Agentic Commerce is Dead. Long Live Agentic Commerce.

OpenAI killed Instant Checkout this week. LinkedIn punditry is simultaneously mourning, “told you so-ing,” and prognosticating how this re-entrenches the status quo. I’ll confess I’m giddy.

This is good news. The best kind: a wrong idea failing fast, clearing space for the right one to develop properly.

Let me explain why, and why the people calling agentic commerce a mirage are simultaneously correct and completely wrong.

What Actually Died

First, let’s try to stick to the facts of the OpenAI news. OpenAI launched Instant Checkout in September 2025. The idea was that users could discover and purchase products without leaving ChatGPT. OpenAI would sit between merchant and customer, process the transaction, and charge merchants 4% for the privilege, on top of Stripe’s standard 2.9% (better when you’re a big merchant obviously).

Six months later, roughly 12 of Shopify’s millions of merchants had integrated. Users browsed through ChatGPT and then left to complete the purchase elsewhere. Insert Nelson Muntz here. OpenAI is removing Instant Checkout.

What died is a specific business model: a conversational AI platform inserting itself as a new checkout layer and taxing every transaction. OpenAI tried to be the pipe and the toll booth simultaneously. Users rejected the toll booth…and the market was right.

Eric Seufert at Mobile Dev Memo called agentic commerce “a mirage” last September, and his analysis of this specific model is correct. Amazon blocked external agents to protect $68bn in advertising revenue. Shopify opened discovery but locked checkout behind Shop Pay. The platforms with the most to lose moved to defend themselves. That’s rational. That’s what platforms do.

But Seufert makes a category error that most of the current commentary repeats. His entire analysis assumes agentic commerce means agents accessing existing platform inventory, competing with Amazon on Amazon’s terms, for Amazon’s customers. His conclusion follows from that assumption: “you’d need to build a better Amazon.”

I think the question is whether we need Amazon at all.

(Side note: If you’ve ever met me, you know I hate Amazon search results exactly because of the way they protect advertising and search rankings. They give you horrible search results…arrrgh!!!)

The Alibaba Elephant

Before dismissing everything, it’s worth acknowledging what Alibaba proved.

Alibaba’s Qwen App completes food orders, travel bookings, and purchases inside a single conversational interface. It works. It works because Alibaba owns the AI model, the marketplace, the payment rail, and the logistics. All of it. Qwen is an agent for Alibaba. Users get efficiency. Alibaba gets the behavioural data, the transaction relationship, the payment capture, and the loyalty lock-in.

This is platform-centric agentic commerce executed competently. OpenAI tried to replicate it without owning the stack.

The Western internet has no Alibaba. Amazon comes closest and Amazon is building walls. Google, Shopify, Stripe, and the AI platforms are all competing rather than consolidating. Does anyone thing that’s going to change? No. The fragmentation that makes a centralised model impossible in Western markets is the same fragmentation that creates the need for neutral infrastructure to grease the wheels of commerce here.

Why Aggregation Existed, and Why That’s Changing

Marketplaces and malls existed because connecting buyers and sellers used to be expensive. Expensive for both parties.

The cost wasn’t just technology. It was discovery. Lost opportunity cost. A buyer in any given city couldn’t know about every merchant with relevant inventory. A merchant couldn’t reach every buyer with relevant intent. One solution is the mall. Marketplaces. Malls solved that problem by putting everyone in the same physical space. Amazon solved it by building a database large enough that if you searched for almost anything, you found it.

Both solutions require an intermediary because the alternative, buyers and sellers finding each other independently, cost more than the toll.

Here’s what changed. Connecting millions of buyers to millions of sellers is now a mathematical and database problem, and we’ve solved both. The compute is cheap. The infrastructure is scalable. The part that was hard, finding the right match across a large population, is exactly what the current generation of AI does well.

There’s a further reduction that most people miss. Agents don’t arrive with browsing behaviour. They arrive with structured intent. A human searching Amazon has latent, ambiguous intent isn’t expressed and that the platform can’t infer. A buying agent arrives with a declared constraint set: product requirements, budget, delivery window, payment method, supplier certifications. The matching problem collapses from “find relevant options among millions” to “filter against declared constraints.” You’re matching one structured query to a filtered subset. The computational lift drops by orders of magnitude before a single database query runs.

The marketplace captured value because discovery was expensive. When discovery becomes cheap and structured, the marketplace loses its justification. What replaces it is a directory.

The Yellow Pages Moment

This isn’t a new idea. It’s a very old one, applied at a scale that wasn’t previously possible. I’ve talked about the idea of agents using dynamic directories before and we have historic analogues.

The Yellow Pages wasn’t a marketplace. It didn’t intermediate transactions or extract a percentage. It was a structured index of capability declarations. You looked up plumbers in your area, got a list, called one. The directory got paid for listing, not for the plumbing job.

That model was displaced by search because search handled unstructured queries better than a categorised directory. Unstructured human queries needed the inferential power of a search engine to be useful.

Agents don’t submit unstructured queries. They submit structured intent. The directory model is viable again, at a scale the Yellow Pages could never have reached, with matching quality the Yellow Pages could have never approached. Transaction has been, and always will be, the easy part of a commercial interaction. Finding the right partner in that directory becomes more important than the checkout process. That’s the directory…where buyers present structured intent and sellers present matching propositions.

The Missing Half

Buying agents exist. Selling agents don’t…yet. That’s what’s missing in this directory example.

Every conversation about agentic commerce focuses on the buyer side. Agents that shop on your behalf, find the best price, execute the transaction. The seller sits passively waiting to be discovered, feeding product data into whatever protocol the platform demands, paying to be visible. And hoping that paying for visibility actually results in a chatbot recommending their product.

I think selling agents are inevitable, and I think they’ll matter more than buying agents. At least in consumer markets.

People (not all, but many) like to shop. There’s a chemical component to it. The anticipation, the discovery, the decision. Consumer buying agents will exist and will handle routine replenishment, but discretionary spending involves a cognitive and emotional process that many people won’t fully delegate. The dopamine loop of finding something you love is something consumers will preserve.

Selling is different. Nobody has an emotional attachment to the process of generating a proposal, qualifying a prospect, or responding to an RFQ. These are cognitive tasks. Cognitively demanding, time-consuming, and exactly the kind of thing that automation has always targeted.

Look at the pattern across economic history. We automate manual tasks first. Physical labour, then repetitive cognitive work, then complex cognition. We now have the tools to automate complex cognition. Deciding which cognitive tasks get automated first is about understanding where the cost of human labour is highest relative to the value it produces.

In B2B commerce, that cost is enormous. Supplier qualification, third-party risk management, contract negotiation, compliance verification, procurement workflows, dynamic pricing across thousands of SKUs and dozens of customer segments. These are all cognitive tasks performed by humans today or non-cognitive tools following rules. Selling agents don’t require AGI; they require the same language and protocols as buying agents, just pointed the other direction.

Human brokers exist. Human sales agents exist. Human procurement officers exist. The automation of brokering and agency is not a new concept. It follows exactly the same logic as every labour-saving advance for three (probably more) centuries. The people who say selling agents won’t exist are making the same argument as the people who said cars were impractical and horses were more reliable.

The easily automated tasks will be automated first. Then the next layer of complexity. Then the next.

What This Week’s News Actually Means

OpenAI didn’t retreat from agentic commerce. They retreated from a model that required them to own the checkout layer and tax every transaction.

ACP survives as a protocol. The scope is narrower: routing through apps rather than in-chat checkout. That’s the correct outcome. Protocols should be neutral. Advertising and paid-discovery platforms should compete on the surface, not own the plumbing.

The pundits writing obituaries are looking at one failed implementation of one version of one model and concluding the category is dead. That’s like watching a mall close in 2005 and concluding that retail commerce is over.

What died is platform-centric robotic (not agentic) commerce. The version where a single company positions itself as the new intermediary, extracts rent from both sides and calls it innovation.

What’s coming is something older and simpler. Buyers with agents that represent their actual interests. Sellers with agents that represent theirs. Infrastructure that connects them without owning the outcome.

We’ve built this before. We just had to wait for the technology to catch up with the idea at scale.

---

Marc Massar is the founder of AURA Labs, building protocol-level infrastructure for agentic commerce.