> ## Content Index
> Fetch the complete content index at: https://thepaymentsoperator.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# OpenAI Rebuilt Its Checkout Twice. The Liability Never Moved.
- URL: https://thepaymentsoperator.com/openai-rebuilt-its-checkout-twice-the-liability-never-moved/
- Published: 2026-11-15T23:27:47.000Z
- Updated: 2026-11-18T11:10:46.000Z
- Description: OpenAI did not retreat from agentic commerce. It changed layers — and left the liability exactly where it was.
- Author: Armando

*Fourteen months, and one sentence about dispute protections that OpenAI does not get to write.*

On 29 September 2025, OpenAI put a Buy button inside ChatGPT. Five months later it was on its way out. Fourteen months later, OpenAI shipped something different in its place.

Most of this industry has filed that as a failure story. That reading is wrong, and the way it is wrong matters. OpenAI did not retreat from agentic commerce. It changed layers — out of the checkout and into the credential. The company that spent a year trying to own the transaction now hands the transaction back to the merchant's own store and keeps the authentication instead.

For anyone who runs payments, that migration is the most instructive thing that happened in agentic commerce this year. A company with no stake in the card rails tried to build a checkout on top of them, found out what a checkout actually costs, and published the retreat in public, with dates.

And inside the retreat there is one sentence that should stop you.

Introducing ChatGPT Wallet on 13 November, OpenAI said that existing card rewards and standard bank dispute protections still apply.

Standard bank dispute protections. Written by a company that does not write dispute rules, about a transaction type those rules do not yet define, inside a product designed to eventually let an agent buy with no human in the loop at all.

That is a promise made with somebody else's balance sheet. Mine, and probably yours.

## What shipped, and what shipped instead

The record is public and it is worth laying out with dates, because the shape of the change is the argument.

**29 September 2025.** OpenAI [launches Instant Checkout](https://techcrunch.com/2025/09/29/openai-takes-on-google-amazon-with-new-agentic-shopping-system?ref=thepaymentsoperator.com) in ChatGPT. Etsy sellers are live; over a million Shopify merchants are listed as coming soon. The plumbing is the Agentic Commerce Protocol, built with Stripe and open-sourced. The merchant keeps orders, payments and fulfillment on its own systems; ChatGPT passes information between the buyer and the store. OpenAI says it will charge merchants a small fee on completed purchases, and that product results are organic and unsponsored.

Read that architecture carefully, because it is a bet with a specific shape. It bets that the chat window can be the place where a purchase is completed, and that OpenAI can take a cut of that purchase without ever touching the money. No wallet. No credential. A toll booth on somebody else's road.

**October 2025.** Walmart joins.

**4 March 2026.** At Morgan Stanley's technology conference, Walmart's executive vice-president for AI acceleration, product and design, Daniel Danker, calls Instant Checkout ["a very temporary moment in time"](https://www.retail-insight-network.com/news/openai-shifts-chatgpt-shopping-plans-to-retailer-run-apps-report/?ref=thepaymentsoperator.com) and says the Sparky experience will travel directly into ChatGPT and Gemini instead. Five months after launch, the largest retailer in the world describes the feature it just adopted in the past tense.

**Mid-March 2026.** Walmart stops using Instant Checkout. By 20 March its own assistant, Sparky, is live inside ChatGPT Plus and Gemini Advanced. A Walmart spokesperson's summary of the lesson, [as reported](https://www.thestreet.com/retail/walmart-fires-openai-in-playbook-changing-move?ref=thepaymentsoperator.com) at the time, is one sentence long: customers want consistency across every touchpoint.

OpenAI's own framing of the same month is gentler and says the same thing. Instant Checkout is moving to Apps, where purchases can happen more seamlessly. The company will focus on search and product discovery inside ChatGPT, where it has seen early adoption. Shopify confirms that discovery stays in the chat but checkout moves back to the merchant's own store, reached through an in-app browser or a separate tab. Etsy, building an app of its own, says an app gives it more control over how shopping looks and feels — and earlier access to shopper data.

**13 November 2026.** OpenAI ships [ChatGPT Wallet](https://the-decoder.com/openai-rebuilt-ai-shopping-from-scratch-with-chatgpt-wallet/?ref=thepaymentsoperator.com). United States, ChatGPT Pro, adults with an eligible US bank card. American Express, Mastercard and Visa generate the tokens. Dots is the launch partner. Stripe, Adyen and Riskified are named as providers merchants can work with. Before each purchase the user sees the items, the total and the chosen card, and approves with a face or a fingerprint. Neither the agent nor the merchant sees the real card number. OpenAI's stated reason for scrapping the first version: it wasn't flexible enough. And the roadmap, stated plainly: eventually, purchases handled entirely by the agent, within set spending limits, product categories and timeframes.

Fourteen months, and the company moved from *we will carry your checkout* to *we will carry your credential and send you home to pay*.

## The number nobody actually disclosed

If you have read anything about this episode, you have read a conversion figure. Instant Checkout converted a third as well as Walmart's own site. Three times worse. Pick your phrasing.

That number does not exist.

OpenAI never disclosed usage figures for Instant Checkout. Walmart never published a conversion rate for it either. What exists in the reporting is softer and more careful: that orders through OpenAI converted well below Walmart's typical channels, with no figure attached. The one quantity in circulation — that purchases through Sparky inside ChatGPT complete at roughly seventy per cent of the rate on Walmart.com — reaches us secondhand, and describes the replacement rather than the thing replaced. I could not stand it up at its original source, so treat it as reported, not as established.

I am spending a paragraph on this because the hole where the number should be is the most useful data point in the story.

Two parties had every commercial reason to publish a good result. A retailer with a channel that converted well would say so, loudly, at the investor conference where it instead used the phrase *a very temporary moment in time*. A platform charging a fee on each completed purchase would say so too, in the release where it instead announced a pivot to search and discovery. Both had the number. Neither showed it. In this industry, that is not silence. That is a disclosure.

And what we have in place of a figure is better evidence than a figure would be: three independent decisions, inside one quarter, by three merchants with nothing in common except the problem.

Walmart ran it for roughly six months with real traffic at a scale nobody else in the pilot could match, and chose to build its own assistant and rent out the discovery layer instead. Shopify chose to leave discovery in the chat and bring checkout home to the merchant's store. Etsy chose an app, and said out loud why: control over how shopping looks and feels, and earlier access to shopper data.

Different economics, different sizes, different risk tolerance. Same conclusion, same quarter. One merchant walking away is an anecdote. Three walking in the same direction is a finding.

## What a checkout page is actually carrying

The comfortable explanation for all this is user experience. People like pictures. Chat is bad at browsing. Shopping needs trust and comparison, and a conversation gives you neither.

That explanation is true and it is not the interesting part. The interesting part is that a checkout page is not a form for collecting a card number. It is the merchant's risk surface, and almost nothing on it is optional.

Start with the name. The checkout page is where the merchant's identity attaches to the transaction, and that identity travels to the cardholder's statement as a descriptor. When the purchase happens somewhere that is not the merchant's store, the buyer's memory of where they bought and the string that appears on their statement begin to drift apart. Every dispute team on earth knows what that drift produces, and it is not fraud. It is *I don't recognize this charge*, which costs the merchant a representment cycle to win a transaction that was never in question.

Then the terms. Returns windows, cancellation policy, shipping commitments, what counts as delivered — these are not legal decoration. They are the evidence a merchant relies on when a cardholder claims the goods never arrived or were not as described. The defensibility of that evidence depends on it having been shown at the moment of commitment. Move the moment of commitment into a chat window and you have not simplified the purchase. You have weakened the merchant's file.

Then the fraud stack. A mature merchant's decisioning is tuned on its own traffic: device fingerprints, velocity patterns, address history, basket composition, the quiet correlations that only exist in that merchant's data. Move the transaction to a surface the merchant does not instrument and those models lose the inputs they were trained on. The merchant keeps the chargeback liability and loses the signal that was holding it down. That is not a trade anyone with a functioning risk team accepts.

And then the things that actually make the sale, which vary enormously by market. Shipping choice. Loyalty. The saved card. Instalments — and if you sell anywhere in Latin America, instalments are not a feature, they are the purchase. A checkout page that cannot offer twelve payments without interest is not a checkout page here, whatever it looks like.

Seen against that list, *our customers want consistency across every touchpoint* stops reading like corporate blandness. It is the polite version of a technical objection: your window breaks my descriptor, blinds my fraud model, weakens my dispute file and drops my payment options, and I am the one who pays when any of those fail.

The first architecture asked merchants to surrender the surface and keep the liability. The remarkable thing is not that they declined. It is that anyone expected otherwise.

## The second architecture is a retreat to the rails

Now read ChatGPT Wallet the way someone who provisions tokens for a living reads it.

The networks mint the credential. American Express, Mastercard and Visa generate the token, which is to say this runs on network tokenization — the same machinery that sits under the wallets already in every pocket, and under card-on-file provisioning at every large merchant. A partner handles the United States. The merchant's own provider processes the transaction. A fraud vendor is named in the stack. OpenAI holds the user relationship, the registered card, the consent moment and the biometric.

Strip the branding off that and the shape is unmistakable. OpenAI did not invent a new category. It built a wallet.

That is a far better position than the one it abandoned, and it is a position this industry already knows how to evaluate. The question you ask a new wallet is never whether it can move a payment — of course it can, the rails do that part. The questions are narrower and harder. Who is the token requestor. What authentication is being asserted, and by whom. What liability shift does that assertion earn. And when the assertion turns out to be wrong, who absorbs the loss.

On the first three, the launch version answers well. A user who is shown the items, the total and the selected card, and who then approves with a face or a fingerprint, has performed a real authentication event on a device bound to them. That is a stronger artifact than most card-on-file transactions we authorize today, and I would rather defend it than defend a stored subscription credential. Credit where it is due: this is a serious piece of design, and it is more conservative than what OpenAI shipped fourteen months ago.

It is also, almost exactly, where Anthropic started. The commerce blueprint I wrote about here in September shipped prebuilt shopping and merchant agents and deliberately shipped no checkout protocol, no payment infrastructure and no wallet — handing the cart off and letting authorization happen on rails it chose not to touch. At the time that omission read as caution, or as a gap. Fourteen months of OpenAI's experience suggests it was a conclusion, reached earlier.

So two of the largest model companies in the world now agree, by different routes, that the model layer does not own the transaction.

Which would be a tidy place to end, except for the last line of the Wallet announcement. The roadmap is purchases handled entirely by the agent, within set spending limits, product categories and timeframes.

Read that as an operator and the biometric disappears. What remains is a stored credential, a standing mandate, and a third party initiating a transaction inside limits a cardholder agreed to at some earlier date. We have a name for that family of transaction. We have rules for it. And the rules we have were written for a merchant charging a customer it already had a relationship with — not for an agent shopping across merchants the cardholder has never dealt with, on a mandate given weeks ago.

The launch version is defensible. The roadmap is the part that lands on our desk.

## "Standard bank dispute protections still apply"

So back to the sentence.

The rewards half of it is true and uninteresting. A tokenized transaction carries the economics of the underlying account; that is plumbing, and it works.

The dispute half has three problems, and they get worse in order.

**"Standard" is carrying more weight than it can hold.** Dispute rights are not a property of a card. They are a property of a transaction — of how it was initiated, how the cardholder was authenticated, and which reason codes it is therefore eligible for. The same plastic, in the same wallet, yields materially different rights depending on how the transaction was classified on the way through. So *standard protections apply* is a conditional statement pretending to be a flat one. It is true if, and only if, an agent purchase is classified into a category that has those protections. Fourteen months into this, nobody has said which category that is.

**The promise is made by a party with no authority to make it.** OpenAI cannot grant dispute rights. It does not write network rules and it does not adjudicate claims. The networks write the rules; issuers apply them, case by case, under liability allocations neither party chose. Nobody at any model company signs off on a reason code. Which means the sentence describes an outcome that other people — us — will be required to deliver, on terms we were not consulted about, to a cardholder who read the promise and reasonably believed it.

Picture the first one of these that goes wrong. A cardholder disputes an agent purchase, the dispute is declined because the transaction was classified in a way that stripped the protection, and the cardholder goes back and rereads the launch page that told them standard bank dispute protections still apply. They will not be angry at OpenAI. They will be angry at the bank, in the branch, at the call center, on the regulator's complaint form. The reputational cost of that sentence has been booked to an entity that did not write it.

**And the roadmap removes the thing the promise rests on.** The launch version's defensibility comes almost entirely from the biometric: a person, on their device, shown a total, consenting to a specific purchase. Take that away — which is the stated plan, inside spending limits and categories and timeframes — and you are left defending a card-absent transaction where the identifiers that win disputes today point somewhere unhelpful. The device is not the cardholder's. The address the request came from belongs to infrastructure. The pattern of prior undisputed purchases from the same cardholder, which is exactly the evidence a merchant leans on in a card-absent fraud claim, now describes a server.

I made this argument in September about a different company's architecture, and I would rather not keep making it. What the industry needs is not another protocol for recording what an agent was permitted to do. We have several of those now, and they are good. What is missing is an initiation category that can carry *an agent bought this, under this mandate, with this much human involvement*, and a set of reason codes that resolve differently depending on the answer.

Until that exists, every agent dispute will be forced into a box built for something else. And the box is not a filing convenience. The box is who pays.

## From a Panamá desk

Everything above happened in the United States. Wallet launched there, for US subscribers, with US bank cards. The natural reaction from a desk in this region is that we have a year or two before it matters.

That reaction is wrong, and the reason it is wrong is structural rather than technological.

In a dollarized economy, cross-border card-absent purchasing is not an edge case our cardholders occasionally reach for. It is ordinary behavior. A cardholder here buys from a US merchant with the same indifference they buy from a store in the neighborhood, because the currency does not change and nothing in the flow tells them they crossed a border. Which means an agent does not have to launch in our market to generate agent-initiated transactions on our BINs. It only has to work for somebody holding one of our cards, and our cards work.

So the sequence runs backwards from the way it is usually described. We will receive this transaction category before we can offer it as a product. The tooling, the documentation, the vendor roadmaps and the local rule-writing all arrive later; the authorization request arrives first. That is the normal order of events in this region, and it is worth saying plainly rather than being surprised by it again: we inherit the transaction before we inherit the means to handle it.

Then the geometry of the first serious dispute. The merchant is across the border. The acquirer is across the border. The agent platform is somewhere else entirely, bound by a consent the cardholder gave it in a jurisdiction that is not ours. The issuer is here. Four parties, four legal homes, one claim — and exactly one of them has a branch the cardholder can walk into on a Tuesday morning. We are the only participant in that chain with a physical address the customer knows.

Instalments are worth a line of their own. The single feature that converts a purchase in most of this region cannot be expressed in any agent flow shipped so far. That will not reduce the volume we see; it will shape it. Expect the agent channel to select, at first, for exactly the purchases that do not need financing — small ticket, single payment, low consideration. The surprise in the numbers will be the mix, not the total.

And then the question I would actually put on a roadmap, ahead of any policy decision: can we see it at all? If an agent-initiated transaction lands on our BIN today, classified as ordinary e-commerce because no better classification exists, we will authorize it, we will not count it, and we will learn the category exists when the disputes arrive six weeks later in a queue that has no label for them. There is no policy to set until there is measurement. Visibility first, rules second.

Issuers and processors in markets our size do not write network rules. We inherit them, implement them, and absorb what they allocate to us. Which is precisely the argument for being in the conversation during the window when they are still being drafted — a window that, on the evidence of the last fourteen months, is open right now.

## My judgment

The withdrawal is not evidence that agentic commerce failed. It is evidence of where the boundary sits, and it is the first piece of real evidence anyone has produced at scale.

The boundary is this. Agents are genuinely good at discovery — at reading a catalog, narrowing a field, answering the question behind the question. That work belongs in the chat and it is going to stay there, which is exactly why Walmart was happy to put Sparky inside ChatGPT the moment it stopped letting ChatGPT run its checkout. Discovery was the part worth renting out.

Agents are bad at being the place of purchase, because the place of purchase is not a screen. It is a risk surface with a descriptor, a terms disclosure, a fraud model and a liability owner attached to it. You cannot relocate that by building a nicer interface in front of it.

The sector spent a year pricing the second as though it followed automatically from the first. It does not. They are different businesses, with different owners, and the fourteen months just spent discovering that were expensive for everybody except the people selling tokenization.

Which brings me to the uncomfortable part, aimed at my own layer.

We did well out of this, in the narrow sense that our part of the stack turned out to be load-bearing and nobody could route around it. Two of the best-funded engineering organizations in the world tried to build on top of card payments, and both ended up handing authorization, tokenization and dispute handling back to the people who already had them.

That is not a victory. It is a warning wearing a compliment. The liability stayed exactly where it was, and in fourteen months of very public experimentation, with four frameworks shipped and two architectures built and abandoned, our layer produced precisely nothing new. The same reason codes. The same initiation categories. The same unanswered question about who is the cardholder of record when no human pressed the button.

The interface got rebuilt twice this year. The liability did not move an inch, and it did not get one degree clearer. That is the scoreboard, and we are the ones holding it.

## Three questions for the people who own this

**1.** Can your authorization flow, today, tell a wallet-tokenized agent purchase apart from an ordinary stored-credential e-commerce transaction? If the answer is no, you are not waiting for this category to arrive. You are already receiving it and recording it as something else.

**2.** When the biometric comes off — purchases inside a standing limit, no human in the loop — which initiation category will you apply, and does that choice change the cardholder's dispute rights? Name the person at your institution with the authority to decide it, and tell me whether anyone has actually asked them.

**3.** A cardholder arrives with a declined agent dispute in one hand and, in the other, a screenshot of a launch page promising that standard bank dispute protections still apply. What does the person behind the counter say to them?

If you have an answer to the third one, I would like to hear it, because I do not have a good one yet. Reply to this email with a line or two. The sharpest answers will shape the next issue.

## Sources

**TechCrunch, 29 September 2025** — OpenAI launches Instant Checkout with Etsy and Shopify, built on the Agentic Commerce Protocol with Stripe. [Link](https://techcrunch.com/2025/09/29/openai-takes-on-google-amazon-with-new-agentic-shopping-system?ref=thepaymentsoperator.com)

**Retail Insight Network, March 2026** — OpenAI shifts ChatGPT shopping to retailer-run apps; Daniel Danker's remarks at Morgan Stanley's technology conference on 4 March; Etsy and Shopify responses. Reports the original CNBC account of 20 March 2026, which I was unable to open directly. [Link](https://www.retail-insight-network.com/news/openai-shifts-chatgpt-shopping-plans-to-retailer-run-apps-report/?ref=thepaymentsoperator.com)

**TheStreet, 22 March 2026** — Walmart ends the Instant Checkout integration and takes Sparky into ChatGPT and Gemini; the spokesperson's statement on consistency across touchpoints. [Link](https://www.thestreet.com/retail/walmart-fires-openai-in-playbook-changing-move?ref=thepaymentsoperator.com)

**The Decoder, 13 November 2026** — ChatGPT Wallet: tokenization by American Express, Mastercard and Visa, biometric per-purchase approval, named payment providers, and the roadmap toward agent-handled purchases within set limits. [Link](https://the-decoder.com/openai-rebuilt-ai-shopping-from-scratch-with-chatgpt-wallet/?ref=thepaymentsoperator.com)

On the figure I declined to use: the comparison circulating as *Instant Checkout converted three times worse* does not appear in any primary disclosure. OpenAI has not released usage figures for Instant Checkout, and Walmart has not published a conversion rate for it. The related number — Sparky purchases inside ChatGPT completing at roughly seventy per cent of the rate on Walmart.com — reaches me through secondary reporting I could not confirm at its origin, and in any case measures the successor rather than the feature that was withdrawn. It is reported here as reported, and nothing in the argument above depends on it.