The Deutsche Bundesbank opens its section on market structure with a conditional that deserves to be read slowly. If AI agents are able to switch between providers, protocols and payment rails with almost zero friction, the traditional lock-in effects in payments and the price-setting power of established intermediaries lose their significance.
Everything in that sentence happens at the decision layer. Switching is a choice. Zero friction means the choice is free to make. The question for this lesson is who gets to influence it.
Four parties with a hand on the choice
The model provider. The Bundesbank's sharpest warning is here: agentic payment providers could indirectly influence transaction decisions, such as the choice of merchant or means of payment, through the models used. If the model has been tuned, prompted or paid to prefer a merchant or an instrument, the agent's "choice" at steps 2 and 4 is partly the provider's. The limited transparency of model output makes that hard to detect, which is why the Bundesbank pairs this concern with its concentration concern. Few models, little auditability, large influence.
The merchant. Merchants can guide the use of specific payment instruments and influence agent preferences through attractive conditions. A discount for paying by instant transfer is an instruction to every agent optimizing for price. The Bundesbank also expects merchants to need machine-readable catalogs and agent-compatible interfaces, which is the Module 2 observation again: you cannot influence a choice you are not visible to.
The gateway and the payment service provider. At step 4 the gateway decides which instruments are offered at all. An instrument not offered is not chosen. Payment service providers, the Bundesbank notes, could use the same routing logic for their own and their customers' payments, moving decisions that were based on contractual and account relationships and network dependencies onto a data-driven, dynamic basis.
The customer, through the mandate. The mandate's limits field is the customer's hand on the choice: permitted instruments, preference order, ceilings. It is the only one of the four that is written down and verifiable. Which is the argument for writing it carefully, and for products that surface it rather than burying it in a settings page.
Data access becomes a competitive factor. Asymmetric availability of data may lead to market distortions.
That is the Bundesbank on why the steering fight will be won on information. The agent chooses from what it can see. Whoever supplies the better feed wins more choices.
The case: Instant Checkout, and who actually won
The clearest public test of the decision layer so far came from OpenAI and Stripe. The Agentic Commerce Protocol, published in September 2025, powered Instant Checkout in ChatGPT: the agent found the product, the agent took the payment, inside the chat, with a shared payment token passed to the merchant. Discovery and checkout in one place, under one party's control.
On March 24, 2026 OpenAI discontinued that first version. Its own explanation was that it did not offer the level of flexibility it aspired to provide. Merchants would run their own checkout; ChatGPT would handle product discovery. I wrote it up as OpenAI Abandoned Instant Checkout. The Merchants Won, and the title is the point for this module. Step 2, product selection, stayed with the model provider. Steps 3 and 4, the delegated checkout and the payment method, went back to the merchant and its gateway.
Read that through the Bundesbank's four parties. The model provider kept the choice of what to buy. The merchant recovered the choice of how to pay. The split landed exactly on the boundary between the two decisions in the last lesson. The earlier evidence pointed the same way: Walmart's Instant Checkout had not found users, and Anthropic's shopping agent left the checkout out altogether.
The protocol survived the product. ACP's specification remains published, and Module 5 covers it alongside UCP as the two de facto standards the Bundesbank sees at transaction execution. What the episode settled was narrower and more useful: who holds which decision is negotiated, product by product, and the merchant has more say over step 4 than the first wave assumed.
Practical exercise: write the decision policy
Allow 30 minutes. Take the household mandate you wrote in Module 1.
- Step 2 policy (10 minutes). For product selection, write the agent's rules in order: what counts as the same product across the two supermarkets, how it trades price against delivery date, when it substitutes, and when it asks. Then write one line on what a merchant could put in its catalog that would change the agent's choice.
- Step 4 policy (10 minutes). For payment method, write the preference order the mandate implies, the tie-break, and the conditions under which the agent may depart from it (the pharmacy stops accepting instant transfer, from Module 1's case B). State the ceiling above which the agent must hand the decision to a person.
- The deterministic check (10 minutes). Write the test that runs after the agent has chosen and before anything is initiated: amount within ceiling, instrument on the permitted list, merchant in scope, mandate not revoked. Make every line a yes or no. If any line needs the model to answer it, move it into the policy above and replace it with something that does not.
Keep both documents. In Module 4 the check becomes what the payment service provider verifies at step 6.
Key takeaways
Two decisions, two counterparts. Product with the merchant at step 2; instrument with the gateway at step 4. The chart colors both as decision-making.
No standard, only a framework. The decision layer has no payment-specific standard. The agentic framework, the model and the prompt define what the agent can choose.
Probabilistic in, deterministic out. Agents decide by likelihood. Before initiation the choice must become a verifiable instruction checked against the mandate in code. The Bundesbank's governance rule and the MAS runtime design are the same shape.
Herd behavior and collusion are decision-layer risks. Homogeneous models and data make correlated choices. Neither failure is visible inside one transaction.
Four hands on the choice. Model provider, merchant, gateway, customer. Only the customer's, the mandate, is written down and verifiable.
The merchant kept step 4. Instant Checkout's retreat left discovery with the model provider and payment with the merchant.
Sources and further reading
- Deutsche Bundesbank: Agentic payments: when payments learn to think, Monthly Report, September 2026, sections 1, 4, 5 and 6.
- Monetary Authority of Singapore: Safeguards for Agentic Finance at Runtime, July 2026.
- Stripe: Stripe and OpenAI launch Instant Checkout and the Agentic Commerce Protocol, September 2025.
- OpenAI: Powering product discovery in ChatGPT, March 2026, the discontinuation of the first Instant Checkout.
Next: Module 4, Authorization and Identity. Steps 5 and 6: the agent initiates, and the payment service provider demands proof of who it is and what it may do.
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.