The agent can choose the merchant while the card network still gets paid. PayPal's branded checkout is exposed, but Braintree gives shareholders a second business to price.
The checkout screen lists a merchant, items, a total and a card. A fingerprint approves the purchase. Then the merchant decides whether an order actually exists. There is something peculiar about a checkout that begins with an agent and ends with a human signing off on its choices. You still hold the card. Someone else is getting closer to deciding where it gets used. For an incumbent payment button, that is an awkward place to be.
Agents threaten PayPal at the instant of choice, even when Visa or Mastercard carries the payment. An agent selecting a merchant and initiating a purchase can bypass branded PayPal checkout while preserving the networks underneath. That threatens transaction revenue from PayPal products and services, the company's closest disclosed proxy for branded checkout economics, while Braintree gives it a processing hedge. If you own PayPal, those are separate bets: one on people choosing the brand, the other on merchants still needing its plumbing.
Choice Layer Capture is what happens when software takes control of purchase initiation without taking over the financial rails. In OpenAI's 13 November 2026 launch, OpenAI's wallet starts with eligible US Pro users inside dots, its agent surface. Users review items, price and card before biometric approval. OpenAI expects fuller delegation under spending limits over time. The agent leads the shopper to payment before PayPal even gets a chance to propose its button.
The wallet terms make the separation explicit. Merchants remain sellers of record. The credential is tokenized instead of the card number. OpenAI holds no purchase funds, creates no stored balance and leaves card issuer rules in place. Visa's fiscal 2026 filing says over 60% of ecommerce transactions were tokenized as of 30 September 2026. Its $45,903m of fiscal 2026 revenue over $15,338bn of payments volume measured through 30 June 2026 works out to about 0.30%, with the reporting periods offset. That is a network revenue ratio. You do not have to replace a network to replace the checkout invitation.
Start with PayPal's own quarter. In the quarter ended 30 September 2026, transaction revenue from PayPal products and services fell approximately $240m year on year. Total payment volume was $487,138m, and transaction revenue of $7,711m gives a 1.5829% blended take rate ($7,711m ÷ $487,138m). In a scenario where 1% of quarterly volume leaves branded checkout, $487,138m × 1% = $4,871.38m represents $4,871.38m × 1.5829% = $77.1m of transaction revenue. In a scenario with the 0.87% transaction expense rate deducted, the amount left is $4,871.38m × (1.5829% − 0.87%) = $34.7m a quarter. In a scenario that also subtracts the 0.08% loss rate, the amount left is $4,871.38m × (1.5829% − 0.87% − 0.08%) = $30.8m. Both are hypothetical sensitivities, not an observed agent loss.
That sensitivity uses PayPal's companywide transaction rate; branded checkout yield is undisclosed. The PayPal-products line is the closest disclosed exposure, and it includes services besides checkout. Treating $77.1m as an agent's cash would be nonsense. The risk is replacing a consumer relationship with processing revenue inside the same company. You can preserve the transaction and still surrender the reason a customer chose you.
The merchant has a second problem. OpenAI's launch post suggests that businesses upload product feeds, configure their CDN or firewall to recognise signed requests, and add plugins or WebMCP tools. It says risk systems may still flag legitimate purchases made by agents. OpenAI's commerce guide says Wallet works with existing checkout pages without a Wallet-specific integration. The help article says "Ready to check out" confirms approval, while the merchant still has to place the order. Risk-system acceptance becomes commercial work, rather than mere back-office plumbing. As a processor, Braintree has a route to that work. The risk team becomes part of the checkout product.
Lily Varon of Forrester is a convincing opponent of the notion that consumers are ready to delegate payment. On 15 June 2026 she cited an April survey of online adults in the US, UK and Canada. "Three-quarters of these consumers say they are uncomfortable allowing an AI agent to complete a purchase and pay autonomously, even if they could set spending limits and rules in advance," she writes. She also points out that after more than a decade of Apple Pay, just one in four US online adults surveyed had used it in the previous three months. That is an argument for patience. This wallet still asks for biometric approval on each purchase. You are watching permissioned checkout, not automatic purchasing.
PayPal's quarterly filing supplies the sharper financial objection. In the quarter ended 30 September 2026, transaction revenue from Braintree rose approximately $270m year on year and from Venmo approximately $100m, while transaction revenue from PayPal products declined about $240m. PayPal attributes that product decline to mix, merchant incentives and foreign exchange fees. The filing gives Braintree's revenue gain, not its margin. PayPal's blended 1.58% transaction rate and Visa's approximate 0.30% network revenue ratio measure different businesses; subtracting them invents a revenue transfer. After all those numbers, the key one is absent. PayPal's 10-Q never uses the word agentic, and neither it nor Visa's 10-K discloses agentic volume or revenue; OpenAI discloses no wallet volume or revenue. Its 1.2B weekly audience as of 13 November 2026 measures reach alone.
There is a historical reason to resist paying for a disaster before it arrives. eBay announced on 31 January 2018 that Adyen would become its primary payment processor, while PayPal stayed an option. PayPal's 2021 filing shows eBay-platform customers falling from approximately 14% of PayPal revenue in 2019 to 6% in 2021. Yet PayPal revenue climbed from $17,772m to $25,371m, an increase of ($25,371m ÷ $17,772m − 1) × 100 = 42.8%. eBay's take rate rose from 9.12% in 2019 to 11.19% in 2021. The analogy breaks because eBay owned its marketplace and could migrate sellers; OpenAI's merchants remain in control of their sales. The precedent kills the easy short. The checkout concern survives.
Choice Layer Capture explains how a growing processor can mask a weakening checkout franchise. In the quarter ended 30 September 2026, PayPal's payment volume rose 6% year on year while transaction revenue from PayPal products and services declined approximately $240m. It held $13,337m in cash, cash equivalents and investments at quarter end. This business has the means to adapt; those resources are worth something. So is a customer relationship. Once an agent chooses the merchant, PayPal can keep the processing job while the buyer stops choosing its brand. A sound business can still be the wrong part to pay up for.
I would value PayPal as a resilient processor with a shrinking branded checkout business. Braintree deserves its hedge value, and the filing gives no Braintree margin to credit dollar for dollar. PayPal's fiscal 2026 Form 10-K, expected in February 2027, would prove me wrong if it showed year on year growth in transaction revenue from PayPal products and services. Until that happens, you should pay for a processor and discount the consumer payment brand, however reliably the underlying networks work.