Look again at the Bundesbank's architecture chart and read the bullet lists rather than the protocol boxes. At the communication layer the existing and potential actors are technology corporations, AI providers and cloud and infrastructure providers. Payment service providers appear one layer in, at transaction execution, and payment networks two layers in, at authorization and identity.
That ordering is the argument of this lesson. The Deutsche Bundesbank is saying that the outermost layer of the payment, the one every other layer has to pass through, is owned by companies that do not provide payment services.
Control without a payments license
The Bundesbank puts it directly: technology and AI providers could take on central roles in the emerging ecosystem without providing traditional payment services themselves. Their influence rests on control over decision-making, communication and authorization. As operators of the platforms agents run on, they determine which protocols and interfaces are available.
Follow the mechanism. An agent runs inside a framework. The framework decides which tools the agent can see, which protocols it speaks, and what actions it may take before any payment protocol is ever invoked. The Bundesbank calls this an upstream control instance. It sits before the mandate is checked, before identity is proved, before an instruction is formed. If the framework does not expose an instant credit transfer, the agent will never choose one, however good the price.
Those who control these layers have an influence not only over access to payment services, but also over the relevant data flows.
That second clause matters as much as the first. The communication layer sees every request and every response. Whoever runs it knows what agents ask for, what merchants offer, and what gets chosen.
Three kinds of actor, three kinds of interest
The Bundesbank distinguishes the incumbents from the newcomers from the infrastructure.
Payment networks and payment service providers are faced with integrating agent-based actions into processes built for human interaction, in order to keep their roles with their own standards. Their authentication and authorization logic assumes a person. Their response has been protocols and frameworks that open existing infrastructure to agents while keeping influence over identity and authorization standards. The Bundesbank also notes regional players building their own solutions, with the fragmentation risk that implies.
Technology and AI providers compete by providing the decision logic, the platforms and the interfaces. Broad partner alliances, the Bundesbank observes, can generate strong network effects and establish standards early. The launch lists attached to the protocols in Modules 1 and 4 are that observation in practice.
Cloud and infrastructure providers are the odd group out. Their interest in agentic payments is independent of which layer or protocol wins: more agents means more compute, more interfaces, more memory, more data processing. They can influence the identity and communication layers by implementing payment protocols at the network level, and the Bundesbank singles out one possibility as particularly attractive: verifying agent identities at the protocol level, similar to the certification of web domains. Because their economics do not depend on any one standard, they rarely compete head-on with the other two groups, which makes them useful partners to everyone.
Where this ends up
The Bundesbank draws two conclusions, and it is explicit that they are concerns rather than predictions.
Concentration. If a small number of central protocols become dominant, a small number of technology providers could control core functions from interpreting a user's intention to initiating and executing a payment. Payment service providers and users alike become dependent on external systems. Worse, an agentic payment provider could influence transaction decisions indirectly, through the model: the choice of merchant, the choice of payment instrument. The limited auditability of model output makes that influence hard to see.
Sovereignty. The providers currently offering agentic payment solutions are mainly non-European. For an infrastructure the Eurosystem treats as critical, the Bundesbank wants resilience and sovereignty through providers that are European or regulated in Europe, and sees interoperability, open interfaces and easier market entry as the way to get supplier diversity.
For a product team none of this is a reason to wait. It is a reason to know which gatekeeper you are standing behind, and what it can see.
Practical exercise: map your exposure
Allow 30 minutes. You will need access to your own product's documentation, or a competitor's public documentation if you are not currently shipping.
- List the operations (10 minutes). Write down every payment-related operation your product performs: list methods, quote, initiate, authorize, refund, check status. For each one, note whether an agent could call it today, and through what: a public API, an MCP server, a UCP profile, an agent card, nothing.
- Name the gatekeepers (10 minutes). For each operation an agent could call, name the platform or framework the agent would have to be running on to reach it. If the answer is "any", say what makes that true. If the answer is one company, write its name.
- Find the data (10 minutes). For the operations exposed, list what the integration point can observe: the request, the alternatives offered, the choice made, the price. Decide whether you are comfortable with that party seeing it, and whether your customer would be.
Keep the list. Module 4 asks which of those operations can verify a mandate and an identity, and Module 6 asks what happens to your margins when an agent can switch away from you at zero cost.
Key takeaways
Steps 1 to 3 are conversations, and the agent is party to two of them. The customer instructs the agent; the agent selects from a merchant; the merchant delegates checkout to a gateway without the agent in the room.
The communication layer is the gatekeeper. An agent can only use what has been exposed to it. Exposure is access control and operational translation at once.
MCP and A2A are pipes, not payments. They connect an agent to tools and to other agents. Mandate, identity and money live in the layers inside them.
Most merchants have not opened the door. On the October 17 scan, most of the 92 largest US storefronts had no published route an agent could use to reach step 2.
The gatekeepers are not payment companies. The Bundesbank places technology, AI and cloud providers at this layer, and warns of concentration and dependence on non-European providers.
Sources and further reading
- Deutsche Bundesbank: Agentic payments: when payments learn to think, Monthly Report, September 2026, the supplementary box on architecture and section 4 on market structures.
- Model Context Protocol specification, Anthropic.
- Agent2Agent Protocol, Google.
- Major Matters: Merchant readiness tracker, weekly scan of the 92 largest US storefronts.
Next: Module 3, Decision. Steps 2 and 4: what the agent chooses, what steers the choice, and why a probabilistic decision has to be made deterministic before it reaches a rail.
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.