Skip to content
In the news TRM Labs × Finray — audit-ready crypto transaction monitoring for banking
Finray

Glossary · Agentic finance

What are agentic payments?

Agentic payments are payment instructions produced by an AI agent under a delegation: the agent prepares or submits them for a person or company, inside limits that principal set. The institution that moves the money still has to authenticate, authorise, record and be able to explain each payment.

Updated

What are agentic payments?

Agentic payments are payment instructions produced by an AI agent under a delegation: the agent prepares or submits them for a person or company, inside limits that principal set. The agent interprets a goal, for example “pay the approved supplier invoices due this week”, and produces payment instructions. The regulated institution that holds the account or issues the credential still has to authenticate, authorise, record and be able to explain each payment.

The term describes who or what starts the payment, not a new payment rail. The money moves over the same rails as before: cards, bank transfers, instant payments or, in some designs, stablecoin transfers. What changes is the nature of the instruction and the evidence needed to trust it.

How do agentic payments work?

A typical flow has five steps, whatever the rail.

  1. Delegation. The principal gives the agent a task and a set of limits: amounts, payees, time windows, and sometimes instruments. In protocol terms this is often expressed as a mandate.
  2. Interpretation. The agent reads context such as invoices, a shopping request or a budget, and decides what to propose.
  3. Preparation. The agent produces a specific instruction: payee, amount, currency, reference.
  4. Checks and approval. Limits, payee checks, screening and, where required, a person’s approval are applied. These checks should run in deterministic code, not in the model.
  5. Execution and record. The payment executes, ledger postings are made, and the evidence is stored: what was delegated, what was proposed, what was checked, who approved.

The first and fourth steps are where agentic payments differ most from conventional ones. A human clicking “pay” is replaced by a delegation recorded earlier and a set of checks applied later.

Why do agentic payments need new evidence?

Existing payment systems tend to assume a person is present at a trusted screen when a payment is authorised. The Google Cloud announcement of the Agent Payments Protocol (AP2) puts it this way: the rise of autonomous agents and their ability to initiate payments breaks that assumption. It names three questions that have to be answerable: authorisation (did the user give the agent authority for this purchase), authenticity (does the request reflect the user’s real intent) and accountability (who is responsible if the transaction is fraudulent or wrong).

AP2’s answer is a chain of signed records, called mandates, from the user’s instruction through to payment, described as an audit trail that answers those questions. The AP2 project site calls the underlying objects tamper-evident, cryptographically signed digital credentials. Other initiatives take different approaches. x402, for example, uses the HTTP 402 Payment Required response so that a client can pay for a request and retry it, with stablecoin payments as the primary use case. These are different tools for different problems, and none of them removes the need for the institution’s own checks.

What risks do agentic payments introduce?

RiskWhat can go wrongTypical control
Over-delegationThe agent has authority well beyond the taskNarrow mandate; per-payment and cumulative limits
Manipulated inputText in an invoice or web page redirects the paymentCompare execution to the approved proposal; reject mismatches
Duplicate paymentThe agent retries after a timeoutIdempotency key on every instruction
Unknown payeeThe agent pays a new or look-alike accountPayee allowlist; approval for new payees
Weak accountabilityNobody can show who authorised a paymentStored mandate, agent identity and approver identity
Opaque decisionsA reviewer cannot see why a payment happenedAppend-only decision log linked to ledger postings

What do regulators say?

There is no single rulebook for agentic payments. Existing rules apply to the payment and to the institution’s use of AI.

In the EU, PSD2 requires a payment service provider to apply strong customer authentication where the payer initiates an electronic payment transaction, and for electronic remote payments requires authentication elements that dynamically link the transaction to a specific amount and payee. It also sets out rules on the payer’s liability for unauthorised transactions. The directive does not mention AI agents, and how its provisions apply to an agent that initiates payments is a matter for the institution’s compliance function and its supervisor.

In Switzerland, FINMA’s Guidance 08/2024 sets out expectations that supervised institutions identify, assess, manage and monitor the risks arising from AI, with operational and model risks, data risks, IT and cyber risks, third-party dependency and legal and reputational risks among its focus areas. FINMA describes the document as guidance and notes that most institutions are early in building the relevant governance.

What is a worked example?

The numbers are synthetic and for illustration only. The example illustrates the control pattern; it is not a description of any one product’s features. A company lets an agent prepare supplier payments. The bank-side mandate sets a per-payment limit of CHF 5,000, a daily limit of CHF 20,000 and approved suppliers only.

The agent proposes EUR 3,100 to an approved supplier. Converted at an illustrative rate of 0.94, this is CHF 2,914, inside the limits, so it executes and is posted. It then proposes CHF 7,400 to another approved supplier. That exceeds the per-payment limit, so the instruction is held until a named approver confirms it. A third proposal, CHF 900 to an account that is not on the approved list, is rejected and logged. If the agent resends the first instruction with the same idempotency key, the payment path returns the original result instead of paying again.

Every outcome has a record of the mandate in force, the proposal, the control decision and any approval. None of it depends on the agent’s own description of what it did.

What belongs outside the model?

Controls that must hold every time should be enforced by deterministic code in the payment path, with results logged:

  • Limits. Per-payment, daily and cumulative limits, evaluated against the mandate in force.
  • Identity. Which agent, acting for which principal, under which credential.
  • Approvals. Thresholds and rules for routing instructions to a person, with the approver distinct from the initiator.
  • Idempotency. A unique key per instruction so retries cannot create duplicates.
  • Screening. Sanctions and risk checks that do not depend on the agent.
  • Audit trail. Append-only records tying the decision to the ledger posting.
  • Suspension. A way to revoke an agent’s authority immediately.

A language model can be asked to respect a limit, but a request is not an enforcement mechanism.

How does Finray relate to agentic payments?

Today. Finray’s platform pages describe building blocks relevant to the controls above. Corebanq provides an atomic double-entry ledger, payment operations across SIC, SEPA, SWIFT and instant rails with structured exception workflow, an ISO 20022 translator with idempotency rules, and role-based access control where every action carries the actor and the policy version. XZiel provides transaction risk monitoring with policy checks, risk scoring and audit logging. Ordinis provides approval routing and evidence capture.

Not yet. Finray does not describe a product in which an AI agent initiates payments, and it does not claim agent integrations through the Model Context Protocol. Agent mandates and agent identity are design topics for an institution to assess with any vendor.

For a longer treatment, see the note Agentic payments in regulated institutions. Related terms: agentic commerce, agent mandate, human-in-the-loop approval and Model Context Protocol.

Frequently asked questions

What are agentic payments?
Agentic payments are payment instructions produced by an AI agent under a delegation: the agent prepares or submits them for a person or company, inside limits that principal set. The agent does the interpretation and preparation; the institution that holds the account or credential still decides whether the payment executes.
How are agentic payments different from automated or scheduled payments?
A scheduled payment repeats a fixed instruction. An agent interprets a goal, such as paying approved invoices, and decides which payments to propose at run time. That flexibility is why the limits, approvals and audit trail have to be explicit rather than implied by a fixed schedule.
Are agentic payments regulated?
A payment an agent initiates is generally subject to the same rules as the same payment initiated by a person on the same rail. In the EU, PSD2 sets authentication and liability rules for payment services, and its recitals say the definition of payment services should be technologically neutral. Whether a particular agent arrangement counts as a regulated payment service depends on the facts and the jurisdiction, and should be confirmed with counsel and the supervisor.
Can an AI agent hold money?
An agent is software acting under delegated authority. Accounts and credentials belong to the person or company that opened them, and the agent uses them within a mandate. It does not become an account holder.
What is the safest way to start?
Start with the agent preparing proposals that a named person approves, measure how often approvers change them, and widen delegation only under limits that are enforced outside the model and under change control.

Sources

Published · Updated · All terms

Certificate of Registration NQA · UKAS Management Systems
ISO/IEC 27001:2022 Certificate of Registration issued by NQA to Finray Technologies Ltd, certificate number 215646, valid 21 October 2025 to 21 October 2028
Search
Type to search across Finray, products, company, and journal.

    Press Esc to close · ↵ to open the highlighted result.

    Book a briefing Step 1 of 3

    Your details

    Who is requesting the briefing.