The Labs Are Building the Front Door, Not the Store

In March, OpenAI killed Instant Checkout, its button for buying from merchants inside ChatGPT, after the merchant fees made it unsellable. Three months later the feature came back in a different shape. In June OpenAI signed a deal putting payments inside ChatGPT and OpenAI's Atlas browser, but with a telling division of labor: OpenAI supplies the deciding — the part where the software figures out what you want. Authorization, fraud monitoring, disputes, and the tokenized card credentials — digital stand-ins for the card numbers — all stayed on Visa's rails, the pipes money actually moves through ^[1]. Read one way, that is a retreat. OpenAI tried to own the transaction, couldn't make the economics work, and ceded the rail to the card network. Which leaves a question hanging over the whole industry's big pivot. If execution — actually doing the work, actually completing the sale — was supposed to be the prize, why hand the transaction away? The reason everyone rushed toward execution in the first place is that the other thing these companies sell stopped being scarce. In August OpenAI cut fees on its Luna model by 80 percent, Anthropic answered with a half-cost model, and xAI and Meta followed, all of it to hold share against self-hostable Chinese open-source models like Alibaba's Qwen that businesses can run themselves ^[2]. The pivot is also running against the labs' own deadline: OpenAI has ruled out an IPO this year, Anthropic has pushed its own from October to November, and analysts describe a circular funding cycle — Nvidia and SoftBank invest in OpenAI, which sustains Oracle's debt-fueled data-center spending — that keeps turning only if the labs can show a durable revenue story ^[3]. Anthropic's newest release makes the commodity status explicit: Claude Sonnet 5.5, out today, is advertised as doing the same work more cheaply rather than advancing the frontier, and enterprise customers are already 80 percent of the business ^[4]. Some coverage adds a second push: the U.S. government blocked Anthropic's Fable 5 frontier model over bioweapons-capability concerns, and the Pentagon terminated Anthropic's contracts ^[4] — whether that pushed the labs toward safe, sellable models like Sonnet 5.5 is the framing of the reporting, not an established mechanism. And the buyers treat the models as interchangeable anyway: Scale AI's benchmark ranks the leading systems by which profession and region each one suits, the way you'd comparison-shop tool brands ^[5]. So the question from the failed checkout sharpens. If the intelligence is a commodity and the frontier is blocked, the value has to live somewhere else: in doing the work. Follow a single purchase and ask, at each step, whose system holds what. An agent, here, is just a program that acts on a person's behalf: booking, buying, filing. Start with the rails. Mastercard, launching autonomous shopping this month, issues a fresh virtual card number for every purchase an agent makes, so a program that buys things can't form a spending habit and no agent can spend without the network clearing it ^[6]. Visa's Intelligent Commerce Connect is deliberately built to work with non-Visa cards too, engineering itself into every agent-payment stack so the authorization runs on Visa's rails no matter which bot does the picking ^[7]. The networks aren't being displaced. They're issuing the credentials and keeping the authorization. Then the data. Snowflake paid Anthropic $200 million and embedded Claude inside its own Cortex environment so finance and security workflows run without the data leaving Snowflake's walls. The intelligence travels to the data, not the reverse ^[8]. Claude's desktop app ships through Amazon Web Services, Google Cloud, and Microsoft Foundry, inside the IT controls of the companies that already hold the data ^[9]. Bank of America spends $13.5 billion a year on technology and has reserved AI for administrative efficiency rather than direct investment advice, keeping the client relationship human and building with Salesforce and Zoom instead of the frontier labs ^[10]. A bank spending billions to keep the AI out of the advice. Then the workflow. Salesforce has gone agent-first ^[11], and its Agentforce revenue is up more than 240 percent because it sells agents from inside the software companies already run ^[12]. Workday is pitching an agent System of Record, the registry that would keep track of every agent a company runs ^[13], and Oracle is shipping banking agents for loan origination and credit decisions ^[14]. Every incumbent that owns the workflow is folding the agent into its own product, so the labs' models run somewhere in the middle of someone else's software. Even Anthropic's own enterprise product head draws the line in the same place.

We’re providing infrastructure and intelligence so our partners or our customers can bring their business knowledge, their expertise, their trusted relationships and their customers to the equation. — Scott White

And in April, the disarray. Anthropic blocked third-party harnesses — outside tools that run Claude for you — like OpenClaw from flat-rate Claude subscriptions ^[15], forcing their builders onto metered pay-as-you-go keys because autonomous agents were burning more computation than a flat fee could justify. OpenAI answered by hiring OpenClaw's creator and pledging to support the OpenClaw foundation.

The future is going to be extremely multi-agent and it’s important to us to support open source as part of that. — Sam Altman

On the question of who gets to build agents on top of their models, the two labs came down on opposite sides. The money with the clearest view of the economics is not convinced: Thoma Bravo declined both labs' joint ventures outright, questioning whether enterprise AI deployment has a durable profit profile ^[16]. So what are the labs actually building? Not the back office. The front door. Shopify moved its storefront inside ChatGPT last October, months before the checkout died ^[17]. Anthropic sells private plugin marketplaces so companies build and distribute their own agents, reaching into the spreadsheets and chat channels where transactions begin ^[18]. Visa will charge merchants for making their catalogs discoverable inside AI environments ^[7]. And a measurement industry is sprouting around whether your brand shows up when a chatbot is asked where to buy ^[19].

AI search is the new front page of the internet, and most brands don't even know if they're showing up. — Tifenn Dano Kwan

That is routing work, metered by the token, the unit of computation a model sells, and it is exactly what the harness lockout was protecting. The transaction stayed on the card networks' rails; the data stayed inside the incumbents' walls. What the labs are building is the narrowest slice left, and possibly the most valuable one: deciding what a customer sees before the transaction, and charging for the routing. But the doorway is built ahead of the demand, and ahead of the ability. Seven percent of shoppers are willing to let an agent buy for them — and only under predefined conditions ^[6]. As of mid-September, the general chatbots still can't reliably do live price shopping at all; the actual execution is handled by specialized fee-charging tools and old aggregators like Google Flights ^[20]. The labs have built the doorway, and nobody has yet agreed to walk through it.