to agents trying to visually navigate a page. It's early data, but the direction makes sense — it lets agents interact with your site programmatically rather than needing to "see" it the way a human does. This fact, that agents would have to take a screenshot, process it, decide which button to click and then proceed, one painful step after the other is what has slowed computer-using agents down thus far.

Underneath all of this, a set of open standards is emerging that defines how agents discover products, negotiate prices, and complete payments — essentially a shared language between websites and AI agents. And it's moving fast. In January, Shopify and Google launched the Universal Commerce Protocol with over 20 partners including Visa, Mastercard, Walmart, Target, and Stripe. OpenAI and Stripe open-sourced the Agentic Commerce Protocol under Apache 2.0 — meaning anyone can adopt, modify, and build on it without licensing fees or permission, a clear signal they want this to become shared infrastructure rather than a proprietary advantage. Google shipped AP2, its agent payments protocol. Microsoft launched Brand Agents. All in the space of a few weeks.

The fact that these landed together matters. Google, Microsoft, OpenAI, Shopify, Stripe, Visa, and Mastercard are direct competitors in most contexts. They don't usually converge on shared open standards at the same time. When they do, it's because they've each independently concluded that the market is coming and that owning a proprietary standard matters less than getting the infrastructure layer to work. The competition moves up the stack — to whose agent is best, whose surface converts most, whose data is cleanest. The plumbing becomes shared. That's the stage we're at.

What This Means For Your Traditional Conversion Journey

When a human shops on your site, you can see the journey — where they dropped off, what they clicked, how long they spent on your pricing page. With agent commerce, the invisibility works at two levels. If your product data is messy or your site is hard for agents to parse, you don't get recommended in the first place — the agent surfaces a competitor instead, and you never knew the comparison happened. And if an agent does attempt a checkout and hits a wall (a CAPTCHA, a broken form, an account creation gate), the purchase just fails. The shopper gets a notification. There's no abandoned cart email on your end. You have no idea the attempt was made.

Most sites weren't built for this. Think about a typical product page for a pair of trainers — you pick a size from a dropdown and a colour from a swatch. Those are variant selectors, and on many sites they're built with JavaScript that loads dynamically, which means an agent can't reliably read what options are available or confirm which one it selected. The agent thinks it picked a size 10 in black, but the site actually defaulted back to size 8 in white.

Then there's dynamic pricing — a product shows as £49.99 on the product page, but by the time the agent reaches checkout, shipping, tax, or a removed promo has shifted the total, and the agent either fails the price check set as a guiderail or completes a purchase at the wrong amount. CAPTCHAs block agents entirely, even on checkout pages where a legitimate customer has already authorised the purchase. And if your site requires account creation before payment, Google's Buy for Me won't even attempt your checkout — guest checkout is a hard requirement. Browser-based agents like Operator can fill registration forms, but they pause for human input on passwords and sensitive fields, which adds friction and defeats the point of an autonomous purchase. Either way, you're losing the speed advantage that makes agent commerce valuable. These are standard features of most e-commerce sites, and every one of them is a point where an agent fails and a sale disappears.

So What Should You Be Doing This Week?

If you're on Shopify, you're already enrolled in UCP and Copilot Checkout. In practical terms, that means Shopify has handled the technical integration for you — your product catalogue is already discoverable by AI agents across Google, Bing, Edge, and ChatGPT. Your products can appear in AI shopping conversations right now, and when an agent completes a purchase, you remain the merchant of record — you get the payment, own the order data, and keep the customer relationship. There's nothing to install or apply for (though you can opt out from your Shopify admin). The bottleneck is your product data — but probably not in the way you'd expect. If you're selling products, you obviously have SKUs, prices, and availability in your backend systems. The issue is whether that data is exposed on your pages in a format agents can actually read. When an agent compares products across retailers, it reads structured data embedded in the page markup (JSON-LD schema) — things like exact price, stock status, SKU, and product descriptions. Shopify auto-generates some of this, but it's often incomplete: SKU is treated as optional and frequently missing, availability doesn't always sync with actual stock, and descriptions can be thin or generic. Until now, getting this right has been an SEO nice-to-have — it helped with Google rich results but wasn't essential for sales. With agent commerce, it becomes the primary way your products get discovered and compared. If your schema is incomplete or out of sync with what's on the page, agents will surface the competitor whose data is cleaner.

If you're not on Shopify, the protocols mentioned above are all open standards and you don't need to adopt all of them at once. Here's a quick way to think about which ones matter to you:

  • UCP (Universal Commerce Protocol) is about product discovery and checkout through Google's ecosystem. If Google search and shopping is where your customers find you, this is the one to prioritise. It's designed to layer on top of existing platforms — WooCommerce, Magento, and custom stacks can integrate without migrating.
  • ACP (Agentic Commerce Protocol) is about checkout inside ChatGPT, powered by Stripe. If your audience uses ChatGPT for product research and buying (and an increasing amount probably do!), this is where to focus. Unlike UCP, which covers the full journey from discovery to payment, ACP focuses specifically on the checkout step — it's how a product in a ChatGPT conversation gets a "Buy" button that completes a real transaction. If you already use Stripe for payments, much of the integration work is done, since ACP is built on Stripe's infrastructure.
  • WebMCP is broader — it lets any website expose structured tools (like "add to cart" or "book appointment") directly to agents through the browser. It's relevant to any business with an interactive website, not just e-commerce. It's the newest of the three and still in early preview, but it's a W3C standard with Google and Microsoft behind it, so it's worth tracking.

None of these require ripping out your existing stack. But they do require someone to own the question.

So What Should You Do Next?

If you're a product lead or CTO reading this, the questions you're probably asking are: which of these protocols actually matters for our platform and our customers? How much engineering time does adoption take? What's the priority order? And if you're a CEO or commercial leader, the question is simpler: are we already losing sales to this, and how would we even know?

The honest answer is that most businesses don't know yet — because there's no analytics dashboard for "agent tried to buy from you and failed." The data doesn't exist in your current stack.

The most useful thing you can do right now is run an agent through your checkout flow and see where it breaks. I've been doing this for clients over the last few weeks, and the pattern is consistent: sites that look perfectly functional to humans are deeply broken for agents. The fixes aren't always complex, but you can't fix what you haven't measured.

If you want a quick starting point, there's a free readiness check at anotherflock.com — enter your domain and it'll show you which protocols you already support, where your product data has gaps, and what agents can and can't see on your site. It takes about 30 seconds and is useful as a baseline to share with your product or engineering team and start the conversation internally about what to prioritise.

McKinsey projects agentic commerce could reach $1 trillion by 2030. The businesses that start optimising for agent transactions now are going to compound that advantage as this traffic scales.

What's your experience been? Have you tried buying something through an AI agent yet? I'd love to hear how it went.

All articles