Agentic Commerce, One Year Later: Merchants Push Back, Businesses Get a New Customer, and the Infrastructure Race Begins

A follow-up to Intelligent Commerce: the Next Financial Revolution, June 2025

When I wrote about agentic commerce and programmable money last year, the conversation was mostly theoretical. Fifteen months later, the details are coming into focus, and so are the strategic questions they raise for merchants, consumers, and the future of commerce. We discussed this live on Friday with a curated set of builders, operators, investors, and advisors in payments, AI, and commerce. What's clear is that this is a space to watch and build in more than ever. What's not clear is the timeline to mass consumer adoption, how merchants retain loyalty, and who wins.

OpenAI and Anthropic Are Not Building the Same Thing

OpenAI's Instant Checkout, built on the Agentic Commerce Protocol it co-developed with Stripe, lets a shopper complete a purchase inside ChatGPT itself, already live for over a million Shopify merchants alongside Etsy sellers. The transaction happens on OpenAI's surface; the merchant fulfills, but the moment of purchase now belongs to the chat window, not the merchant's own site.

Anthropic took the opposite bet. Claude's shopper and merchant agents run inside the retailer's own app or website, helping with product questions, cart-building, and checkout, but stop short of completing the purchase. The customer finishes through the retailer's existing checkout, and the merchant keeps control of payment and the shopping experience.

This is not a small distinction. It's the whole ballgame. One model treats the merchant as a fulfillment backend to an AI-native front door. The other treats the AI as a layer that makes the merchant's own front door smarter, without taking the relationship away. OpenAI's model reportedly takes a cut of completed sales; Anthropic monetizes through token and compute usage rather than a transaction fee.

I'm biased toward Anthropic's approach: it bolsters merchant success inside the merchant's own ecosystem rather than routing the relationship, the data, and the loyalty through a third party's surface. Anthropic is citing partner-reported results of up to 35% larger carts and a 60% higher likelihood of completing a purchase, though those figures come from a single partner and haven't been independently verified. The counterpoint that emerged at breakfast: what if consumers simply prefer to skip the merchant's site altogether? Fair question, but it feels far off even generously squinting. As one consultant put it, most large enterprise merchants they advise see so little consumer agent usage for shopping that they aren't taking any action at all. That's how early this still is.

Founders should watch which model wins enterprise trust over the next two quarters; it tells you whether merchants ultimately prize reach or control more. I'm rooting for small business, diverse user interfaces, and design excellence over all of us haggling with ChatGPT to do a better job fetching what we're trying to buy. There's zero delight in that!

Businesses Will Interact With Agents Whether They're Ready or Not

This isn't only a consumer shopping story. It's a fundamental shift in what "a customer" even is, and it's already breaking things. Take Resy: bot and broker accounts show no-show rates roughly four times higher than those of the general population, and Resy's understandable initial response was to detect and deactivate them. That was the right call as long as "bot" almost always meant a bad actor. It doesn't anymore. When a legitimate personal assistant is booking a table using the same automated patterns a scalper bot uses, blunt account deactivation becomes a customer experience problem, not just a fraud control. I know American Express well; they accept nothing less than a first-class user experience, and legitimate account deactivation is far from delightful.

This is exactly the frontier Lorikeet is building on. Lorikeet built B2A specifically for an AI assistant to query on a user's behalf, with a usage policy that requires user consent before an agent can act, a governed lane distinct from the fraud lane. Every consumer-facing business will need a version of that distinction soon: how do you let a customer's agent in and delight them, while keeping scalper bots out?

Beyond the fraud risk, there's the customer experience itself. As one person said Friday, we have a mismatch: merchant sites were built for consumer discovery, upsell, checkout, and loyalty. Merchant sites will need to evolve now that some share of their traffic is AI assistants shopping on a consumer's behalf.

Stripe Is Quietly Building the Rails, and That Should Worry Every Commerce Platform

Stripe's Agentic Commerce Suite lets businesses publish their product catalog once and sell through any agent that plugs into it, while remaining the merchant of record. It layers in Shared Payment Tokens to distinguish legitimate agent transactions from fraud, and it's building toward open protocols like the Machine Payments Protocol and the Universal Commerce Protocol. 78% of companies on the Forbes AI 50 list build on its rails. Stripe isn't picking a side between OpenAI's in-chat checkout and Anthropic's merchant-controlled model. It's building the plumbing underneath both. As usual, Stripe is positioning itself as the layer every commerce platform will have to interoperate with, regardless of which agent architecture wins.

That's what makes me wonder about Shopify's long-term position. If Stripe becomes the default settlement and discovery layer for agentic transactions, the durable value in "a Shopify" starts to erode.

Here's what I think is actually happening at Shopify: they're not betting on Anthropic's architecture. They're wired into OpenAI's ACP, Google's UCP, and Stripe's protocol simultaneously. Shopify is making itself protocol-agnostic so it's in every agent, regardless of who wins.

What Happened to "The Customer is Always Right"?

We're at an interesting inflection point for builders and marketers alike. Building with the customer at the center was the winning strategy for a long time, but building for the individual may become this generation's "faster horse" versus car.

We're clearly in the early innings of this transformation. While the AI assistant war feels like it's in full force, Muse from Meta isn't even a month old, Instinct is two months old, and Amazon is already blocking Muse while likely planning to take on the rest soon enough. It's still early, not only because of the novelty of it all, but also because the consumer population testing these tools remains only a tiny fraction of the overall shopping population. As someone said at our discussion Friday, shopping online is already quite easy for many things, so for most consumers, whether AI assistants solve a real problem is still an open question. In this way, many merchants are listening to the customer… which means not changing a thing… yet.

What's Still Left to Build

None of this is close to finished. A few gaps stand out as the real white space right now:

• Fraud tooling purpose-built for agent traffic. Traditional bot-detection heuristics weren't built to distinguish a legitimate personal agent from a scalper. As someone said Friday, KYA (know your agent) needs new parameters, and the sheer speed of agent transactions demands new technology.
• Orchestration across a fragmenting protocol landscape. Between ACP, MPP, x402, and the Universal Commerce Protocol, merchants need someone to sit above the standards war, or at minimum orchestrate across them.
• Accessibility for novice AI users. Most of what's shipped assumes a sophisticated user directing a capable agent. The mainstream customer needs a gentler on-ramp and demands a "killer" experience with ready-made distribution.
• Trust and observability layers. Merchants and consumers need to understand why an agent made a decision and have tooling to optimize in real time.
• Financial services embedded in the agentic flow. Some have called for embedded lending layers; others, embedded banking. Cheers to the founder who can deliver this in a delightful business-to-agent way!

The infrastructure race of the last fifteen months has been about proving agentic commerce works at all. The next fifteen will be about who builds the plumbing that makes it safe, trustworthy, and the default. Then I expect a new player to come in and win the hearts, minds, and trust of mass consumers (and I don't believe that will be Meta).

If you are building in these spaces, I hope to hear from you; the next step is to have you first interact with my agent, of course.