How this course works
Most explanations of agentic payments start with a protocol map. I am going to start with a single transaction and follow it from the moment a person hands authority to a software agent to the moment the goods arrive. Ten numbered steps, six technical layers, one payment.
The reference flow comes from the Deutsche Bundesbank, whose September 2026 Monthly Report article, Agentic payments: when payments learn to think, lays out a stylized instant credit transfer executed by a fully autonomous AI agent. I use the Bundesbank's numbering throughout so you can cross-reference the original.
Decision-making becomes part of the payment process. In that sense, payments learn to think.
That one sentence from the Bundesbank is the thesis of the whole course. Every module asks the same question of its step: what decision is being made here, by whom, on what authority, and who carries the risk if it is wrong?
At each layer I also name the protocols and standards in play, as the Bundesbank catalogs them: MCP, A2A, AP2, TAP, Agent Pay, ACP, UCP and x402. I describe what each one does. I do not rank them. Where the Bundesbank assesses which protocols are leading, I report that assessment as theirs.
If you want the wide view before the close-up, my guide to AI agents, Software That Acts, covers how agents work in general, and What Are Agentic Payments? covers the vocabulary this course assumes.
Module 1: The Mandate
Step 0 in the Bundesbank flow. Layer: authorization. Actors: the customer (the payer) and the AI agent.
Module 1 runs across four lessons. By the end of it you will be able to:
- Explain why an agentic payment begins at step 0, before any purchase order exists, and why that makes decision-making a technical layer for the first time
- Describe the autonomy spectrum the Bundesbank uses, from agent as execution tool to agent as decision-maker, and place a given product on it
- Specify a payment mandate using the five fields a machine-readable mandate has to carry
- Locate the mandate in the six-layer architecture and explain why the authorization layer is strategically the most contested
- Describe how AP2, TAP and Agent Pay each approach mandates, factually and without ranking them
- Identify the three failure modes that originate at step 0 and the regulatory questions they raise under PSD2
The payment that starts before the purchase
In the traditional model there is no step 0. A card payment starts when the cardholder presents a card. A credit transfer starts when the payer keys in an IBAN. Even a direct debit, the closest existing analog, starts with a mandate but then runs on fixed amounts and fixed dates. The person decides, then the system executes.
The Bundesbank's agentic flow puts a new step in front of all of that. Step 0 is Mandate: the customer grants the AI agent the authority to act. Only then does the agent receive a purchase order (step 1), select a product (step 2), take over the checkout (step 3) and work through payment method selection, initiation and authorization (steps 4 to 6). The money itself does not move until step 7.
That ordering matters because of what the agent is allowed to do in between. With a high level of autonomy the agent can select the payment instrument, the timing and the rail, and can execute within the mandate without the person approving each payment. The Bundesbank's phrase for this is that decision-making becomes part of the payment process.
In the traditional architecture, decision-making was never a technical layer. Selection and payment were a person's job. Autonomous agents move that job into software, and software needs a specification.
The mandate is that specification. Everything downstream, from which merchant the agent can talk to, to whether a payment service provider will accept the instruction, to who is liable when something goes wrong, inherits its limits from what was written at step 0.
The autonomy spectrum
The Bundesbank is careful to say that agentic payments are not one thing. How much the architecture changes depends on how much authority the agent holds. I find it useful to think of this as a dial with three settings rather than a switch.
The Bundesbank expects practice to settle on the middle setting, with autonomy restricted by predefined frameworks such as maximum amounts or permissible payment instruments. Their comparison is to the mandate in direct debit. That comparison is the right one to hold on to, because it tells you what a mandate for an agent has to improve on: a direct debit mandate names a creditor and an account and little else. An agent mandate has to describe behavior.
Why this matters for product people
Where your product sits on the dial determines which regulatory questions you inherit. The Bundesbank's regulatory annex says low-autonomy models, where the customer still authorizes each payment, are generally compatible with PSD2 already, possibly with licensing requirements. High-autonomy models, where the customer leaves the loop, raise the open questions covered in the last lesson of this module. Pick your setting deliberately.
Next lesson: the five fields a mandate has to carry before a payment service provider can use it.