# What is an agent mandate?

An agent mandate is a recorded grant of authority to an AI agent: what it may do, up to which amounts, with which counterparties and for how long, and who granted it. It is the artefact that lets an institution check an agent's request against what was delegated.

- Source: https://finray.tech/intelligence/glossary/agent-mandate/
- Published: 2026-11-21
- Updated: 2026-11-21
- Publisher: Finray Technologies Ltd, Cyprus Companies Registry HE 445903

---

## What is an agent mandate?

An agent mandate is a recorded grant of authority to an AI agent: what it may do, up to which amounts, with which counterparties and for how long, and who granted it. Its purpose is to give an institution, a merchant or an auditor something concrete to check an agent's request against.

The word "mandate" is used in more than one way. In banking it can mean a direct-debit authorisation or a legal mandate to act. In agentic payments it usually means a technical record of delegated limits. This entry uses the technical sense and notes where the legal question begins.

## How do mandates appear in the main protocols?

**AP2.** The Agent Payments Protocol site describes two kinds of mandate. A checkout mandate records the items and purchase details negotiated between agent and merchant; a payment mandate authorises payment against a specific instrument. Each has an open stage, which captures constraints and goals such as budget and allowed instruments, and a closed stage, which captures authorisation for a specific finalised transaction. The site describes the underlying objects as tamper-evident, cryptographically signed digital credentials, chained to form a verifiable audit trail for both human-present and human-not-present transactions. The Google Cloud announcement used the terms Intent Mandate and Cart Mandate for the same arc and describes mandates as signed digital contracts that serve as verifiable proof of a user's instructions.

**Card schemes.** Visa's Intelligent Commerce release says consumers can set spending limits and conditions, that credentials are tokenised, and that agent-linked credentials confirm that a chosen agent may act on the consumer's behalf. Mastercard's Agent Pay release says trusted agents will have to be registered and verified and that consumers keep control over what an agent may buy for them. Neither release is a technical specification of the mandate format; read the scheme documentation for current details.

## What does a mandate contain?

| Field | Purpose | Example (synthetic) |
|---|---|---|
| Principal | Who delegated | Treasury department of a company |
| Agent identity | Which agent may act | Agent registered under a named credential |
| Permitted actions | What it may do | Prepare and submit supplier payments |
| Per-item limit | Cap on a single action | CHF 5,000 |
| Cumulative limit | Cap over a period | CHF 20,000 per day |
| Counterparties | Who it may pay | Suppliers on an approved list |
| Validity | Start, end, revocation | 90 days; revocable at any time |
| Approval rules | When a person must approve | Above CHF 5,000; any new payee |
| Version | Which rules applied | v3, effective on a recorded date |

## How is a mandate used at decision time?

When an instruction arrives, the system looks up the mandate in force for that agent and principal, evaluates the instruction against it, and records the result with the mandate version. Anything outside the mandate is rejected or routed to a person. Because the decision cites a version, a later reviewer can tell whether the limits were respected and which rules applied on the day.

This is the point of the structure. A statement such as "the agent was told to stay under CHF 5,000" is a prompt. A mandate that the payment path evaluates is a control.

## What does a mandate not do?

- **It does not replace authentication.** PSD2 requires strong customer authentication where a payer initiates an electronic payment, and for remote electronic payments authentication elements that dynamically link the transaction to a specific amount and payee. A mandate records delegation; it does not by itself satisfy those requirements, and how they apply to agents is a question for the institution and its supervisor.
- **It does not settle liability.** A mandate does not displace the directive's rules on unauthorised transactions, and whether an agent's payment counts as one the payer consented to is a question for counsel. A signed mandate is evidence in a dispute, not a ruling.
- **It does not make the agent trustworthy.** A mandate bounds what the agent can do; it does not ensure what it proposes is sensible. Limits on the downside are the point.
- **It does not have automatic legal effect.** Whether a technical mandate is binding depends on the contract and the law that govern the arrangement.

## What is a worked example?

The numbers are synthetic. The example illustrates the control pattern; it is not a description of any one product's features. A company grants an agent mandate version 3: CHF 5,000 per payment, CHF 20,000 per day, approved suppliers only, approval needed above the per-payment limit.

At 09:10 the agent submits CHF 4,300 to an approved supplier. The check cites mandate v3, finds the amount within limits, and the payment executes. At 09:40 it submits CHF 5,600. This breaches the per-payment limit, so it is held for the named approver, who approves at 10:05. At 11:20 the company's treasurer, after a supplier-fraud warning, revokes the mandate. At 11:25 the agent submits CHF 1,200; the check finds no valid mandate and rejects it.

The log shows each decision, the mandate version, the approver and the revocation time. A reviewer can establish that nothing executed after revocation.

## What belongs outside the model?

- **Evaluation.** Deterministic code checks every instruction against the mandate in force.
- **Storage and versioning.** Mandates are stored with their versions and changes are approved.
- **Revocation.** Revoking takes effect immediately and is itself recorded.
- **Identity binding.** The agent's credential is tied to the mandate; a different agent cannot use it.
- **Signing and verification.** Where mandates are cryptographically signed, verification happens in the institution's path, not in the model.
- **Audit trail.** Each decision is linked to the instruction and any ledger posting.

## How does Finray relate to agent mandates?

**Today.** Finray's platform pages do not describe a mandate object for AI agents. They do describe related foundations: [Corebanq](/platforms/corebanq/) role-based access control in which every action carries the actor and the policy version, and versioned configuration under change control; [Ordinis](/platforms/ordinis/) approval routing as state machines, with policy changes carrying a written approval record.

**Not yet.** Mandate objects for agents, agent credentials and revocation are not described as Finray features, and Finray does not claim agent integrations through the Model Context Protocol. They are design topics an institution should put to any vendor.

Related reading: [agentic payments](/intelligence/glossary/agentic-payments/), [agentic commerce](/intelligence/glossary/agentic-commerce/), [human-in-the-loop approval](/intelligence/glossary/human-in-the-loop-approval/), [Model Context Protocol](/intelligence/glossary/model-context-protocol/) and the note [Agentic payments in regulated institutions](/notes/agentic-payments-control-layer/).

## Frequently asked questions

### What is an agent mandate?

An agent mandate is a recorded grant of authority to an AI agent, stating what it may do, within which limits, with which counterparties, for how long and on whose behalf. An institution can check an agent's request against it.

### How is an agent mandate different from a power of attorney or a standing order?

A standing order repeats a fixed instruction. A power of attorney is a legal instrument with its own formal requirements. An agent mandate, as used in payment protocols, is a technical record of delegated limits that systems can evaluate automatically. Whether it has legal effect depends on the jurisdiction and the arrangement.

### What does AP2 mean by a mandate?

The AP2 project site describes checkout and payment mandates, each with an open stage that captures constraints and a closed stage that captures authorisation for a specific finalised transaction. The Google Cloud announcement describes mandates as cryptographically signed digital contracts that serve as verifiable proof of a user's instructions; the project site now calls the underlying objects tamper-evident.

### Does a signed mandate mean a payment is authorised?

A signed mandate is evidence of what the principal delegated. The institution still decides whether to execute, applying its own authentication, screening and limit checks. In the EU, PSD2 has its own rules on authentication and liability that a mandate does not replace.

### What should a mandate contain?

At least: the principal, the agent identity, the permitted actions, amount limits per item and cumulative, permitted counterparties or categories, the validity period, the approval rules, and how it is revoked. It should be versioned so each decision can cite the version in force.

## 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)
- [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)
