# Agentic payments in regulated institutions: where the control layer sits in the payments value chain

When an AI agent can initiate a payment, the protocols prove who authorised what, but the decision to move money still belongs to the regulated institution. A walk through the value chain, the control points, and a worked example.

- Source: https://finray.tech/notes/agentic-payments-control-layer/
- Author: Finray
- Published: 2026-11-21
- Updated: 2026-11-21
- Reading time: 11 min
- Publisher: Finray Technologies Ltd, Cyprus Companies Registry HE 445903

---

## Short answer: where does the control layer sit?

In an agentic payment, the control layer sits at the regulated institution that holds the customer's account or issues the payment instrument, in the path of every payment instruction and outside the AI model. Protocols and card-scheme programmes can carry evidence of what a customer delegated to an agent. They do not remove the institution's duty to decide whether a specific payment should execute, and to prove that decision later.

This note walks through the payments value chain, marks the points where controls can sit, and explains why the institution's own payment path is the one that carries the most weight. It ends with a worked example using synthetic numbers and a plain account of what Finray software does and does not do today.

## What changes when an AI agent can initiate a payment?

Most payment systems were designed around a simple assumption: a person is present, looking at a trusted screen, and presses a button. The Google Cloud announcement of the Agent Payments Protocol (AP2) states this directly, noting that the rise of autonomous agents and their ability to initiate payments breaks that assumption. The same post frames the open problems as authorisation (proof that a user gave an agent authority for a particular purchase), authenticity (a merchant can be sure the agent's request reflects the user's intent) and accountability (who is responsible if a transaction is fraudulent or wrong).

Three things shift for an institution:

- **The initiator is software acting on a delegation.** The instruction arrives from an agent, not from a person at a screen, so the institution needs a way to know which person or company delegated what.
- **The volume and speed of instructions can change.** A person submits a handful of payments a day; a loop in software can submit many more. Limits that were implicit in human effort have to become explicit.
- **Instructions can be influenced by text.** An agent built on a language model reads content, such as invoices, emails and web pages, that can contain adversarial instructions. The model's output is therefore not a trustworthy statement of the principal's intent on its own.

None of this makes agentic payments a different kind of payment. It makes the payment instruction less self-evidently authentic, and puts more weight on the controls around it.

## Which parties sit in the value chain, and what does each one see?

A card-style agentic purchase passes through several parties. The exact set varies by rail, but the shape is stable enough to reason about.

| Position in the chain | What it typically contributes | What it can and cannot see |
|---|---|---|
| Principal (consumer or company) | Gives the agent a task and limits | Knows intent; may not see each step |
| Agent and agent platform | Interprets the task, finds options, prepares the payment | Sees the content it read; its output can be wrong or manipulated |
| Merchant or payee | Offers goods or services, requests payment | Sees the cart and the agent's claims about authority |
| Acquirer or payment service provider | Accepts the payment on the merchant's behalf | Sees merchant-side data, not the principal's mandate unless it is passed along |
| Scheme or payment rail | Routes and, for cards, tokenises credentials | Sees credential and transaction metadata |
| Issuer or account-servicing institution | Holds the account or credential, decides to authorise or execute, posts to the ledger | Sees the account, balances, history and limits; this is the last point before money moves |

Two observations follow. First, only the last position holds the customer's balance and the institution's own risk limits, so only it can evaluate a payment against them. Second, every earlier position can add useful evidence, but each sees only part of the picture.

## How do the protocols and scheme programmes fit in?

Several initiatives address how an agent's authority can be expressed and verified. They are best read as evidence-carrying layers that feed the institution's decision, not as replacements for it.

- **AP2.** The AP2 project site describes it as an open protocol for the agent economy, available as an extension for the Agent2Agent (A2A) protocol and the Universal Commerce Protocol; its core principles also call it an open extension for A2A and the Model Context Protocol. It uses mandates that are signed, tamper-evident digital objects. The Google Cloud post describes an Intent Mandate that captures a user's request and, for delegated tasks, rules such as price limits and timing, and a Cart Mandate that captures approval of an exact cart and price. The current specification on the project site uses a checkout mandate and a payment mandate instead, each with an open stage that captures constraints and a closed stage that captures a finalised transaction. The post says the sequence from intent to cart to payment creates a non-repudiable audit trail. Whether a given institution accepts that chain as evidence is that institution's decision.
- **x402.** The x402 site describes an open standard that uses the HTTP 402 Payment Required status: a server answers an unpaid request with 402, the client pays, and the request is retried. Stablecoin payments are the primary use case. It addresses how a machine pays a service; it does not by itself describe the institution-side controls that would wrap an account holder's payments.
- **Visa Intelligent Commerce.** Visa's launch release says tokenised digital credentials replace card details, that consumers can set spending limits and conditions, and that agent-linked credentials confirm a chosen agent may act on the consumer's behalf.
- **Mastercard Agent Pay.** Mastercard's launch release says its Agentic Tokens build on its existing tokenisation, that trusted agents will be required to be registered and verified, and that consumers keep control over what an agent may buy on their behalf.
- **Model Context Protocol (MCP).** MCP is a general protocol for connecting AI applications to tools and data. It is not a payments protocol, but a payment-capable tool can be exposed through it. Its specification states that hosts must obtain explicit user consent before invoking a tool and that descriptions of tool behaviour should be treated as untrusted unless they come from a trusted server.

The common thread is that each initiative strengthens one link: who the agent is, what the principal allowed, or how credentials travel. None of them can know the institution's current exposure to a counterparty, a sanctions hit that appeared this morning, or a policy that changed last week.

## Why does the institution's payment path carry the most weight?

Three reasons, all structural.

**Last point before money moves.** Upstream controls can be bypassed, misconfigured or simply absent for a given merchant or rail. The institution's own payment path is the one place every instruction from its customers must pass through.

**Accountability stays with the regulated entity.** Under PSD2, strong customer authentication is required where a payer initiates an electronic payment transaction, and for remote electronic payments the authentication must include elements that dynamically link the transaction to a specific amount and a specific payee. The directive also sets out rules on the payer's liability for unauthorised transactions. It treats a payment as authorised only if the payer has given consent (Article 64), and where a user denies authorising an executed payment it puts the burden on the payment service provider to prove the transaction was authenticated and accurately recorded (Article 72). How these provisions map onto an agent that initiates payments on a principal's behalf is not settled by the protocols above, and this note does not attempt to settle it. The point for design is narrower: whichever way the question is answered, the institution has to be able to show what the customer authorised and how the payment was checked.

**Supervisors apply existing expectations to AI use.** FINMA's Guidance 08/2024 describes expectations that institutions identify, assess, manage and monitor risks from AI, and lists operational and model risks, data risks, IT and cyber risks, third-party dependency, and legal and reputational risks among its focus areas. The EU AI Act, where a system falls within its high-risk category, requires that the system can be effectively overseen by natural persons while in use (Article 14) and technically allows automatic recording of events (Article 12). Application dates are staged, so check the current consolidated text before relying on a date.

## What belongs outside the model?

A language model is good at interpreting a request and drafting a proposal. It is a poor place to enforce a rule that must hold every time. The following should be evaluated by deterministic code in the payment path, with each decision recorded.

| Risk | Control outside the model | Who owns it |
|---|---|---|
| Agent exceeds the delegated amount | Per-payment and cumulative limits evaluated against the mandate in force | Payments product owner; risk sets the limits |
| Payment to an unintended or unknown payee | Payee allowlist or new-payee approval step; name and account checks where the rail supports them | Operations; compliance for policy |
| Prompt injection changes the payee or amount | Compare the executed instruction to the principal-approved cart or mandate; reject on mismatch | Engineering; security |
| Agent retries and pays twice | Idempotency key on every instruction; duplicate detection before posting | Engineering; payment operations |
| Principal denies authorising a payment | Signed mandate, agent identity, and approver identity stored with the instruction | Compliance; legal |
| Sanctions or fraud exposure on the counterparty | Screening and risk scoring before execution, independent of the agent | Financial crime function |
| Approval is rubber-stamped | Approver sees the proposal and the control results; approver cannot be the initiator; sample review of approvals | Control owner; internal audit |
| Agent credentials misused | Scoped, revocable credentials; a switch that suspends the agent's mandate | Security; payments product owner |
| Cannot reconstruct why money moved | Append-only decision log linked to the ledger postings | Finance; internal audit |

The table is a design aid, not a complete control framework. An institution's own risk assessment decides which rows apply and how strictly.

## Worked example: a corporate treasury agent (illustrative)

The figures below are synthetic. The example illustrates the control pattern; it is not a description of any one product's features. A company holds a CHF account with its bank. It lets an agent prepare supplier payments from approved invoices. The bank records this delegation as a mandate with these terms:

- Per-payment limit: CHF 5,000. Anything higher needs a named person's approval.
- Daily cumulative limit for the agent: CHF 20,000.
- Payees: only suppliers already on the company's approved list.
- Instruction identifiers: each payment carries an idempotency key.

On one morning the agent proposes four payments in sequence.

1. **CHF 1,850 to an approved supplier.** Within the per-payment limit; cumulative CHF 1,850. Executes.
2. **CHF 4,200 to another approved supplier.** Within the limit; cumulative CHF 6,050. Executes.
3. **CHF 6,300 to an approved supplier.** Exceeds the per-payment limit. The instruction is held and routed to a named approver, who sees the invoice reference, the proposal and the control results. The approver confirms. Cumulative CHF 12,350. Executes after approval.
4. **CHF 4,900 to a payee that is not on the approved list.** Within the amount limits, but the payee fails the allowlist check. Rejected, and a task is opened for the company to add the payee through its own approval step if it is legitimate.

Then the agent's connection drops and it resends payment 2 with the same idempotency key. The payment path recognises the key and returns the original result instead of paying again.

At the end of the day the institution holds, for each instruction: the mandate version, the proposal, the control decisions, the approver where one was needed, and the ledger posting. A reviewer can answer "why did CHF 12,350 leave the account?" without asking the agent. The explanation does not depend on the model's account of itself.

## How should an institution approach a pilot?

A staged approach keeps the risk proportionate while the evidence builds.

1. **Read-only.** The agent reads invoices and prepares proposals; nothing is submitted to the payment path.
2. **Propose with approval.** Every proposal needs a named person's approval before submission. Measure how often approvers change or reject proposals.
3. **Bounded delegation.** Low-value payments to allowlisted payees execute under a tight mandate; everything else still needs approval.
4. **Widen deliberately.** Extend limits only after control results, rejections and exceptions have been reviewed, and the change itself is approved under change control.

At each stage, the suspension switch and the decision log should work before the limits are raised, not after.

## What does Finray software do today, and what is design work?

**Today.** The Finray platform pages describe capabilities that are relevant building blocks. Corebanq provides an atomic double-entry ledger that rejects unbalanced postings at write time, payment operations across SIC, SEPA, SWIFT and instant rails with exception workflow as a structured object, an ISO 20022 translator with idempotency rules and replay-safe ingress, and role-based access control in which every action carries the actor and the policy version. XZiel provides transaction risk monitoring with a decision engine that includes policy checks, risk scoring and audit logging. Ordinis provides approval routing expressed as state machines and evidence captured as work happens.

**Not yet.** Finray does not currently describe a product in which an AI agent initiates payments through these platforms, and Finray does not claim agent integrations through the Model Context Protocol. Mandate objects for agents, agent-specific suspension switches and agent identity are design topics for institutions to assess, not features described on the platform pages. If you are planning an agentic pilot, the useful question to put to any vendor, including Finray, is which of the controls in the table above are enforced in its payment path today and which are expected to be built by the institution.

For the underlying concepts, see the glossary entries on [agentic payments](/intelligence/glossary/agentic-payments/), [agent mandates](/intelligence/glossary/agent-mandate/), [human-in-the-loop approval](/intelligence/glossary/human-in-the-loop-approval/), [agentic commerce](/intelligence/glossary/agentic-commerce/) and the [Model Context Protocol](/intelligence/glossary/model-context-protocol/). Platform pages: [Corebanq](/platforms/corebanq/), [XZiel](/platforms/xziel/) and [Ordinis](/platforms/ordinis/).

## Frequently asked questions

### Where should the control layer for agentic payments sit?

The control layer for agentic payments should sit at the institution that holds the account or issues the instrument, in the path of every payment instruction, outside the AI model. Protocols and scheme programmes can carry evidence of what a customer authorised, but the institution still has to decide whether to execute, and it needs to be able to prove that decision afterwards.

### Does a mandate replace strong customer authentication?

No. A mandate records what a customer delegated to an agent. In the EU, PSD2 separately requires strong customer authentication for defined cases and, for remote electronic payments, authentication elements that link the transaction to a specific amount and payee. How those rules apply to agent-initiated payments is a question for the institution's compliance function and its supervisor.

### Can the AI model itself enforce spending limits?

An AI model should not be the enforcement point for spending limits. A model can be instructed to respect a limit, but an instruction is not a control. Limits, payee lists, approval thresholds and duplicate detection need to be evaluated by deterministic code in the payment path, with the result logged.

### What evidence should an institution keep for an agent-initiated payment?

For an agent-initiated payment, an institution should keep at minimum: the mandate in force and its version, the identity of the agent and the principal, the proposed payment as the agent presented it, each control decision with the rule and policy version that produced it, any human approval with the approver identity, and the final ledger postings. Together these let a reviewer reconstruct why money moved.

### Is this note legal or regulatory advice?

No. It describes control design patterns and cites primary sources. Whether a given agent-initiated flow is a regulated payment service, and which authentication and liability rules apply, depends on the jurisdiction, the licence and the facts, and should be confirmed with counsel and the supervisor.

## Sources

- [Agent Payments Protocol (AP2), project site](https://ap2-protocol.org/)
- [Google Cloud blog, Announcing Agent Payments Protocol (AP2), 16 September 2025](https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol)
- [x402, open standard for internet-native payments](https://www.x402.org/)
- [Visa, Visa Intelligent Commerce launch release (30 April 2025)](https://investor.visa.com/news/news-details/2025/Find-and-Buy-with-AI-Visa-Unveils-New-Era-of-Commerce/default.aspx)
- [Mastercard, Agent Pay launch release (29 April 2025)](https://newsroom.mastercard.com/news/press/2025/april/mastercard-unveils-agent-pay-pioneering-agentic-payments-technology-to-power-commerce-in-the-age-of-ai)
- [Directive (EU) 2015/2366 (PSD2), EUR-Lex](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32015L2366)
- [Regulation (EU) 2024/1689 (AI Act), EUR-Lex](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689)
- [FINMA Guidance 08/2024, Governance and risk management when using artificial intelligence](https://www.finma.ch/en/news/2024/12/20241218-mm-finma-am-08-24/)
- [Model Context Protocol specification](https://modelcontextprotocol.io/specification/latest)
