Instant Checkout is out. Merchant-controlled commerce is in. The pivot changes who owns the buyer relationship when the buyer is a model, and the payments stack still runs through the people who ship the product.

Six months ago, OpenAI was building the checkout. This week, it handed it back. Checkout.com's read on the pivot, published April 20, is the clearest account we have of what changed and why it matters for anyone building in agentic commerce. The short version is that OpenAI has dropped its closed checkout loop in favour of sending the agent back out to the merchant's own payment flow.

That is a big move. Not because OpenAI lost. Because the argument about where value lives in agentic commerce just got settled, at least for now, in the merchants' favour.

OpenAI's Instant Checkout was the first serious attempt to let an AI agent complete a purchase without leaving the chat. The model found the product, the user confirmed, and OpenAI handled the payment. Stripe was the partner. The merchant got the order. The relationship was with OpenAI.

That model has been walked back. Checkout.com's piece describes the shift as a move toward merchant-controlled checkout, where the agent surfaces the product and then hands the session back to the merchant's native payment flow to complete. The merchant keeps the transaction, the data, the customer record. OpenAI keeps discovery.

We wrote at the time about how ChatGPT's shopping discovery was squeezing merchants on checkout. The squeeze has eased. The question now is whether it eased because OpenAI could not get the economics to work, because merchants refused to integrate, or because the strategic read inside OpenAI actually changed. All three are plausible. The outcome is the same.

If you want to see the new model in production, look at restaurants.

Three restaurant brands just handed their ordering flow to ChatGPT, according to PYMNTS. The user talks to ChatGPT about what they want to eat. The agent surfaces the menu, collects the order, and then passes the session to each restaurant's own payment and fulfilment stack. No closed OpenAI checkout loop. Each restaurant owns the transaction.

That is the shape. Agent in front. Merchant in back. Processor where it has always been.

Moody's Agentic Solutions went live on AWS Marketplace this week too. Different domain, same structural choice. Moody's is not trying to be the front door or the payment layer. It is selling workflows that plug into systems the customer already runs. The agent is a capability, not a destination.

Neither of these is a bet on a single closed checkout. Both are bets on the merchant remaining the owner of the transaction.

SolvaPay, a Stockholm-based agentic payments startup, just closed €2.4 million in pre-seed funding to build payment infrastructure for agentic commerce. Read the pitch carefully. It is not another closed checkout. It is rails that sit between the agent and the merchant's existing payment stack, so the merchant does not have to rebuild anything to accept agent-originated orders.

That is the direction the money is going. Infrastructure that accepts the agent, not infrastructure that replaces the merchant. The €2.4 million is pre-seed, so the thesis will get tested, but the bet is clear enough. If OpenAI had stayed on the Instant Checkout path, SolvaPay's pitch would not hold. It does hold. OpenAI moved.

Our earlier mapping of the agentic commerce protocol space had the two directions sketched out. The closed-loop path and the merchant-native path. The merchant-native path just got the bigger validation.

Three reasons.

Merchants did the math. A closed OpenAI checkout loop means the model sits between the merchant and the customer. The merchant loses the customer record, the repurchase signal, the marketing touchpoint. Any merchant with a loyalty programme, a subscription book, or a CAC to recover declined to integrate on those terms. Without the long tail of small merchants, a closed checkout does not have enough selection to be the default.

Regulation is closer than it looks. The FCA's work on agentic payments has already started sketching what authorisation, liability, and dispute handling look like when the buyer is a model. A model-hosted checkout creates a new regulated intermediary. A merchant-hosted checkout with an agent on top keeps the regulated surface where it already is. The legal path of least resistance is the merchant-native one.

Processors did not roll over. Closed agent checkouts compete with the card networks and processors for the authorisation layer. Stripe partnered with OpenAI on Instant Checkout. Visa, Mastercard, and the acquirers are unlikely to cede the auth layer even to a partner. Once they started pushing back, the economics of a closed loop got harder.

None of these is fatal on its own. All three at once made the pivot rational.

The merchant posture here is not complicated. Three things.

Keep control of the payment flow. Whatever agentic integration you sign up for, do not let the checkout move off your stack. You lose more than the cut. You lose the customer.

Get your catalogue and your inventory into the shape an agent can actually read. That is what isitagentready.com is surfacing in developer circles right now. Most sites are not structured for agent consumption, which means the model picks someone else's product. The tooling to fix this is maturing fast.

Watch what the processors do. The merchant-native direction only holds if the acquirer and the card network support agent-originated transactions as a first-class flow. The early indications are that they will. Visa's Trusted Agent Protocol is one version of what that looks like. There will be others.

The OpenAI pivot is not the end of the argument. It is a snapshot. Three things will tell us whether merchant-native stays the default.

First, whether a second major model provider tries a closed checkout. Anthropic, Google, Meta. If one of them moves, and merchants hold the line, the answer is confirmed. If merchants flinch on the margin, the closed model comes back.

Second, whether the payment networks build agent-specific rails that make merchant-native integration frictionless. If they do, the moat around the merchant-hosted flow gets wider. If they do not, someone else will.

Third, whether the discovery layer is as valuable as it looks. OpenAI has kept discovery and given up checkout. Discovery is worth a lot, but it is a thinner business than transaction. The strategic question is whether the model can monetise discovery well enough to justify the valuation it is being paid.

Which raises the question we started with, in a different form.

If the model does not own checkout, what does it actually own in the commerce flow, and is that worth the valuation it is being paid?

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.