Module 5: Execution and Settlement
Steps 7a to 10 in the Bundesbank flow. Layers: transaction execution and settlement. Actors: the payment service providers of the payer and the payee, the payment system, the merchant, the gateway, the AI agent and the customer.
Modules 1 to 4 ended with a provider at step 6 holding a verified instruction and a verified agent. Nothing has moved yet. Module 5 is about what happens when it does, and about the part of the flow that is easiest to skip in a protocol map: the confirmations that have to travel back before anyone, human or agent, can treat the purchase as done.
By the end of this module you will be able to:
- Walk steps 7a to 10 and say which actor acts at each, in which layer, and what the agent sees
- Explain why the Deutsche Bundesbank expects the settlement layer to be complemented rather than replaced
- Describe what ACP, UCP and x402 each do at the transaction execution layer, factually and without ranking them
- Compare settlement on card rails, instant credit transfer rails and stablecoin rails on finality, reversibility, speed and cost
- Trace the confirmation chain for a given rail and say which confirmation a product should treat as the end of the payment
Four steps, six messages
The Bundesbank's Chart 5.1 draws the end of the flow as a cluster of short arrows at the bottom and three long ones traveling back up. The step numbers run to 10, but step 7 has three parts, so there are six messages to account for.
Read the layer column. Steps 7a to 8 are gray in the chart: settlement, the one layer the Bundesbank says agentic payments touch only indirectly. Step 9 is teal, the communication layer from Module 2, because it is the only one of these messages the agent receives. Step 10 has no color at all. Delivery is not a payment event, and the chart is honest about that.
Now read the actor columns. The agent, at the center of every step from 0 to 6, appears once in the last six messages, as the recipient of step 9. The customer appears once, at step 10, receiving a parcel. Everything in between is banks talking to banks and a bank talking to a merchant. That is the Bundesbank's point about where agentic payments operate: upstream of clearing and settlement, in decision-making, communication, mandates and identity.
The agent chose, initiated and proved. Then it waits, like everyone else, for the system to tell it the money arrived.
The instruction, and what the layer is for
Step 7 does not start with the agent. It starts with the instruction the provider accepted at step 6. The Bundesbank's transaction execution layer generates specific payment instructions and governs how the mandate from the authorization layer is translated into an executable one. Module 4 made the point that the instruction is derived from the mandate. Module 5's point is the other half: the instruction is what the rail acts on, and the rail does not read the mandate.
The chart gives the layer a subtitle that is easy to miss: "Payment order, can also extend beyond authorisation and identity." That is why the authorization and identity boxes are drawn inside the transaction execution box in Chart 5.2. An execution protocol can carry the mandate and the identity proof with the order, or it can carry only the order and leave those to another protocol. The next lesson shows that the three protocols the Bundesbank names here take different positions on exactly that.
The actors the Bundesbank lists at this layer are payment service providers, technology corporations, AI providers and decentralized financial services. At the settlement layer they are payment systems, payment networks and decentralized financial services, and the protocol boxes say only "existing standards" and "x402". Execution is contested. Settlement, for now, mostly is not.
Why settlement is complemented, not replaced
The Bundesbank's sentence on the settlement layer is one of the few flat predictions in the article. It will remain largely independent and is more likely to be complemented than replaced by agentic payments. The reasons are structural: high barriers to entry, network effects and considerable regulatory requirements. Agentic payments are unlikely to establish entirely new settlement infrastructures, but could have an impact on their use.
Two consequences follow, and both are in the text. The change is in the interfaces: structural changes are expected less in settlement itself than in the opening and standardization of interfaces to traditional payment infrastructures for autonomous agents, which is Module 2's gatekeeper question arriving at the bottom of the stack. And the exception is already running: the x402 protocol already supports settlement on new DLT-based infrastructures, such as stablecoins. That "however" in the Bundesbank's text is the reason x402 appears in both the execution box and the settlement box. It is the one named protocol that reaches the money.
A confirmation is not finality
Here is the practical problem Module 5 exists to name. Three different parties send three different confirmations, in three different layers, and only one of them means the money has moved for good.
Step 8, payment confirmation, is sent by the payee's provider once the credit has landed. On an instant credit transfer that credit is final and irrevocable. On a card rail the equivalent message at this point in the flow is an authorization: a hold, not a settled credit, and one the cardholder can still dispute. Same step number, different legal weight.
Step 9, order confirmation, is sent by the gateway to the agent. It is a commerce message, not a payment message. Under the execution protocols in the next lesson it is a checkout session moving to a completed state and an order record being created. It says the merchant is proceeding as though the funds will arrive, not that they are final.
Step 10 is the goods. An agent that reports "done" to its principal at step 9 is reporting the merchant's belief. An agent that waits for step 10 is reporting an outcome. Which one the mandate from Module 1 asks for belongs in the lifecycle field, under reporting, and most mandates I have seen do not say.
The gap between those three is where disputes live. I have written about what the agentic commerce stack is missing when things go wrong in The Agentic Commerce Stack Has Every Layer Except the One That Matters When Things Go Wrong, and the first question in Who Pays When the Agent Pays Wrong? is a question about which of these three messages a consumer can rely on.
One design consequence before the protocols. The stylized flow sends the agent a single message after initiation: step 9. A product that wants the agent to act on finality rather than on a merchant's belief has to route a settlement-layer signal to the agent as well, and the chart does not draw one. On an instant rail the payer's provider knows at step 7b that the debit is final. On a card rail it knows only that an authorization exists. On a stablecoin rail the agent may have submitted the transaction and can read the chain itself. That is the real reason the choice of rail belongs in the mandate's limits field: it decides what "paid" means to the agent, and so what the agent can truthfully report to the person who gave it the mandate.
Next lesson: ACP, UCP and x402. Three protocols, one job: how does the instruction the provider approved become something a rail will carry?
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.