Under What Conditions: Central Banks, Sovereignty and the Closing Question

The Deutsche Bundesbank's conclusion contains the sentence this course has been working toward. The crucial point, it says, is not simply whether agentic payments will be used in the future, but under what conditions they will be used. Everything in Modules 1 to 5 was a condition: a mandate with five fields, a deterministic check, a verifiable anchor, a rail written into the limits. This lesson is about the conditions a central bank adds.

Three roles

The operator row touches Module 5 directly. The Bundesbank expects financial market infrastructures to face higher transaction volumes and frequencies, because agentic payments prefer cost-effective instant payments and favor means that support usage-based and machine-to-machine business models. Some infrastructures have already prepared, and its footnote names the Monetary Authority of Singapore's runtime safeguards from Module 3 and the Canton Network's agentic ledger as examples of the kind.

Then a sentence that is easy to read past. New forms of central bank money, for example a CBDC, and newly developed settlement systems should account for the use of AI agents early on, when they are first being developed. A central bank designing digital money is being told, by a central bank, to design it for software payers.

The systemic worry, stated plainly

The Bundesbank opens its central-bank section with the risk it considers most prominent. Agentic payments could amplify bank run risks, because the capacity for autonomous action combined with collective behavior can jeopardize financial stability. Then the qualification: given 24/7 instant settlement and developments in digital finance, this threat is not, in structural terms, a new one. Supervisors are told to take due account of it all the same.

That is Module 3's herd behavior at the scale of a banking system. In the retail flow it meant every household agent picking the same rail on the same morning. In the deposit market it means every treasury agent moving liquidity at once. The Bundesbank's institutional section adds that agents trained on historical data could fail to recognize an unfamiliar event such as a financial crisis, and that widespread use of similar algorithms could fail to internalize external effects like gridlock and free-riding. Its answers are adjusting agents' target functions, transparency in their responses, calibrating mandate and autonomy to the degree of human involvement, and escalation mechanisms that trigger human review in unusual constellations of data. Every one of those is a mandate field or a checkpoint. The systemic answer and the retail answer are the same answer at different scales.

Sovereignty

The sovereignty paragraph is short and the Bundesbank puts it on its own. Agentic payments could further reinforce the dominance of non-European market participants, which could hamper competition and weaken European sovereignty. Payments are critical infrastructure, and the Eurosystem strives to ensure resilience and sovereignty through providers that are from Europe or regulated in Europe. At present, it says, mainly non-European participants are offering agentic payment solutions. The remedies it names are interoperability, open interfaces and easier market entry.

Look back at Chart 5.2 with that paragraph in mind. Every protocol in its boxes, at every layer, is published by a company headquartered in the United States. The Bundesbank's introduction already said the key question is not only which technology will prevail but who will shape the decision-making and authorization layers. Its conclusion repeats the point: those who shape these layers will be in a position to exert significant influence over how payments work in future.

I would add one thing. The remedies it names, interoperability and open interfaces, are the same remedies Module 4 reached for the hand-off problem between anchors and Module 2 for the gatekeeper problem. A provider that publishes which anchors it verifies, exposes its operations through open protocols and accepts more than one execution standard is doing the sovereignty work whether or not it thinks of it that way.

The rule it wants applied

Three principles close the Bundesbank's regulatory argument.

Same business, same risks, same rules. As a rule, this should apply, to ensure fair competition and avoid regulatory arbitrage. The framework is technology-agnostic by design; the point is that new non-financial actors doing the business of payments should carry the rules of payments.

Prudential proportionality. Innovation that boosts efficiency should be facilitated. A footnote gives the balance: the same rules generally apply to small and large players, but the degree of supervisory compliance required increases with size.

Probabilistic in, deterministic out. Payment systems, and even DLT systems, rest on deterministic rules, unambiguous accountability and legal certainty. Agentic payments decide probabilistically. That is not inherently risky, the Bundesbank says, but it makes outputs unpredictable and hard to reproduce. Customers expect auditable, logical transactions, and if random outcomes, such as the order in which transactions are processed, have economic consequences, there may be legal, supervisory or balance-sheet ramifications. Hence the governance framework from Module 3, translating probabilistic decisions into binding, deterministic processes before execution, and clear rules about what agents may and may not do, especially for payment authorization and personal data where the user no longer authorizes directly.

The third principle is the one the whole course has been built on. The agent may decide by probability. What reaches the rail is an instruction someone can verify.

What three other official papers add

The Bundesbank is not alone, and its reference list points to the others.

The IMF's three layers map onto the Bundesbank's six without strain: intent is decision-making, authorization is the mandate and identity pair, settlement is settlement. Two institutions, writing months apart, drew the same seam between the probabilistic part and the deterministic part. That seam is step 6.

The whole anatomy on one page

Lesson 1 said every module would ask the same question of its step: what decision is being made, by whom, on what authority, and who carries the risk if it is wrong. Here is the answer for every step.

Read the risk column once more. The customer appears in seven of the 11 rows. That is lesson 17's paradox as a flow. Simpler outside, and the outside is where the risk collects unless the conditions in the other columns are written down and checked.

Practical exercise: write the conditions sheet

Allow 40 minutes. This is the course's final exercise. It uses everything you kept: the household mandate, the exposure map, the decision policy and deterministic check, the step 6 check, and the three routed cases. If you did not keep them, use your own product.

1. The mandate (5 minutes). One line per field: principal, agent, scope, limits, lifecycle. Mark any field your product leaves to the model's context window rather than a verifiable record.
2. The anchor (5 minutes). Which of the three anchors from Module 4 can your provider, your merchants and your framework all verify? If none, write which hand-off breaks.
3. The rail (5 minutes). For your most common transaction, name the rail, the finality time, and whether the customer has a route back. Then write the mandate line that makes that choice explicit.
4. The checkpoint (5 minutes). Where, between the agent's decision and initiation, does a deterministic check run? Name the threshold above which a person reviews. If there is no checkpoint, say so.
5. The record (10 minutes). List what is logged at step 6 and what it would prove in a dispute six months later. Then write what it would not prove, using the Bundesbank's footnote: the log shows what, not why.
6. The route back (10 minutes). For the case where the agent was wrong, write who the customer calls, which rule allocates the loss, and whether your product has a dispute path or only a goodwill one. Then write the sentence you would put in the product's terms to say so.

Six lines. If you can write all six without hedging, your product is operating under conditions. If you cannot, you know which condition is missing and which module to reread.

Key takeaways for the course

Agentic payments start at step 0. The grant of authority is new, and everything downstream inherits its limits.

Decision-making became a technical layer, and it has no standard. The agent decides by probability; the framework defines what it can attempt; a deterministic check must sit between the choice and the rail.

The communication layer is the gatekeeper. An agent can only use what has been exposed, and most large merchants had not opened the door when this course was written.

Identity and authorization are one arrow. Three anchors, two mechanisms, each relying on a registry someone runs. The provider at step 6 carries the burden of proof and will not execute what it cannot verify.

Settlement is complemented, not replaced. The execution protocols differ on whether they carry the mandate; the rails differ on finality and the route back. Write the rail into the mandate.

Simpler outside, more complex inside. The named risks land mostly on the customer, the margin pressure on the dearer rails, and the contest on the layers where decisions are made about access, identity, authorization, data and rail.

The condition is the product. Mandate, anchor, rail, checkpoint, record, route back. The Bundesbank's question is not whether, but under what conditions. Those are the conditions.

Glossary

Sources and further reading

• Deutsche Bundesbank: Agentic payments: when payments learn to think, Monthly Report, September 2026. The reference text for the whole course. For Module 6 read sections 2, 4, 5, 6 and 7, and footnotes 9, 29, 30 and 44.
• Financial Stability Board: The Financial Stability Implications of Artificial Intelligence, November 14, 2024.
• Bank for International Settlements: AI agents for cash management in payment systems, Working Paper 1310, November 26, 2025, Aldasoro and Desai.
• International Monetary Fund: How Agentic AI Will Reshape Payments, IMF Note 2026/004, April 2026, Davidovic and Tourpe.
• Monetary Authority of Singapore: Safeguards for Agentic Finance at Runtime, July 2026.
• Google Developers Blog: Under the Hood: Universal Commerce Protocol, January 11, 2026, the business as merchant of record.
• TechCrunch: OpenAI's plans to make ChatGPT more like Amazon aren't going so well, March 24, 2026, the Instant Checkout pivot.
• European Central Bank: Instant payments and TIPS, the finality figures in the glossary.
• Major Matters: The Bench, agents under a signed mandate, and the entry I Put a 20B Model on My Laptop Under a Bank Mandate.
• Major Matters: Who Pays When the Agent Pays Wrong?, Europe's banks are building agents on borrowed infrastructure, and Software That Acts: The Major Matters Guide to AI Agents.

The Bundesbank says the question is not whether agentic payments will be used, but under what conditions. So, for the next payment your product lets an agent make: what decision is being made, by whom, on what authority, and who carries the risk if it is wrong?

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.