Agentic Commerce Isn't Dead. OpenAI Just Tried to Build It in the Wrong Order

With the shift in OpenAI's Instant Checkout strategy, if you've seen some of the headlines, Agentic commerce is dead, apparently.

Someone should tell the 200 million orders Alibaba's Qwen processed over Chinese New Year: milk tea, flights, movie tickets, gifts, apparel, furniture... All completed inside a chat interface. Over 4 million users aged 60 and above made their first-ever AI-assisted purchase, most in two or three messages. The naysayers declaring victory this week got the short-term call right. They're wrong about what it means.

OpenAI did kill (or pivoted, whichever framing you prefer) Instant Checkout. Purchases are now routed back through retailer apps or the retailer's own site. The commentary followed predictably: "Agentic commerce is a mirage". The LinkedIn victory laps started before the news had settled.

I believe that diagnosis is wrong.

What actually died

Let's recap the timeline of events:

Words like "learning with us", the progress made in, six months, and only twelve merchants? That's not a product or a misread on consumer behaviour. It's a pilot that revealed the infrastructure wasn't ready.

What I believe actually happened is the category error being made everywhere last week is treating the failure of a single checkout feature as proof that the entire behavioural shift of agentic commerce is a failure.

The wrong question: was it consumer behaviour?

The instinct is to reach for the demand-side explanation. Consumers weren't ready, trust wasn't there, and people just want to browse.

Now, some of the data reported was that consumer uptake of purchasing in ChatGPT was low. No doubt there is some truth to that: using the platform for discovery does not instantly translate to "I'm willing to purchase in your platform" also. There are many reasons this might not have been an effective path early: product availability, poor UX, limited retailer options, etc. That theoretical list runs for days. No doubt hesitancy to trust agents with credit cards would've been right near the top of reasons, and that's a very valid fear in a space that requires a new behaviour.

My view is this argument collapses under scrutiny. Capgemini research found 53% of US consumers have already made a purchase based on AI recommendations. The first 80% of the buying journey (research, comparison, reviews synthesis, narrowing consideration sets) has already been rewired. Consumers didn't reject agentic checkout, the question is whether the consumer ever got a genuine working checkout experience?

Shopify's own president said the problem was on the supply side. Maintaining real-time stock and dynamic pricing for millions of SKUs inside an LLM is a resource-draining technical challenge that the industry hadn't solved. Safeguards against fraudulent or erroneous purchases become more complicated inside an AI framework. These are infrastructure problems, not demand problems.

Sure, consumer behaviour may not quite be ready. But as we saw in China (naturally, who are always years ahead), that isn't the blocker when the path-to-purchase is seamless for all involved.

The right question: why couldn't OpenAI get merchants live?

OpenAI needed cooperation from parties with rational reasons to refuse.

Amazon blocked external agents to protect $68 billion in advertising revenue. Every platform with something to lose moved to defend it, which is very predictable, rational, expected. The structural preconditions for Instant Checkout to work were:

None of those existed. OpenAI charged 4% on infrastructure they didn't own, couldn't control, and couldn't compel anyone to open up.

The ACP survives in narrower form, which is the correct outcome. Protocols should also be neutral, as a protocol standardises the communication, not the incentives. Amazon still has rational reasons to block external agents. Shopify still protects its payment layer. A common language between parties who don't want to cooperate remains useless. This was a structural preconditions failure, not a consumer readiness failure or a pricing failure.

The referee problem

Marc Massarmakes this point (and others that I have pointed out), brilliantly in hisblog.

Taking the protocol point a step further, ChatGPT solves the matching problem brilliantly: connect buyer intent to relevant products at scale. This a high value problem which has always delivered significant value to the match-maker (examples below). But, the moment it becomes the recommender, it becomes a marketplace participant with a stake in the outcome. Those two roles are incompatible.

Markets work when the referee doesn't have a bet on the outcome. Stock exchanges, commodity markets, real estate listing services all function on this principle. The operator creates value by facilitating efficient matching. When the operator starts favouring one side or extracting rent from position in the market, efficiency degrades and participants lose trust.

OpenAI wasn't just building a checkout feature. It was trying to be the market operator and extract rent from the market simultaneously. A referee who bets on the outcome isn't a referee.

Now, with ads confirmed for ChatGPT in 2026, the structural contradiction is visible. OpenAI couldn't own the transaction layer, so it's retreating to the recommendation layer and monetising through placement. When a user asks for the best hiking boots and the first result is a sponsored placement, the core value proposition (objective reasoning from a neutral expert) is compromised. Call it inference bias: when an AI weights a sponsored product higher in its reasoning, it doesn't feel like a banner ad. It feels like betrayal. Google has spent 20 years managing this tension. OpenAI is about to learn the same lesson, with users who came to ChatGPT specifically because it didn't feel like an ad platform.

The $15 billion Amazon investment in OpenAI announced recently is worth considering also. Hard not to read Instant Checkout's death as OpenAI conceding the transaction layer to Amazon while retaining the discovery layer: a commercial arrangement of sorts where OpenAI provides the brain and stays away from Jassy's cash register.

Proof the concept works

During Chinese New Year, Qwen processed nearly 200 million orders through a conversational interface. Not searches. Orders. All inside a chat interface, many completed by people purchasing through AI for the first time.

The difference isn't just scale. It's purely structural in this instance.

Alibaba owns the AI model, the e-commerce marketplace (Taobao), the payment rails (Alipay), the maps, the travel platform, and the ticketing system. When a user asks Qwen to book a movie, it doesn't redirect them anywhere, it books the movie. Friction between "I want this" and "it's done" approaches zero when one company controls every layer. OpenAI owns none of those layers, which is why it needed cooperation from parties who had every reason to refuse.

The Western internet has no Alibaba. Amazon comes closest and is building aggressively (Rufus is getting more capable by the quarter) but hasn't gone all-in on the integrated stack. Until a Western player assembles comparable integration, AI commerce in the West remains a better search engine: impressive, but fundamentally incremental.

There's a structural shift underneath this that the naysayers are missing. A human searching Amazon arrives with vague intent the platform tries to infer. A buying agent arrives knowing exactly what it wants: product category, budget ceiling, delivery window, payment method, sustainability requirements. The matching problem shrinks dramatically. Marketplaces captured value because connecting buyers to the right sellers was expensive and hard. When a buying agent does that work in milliseconds with declared constraints, the case for a high-margin intermediary gets thinner. Marc Massar at Aura Labs put it plainly: when discovery becomes cheap and structured, what replaces the marketplace is a directory.

Where agentic commerce actually lands first

The B2C story is complicated by something real: the dopamine loop of discretionary shopping. Many consumers won't fully delegate discovery for things they enjoy buying. Routine replenishment is the near-term B2C win: toothpaste, coffee pods, printer ink. The things people already buy on autopilot, where price and availability matter more than the joy of finding something new. That's the first wave.

Discretionary spending is harder. The act of finding something you love (the scroll, the comparison, the moment of deciding) has genuine value to people. Agents will assist that process long before they replace it.

B2B tells a different story entirely. Nobody has an emotional attachment to generating proposals, qualifying prospects, responding to RFQs, or managing supplier qualification workflows. These are cognitively demanding tasks with enormous labour costs: exactly where automation has always arrived first. A procurement officer who spends 40% of their time on compliance verification and supplier onboarding isn't doing that for the love of it. The selling side of commerce is where the clearest economic case for agents lives, and it's the side nobody is talking about.

Buying agents get all the attention. The vision is clear: an agent that shops on your behalf, finds the best price, executes the transaction. But the seller sits passively in this picture: feeding product data into whatever platform demands it, paying for visibility, hoping the algorithm recommends them.

Selling agents flip this. Not an agent that shops for you, but an agent that represents a merchant's interests in the market: evaluating incoming buyer requests against real inventory and margin thresholds, crafting responses tailored to each specific query, declining to participate in markets where the economics don't work. A merchant moves from being a row in someone else's database to being a participant in a negotiation.

The infrastructure for this doesn't exist yet at scale. What does exist, acquirers like Stripe and Adyen who hold transaction performance data, platforms like Shopify who know the product catalogue, ERPs that know fulfilment. Someone needs to point it in the right direction. Right now, everyone is building buyer-facing features. The seller-side opportunity is wide open.

What retailers should actually do

Instant Checkout disappearing from ChatGPT doesn't make the underlying preparation irrelevant. AI agents are already surveying the web. They're reading product description pages, extracting pricing, comparing specifications, synthesising reviews. The discovery layer is live whether or not a checkout button follows. Merchants who haven't structured their product data for machine readability are already less legible to the agents influencing purchase decisions, and that gap widens every quarter.

The practical steps still stand:

Getting this right matters regardless of which platform or protocol becomes the standard. The infrastructure underneath commerce is shifting; the preparation is the same either way.

The question to add to that checklist: readable to whom, and on whose terms?

If the answer is "to the platforms, on their terms," merchants are preparing to supply the next intermediary. Google Shopping charged for placement. Amazon charged for visibility inside search. Social platforms charged for reach. The agentic version of this pattern is already forming: platforms building feed ingestion, ranking algorithms for agent-mediated results, and paid placement inside AI responses.

Agency requires loyalty to a principal. If the agent's principal is the platform, it's the platform's agent, not the buyer's, not the seller's. The label on the tin doesn't change the economics inside it. Amazon's Rufus recommends what's best for Amazon's margins. Every "agentic commerce" product being announced right now is a platform agent in costume, and merchants structuring their data solely for these platforms are building the pipes that will extract value from them.

The deeper concern isn't even visibility. Merchants who've spent years on Amazon understand invisibility, that's fixable with ad spend. What's harder to fix is the loss of the conversation itself.

When a buying agent queries a platform and gets back a ranked list, the merchant never had a chance to present context. No ability to explain why their product fits this specific buyer's constraints. No ability to surface what makes them different. No ability to know why they appeared third, or didn't appear at all. The platform absorbed the buyer's intent, made a judgement call, and handed back a result. The merchant is downstream of a decision they had no part in.

That's not a marketplace. A marketplace connects two parties and lets them negotiate. This is a platform appointing itself the arbiter of what buyers should want, and charging sellers for the privilege of being considered.

The merchants who come out ahead will be the ones who do both: prepare their data for AI readability, and start thinking about what it means to have an agent that represents their interests rather than a platform's. That category barely exists today. It won't be empty for long.

The 90% is already gone

The skeptics got the short-term call right. Instant Checkout was oversold and underbuilt. OpenAI tried to replicate Alibaba's integrated stack without owning any of the layers that make it function.

But the 90% of the buying journey that precedes the transaction (research, comparison, synthesis, consideration-set narrowing) has already moved. Not eventually. Now. And this matters more than the checkout debate suggests.

The consideration set is being formed upstream now. Which brands get evaluated, which products make the shortlist, which suppliers even enter the conversation, those decisions are increasingly being made inside AI interfaces, before a consumer reaches a product page, before an on-site search, before any of the ad surfaces that retail media has been built on. By the time a buyer arrives at checkout, the real decision has often already been made. The 10% that's still unsolved is the transaction. The 90% that's already shifted is where the commercial value actually lived. The parallel to 1994 is instructive: the browser didn't transform commerce by itself. It revealed that everything behind commerce was inadequate for what was coming, and then two decades of infrastructure investment did the actual work. ChatGPT's Instant Checkout played the same role. It didn't kill agentic commerce. It showed exactly how much still needs to be built.

By the time someone cracks the transaction layer, and they will, most likely Amazon, most likely from the inside out, the discovery shift will be so complete that checkout inside a chat interface will feel like the obvious missing piece rather than a leap.

What's the question nobody in this conversation is asking?

Sources: Capgemini consumer AI research 2025; Shopify investor conference Q1 2026; OpenAI spokesperson statement via The Information; Marc Massar, Aura Labs Substack (Feb 2026); Alibaba/Qwen Chinese New Year data; Kiri Masters, The Drum (March 2026)