# What is the four-eyes principle?

The four-eyes principle is a control rule under which a second, independent person must review and approve a sensitive action before it takes effect, so that no single person, and no single software agent, can both prepare and release it.

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

---

## What is the four-eyes principle?

The four-eyes principle means that a sensitive action needs two different people: one who prepares it and one who independently approves it before it takes effect. The aim is to stop errors and fraud that a single person could introduce unnoticed. In finance it is also called dual control, maker-checker or the two-person rule.

The idea is old and simple, yet the details decide whether it works. "Independent" means the second person has no stake in the outcome and cannot be the same account as the first. "Before it takes effect" means the action is held in a pending state until approval; a review that happens after the money has moved is detective, not preventive.

## Where does four-eyes come from in control frameworks?

Four-eyes is a practical form of segregation of duties. The US Government Accountability Office's internal control standard, which builds on the COSO framework, tells management to design control activities so that incompatible duties are segregated and, where that is not practical, to design alternative controls (paras 10.21 and 10.23). It adds that segregation helps address management override but cannot absolutely prevent it, because two or more people can collude (para 10.22).

That last point is worth keeping in mind: four-eyes lowers risk; it does not eliminate it. See [segregation of duties](/intelligence/glossary/segregation-of-duties/) for the wider design question of which duties conflict.

Regulation points the same way without naming the term. DORA, Regulation (EU) 2022/2554, asks financial entities for access control that limits access to what legitimate activities require (Article 9(4)(c)) and for documented change management so that changes are recorded, tested, assessed, approved, implemented and verified in a controlled manner (Article 9(4)(e)). An approval step by a different person is a common way to evidence both. FINMA's Circular 2023/1 on operational risks and resilience, in force since 1 January 2024, puts the focus on ICT, critical data and cyber risks for banks; the release does not itself prescribe a four-eyes rule.

## What does four-eyes look like in practice?

| Area | Maker | Checker | What the checker sees |
|---|---|---|---|
| Payment release | Operator who creates the payment | Approver with release rights | Beneficiary, amount, currency, source of instruction |
| Ledger posting | Bookkeeper who drafts the entry | Reviewer who approves posting | Accounts, amounts, source document |
| Payroll run | Payroll preparer | Second approver | Totals, deductions, changes since last run |
| Policy or limit change | Administrator | Second administrator | Old value, new value, reason |
| Software change | Author of the change | Independent reviewer | The exact diff and test evidence |

The last row matters for engineering teams. Independent review of the exact version that ships is the same control applied to code; see [building bank-grade software with AI agents](/notes/bank-grade-software-with-ai-agents/).

## What is a worked example?

Illustrative example with synthetic numbers. A finance team of five uses a rule that any supplier payment of CHF 10,000 or more needs a second approver, and any change to a supplier's bank details needs one too.

An operator enters a payment of CHF 18,400 to a supplier. The system holds it as "pending approval". An approver opens it, sees that the beneficiary account differs from the last three payments, and asks the operator why. The operator explains that the supplier sent a change notice, and the approver checks that the change itself was approved earlier by a second person. Only then is the payment released.

Two controls worked here: the amount threshold forced a second look, and the separate approval of the bank-detail change stopped one person from redirecting funds by editing master data and then paying. If one person could do both, the threshold alone would not have helped.

## How is four-eyes enforced rather than just documented?

A control that lives only in a handbook depends on goodwill. A control that the system enforces rejects the forbidden path. Useful properties to look for:

- **Identity separation:** the approver's identity must differ from the maker's, and the system refuses otherwise. Checking this in the database as well as the interface closes the gap where a different client or script bypasses the screen.
- **Held state:** the action does not take effect until approval; rejecting returns it to the maker with a reason.
- **Scoped rights:** approving is a distinct permission from preparing, assigned to roles and reviewed periodically.
- **Evidence:** who prepared, who approved, when, and which version was approved, written to an append-only record such as a [hash-chained journal](/intelligence/glossary/hash-chained-journal/).
- **Governed scope:** the thresholds and the on/off policy are themselves changes that need approval, so the rule cannot be quietly weakened.

## Does four-eyes still apply when AI agents prepare the work?

Yes, and arguably more. OWASP's guidance on excessive agency names excessive autonomy as a root cause: a system that does not independently verify or approve high-impact actions. Its listed mitigations include human-in-the-loop control, where a person approves high-impact actions before they are taken, together with complete mediation, meaning authorisation is checked in the downstream system rather than trusted to the model.

In four-eyes terms: an agent may be a maker, drafting a posting or a reconciliation proposal. The checker is an accountable person with their own identity, and the system, not the model, decides whether the approval is valid. Retries of an approved action must also be safe, which is where [idempotency](/intelligence/glossary/idempotency-in-payments/) comes in.

## What can go wrong?

- **Rubber-stamping:** approvals granted in bulk without reading. Mitigate with sensible limits, showing the checker the facts that matter, and sampling approvals.
- **Collusion:** two people acting together. Four-eyes does not stop this; rotation, limits and after-the-fact review can reduce it.
- **Shared credentials:** if two people share a login, there is one pair of eyes. Individual identities are a precondition.
- **Bottlenecks:** an absent approver blocks work, which creates pressure to share access. Plan deputies instead.
- **Scope creep in either direction:** too many approvals dilute attention, too few leave gaps. Risk-based scoping, reviewed regularly, is the usual answer.

## What does Finray do today, and what is not yet there?

Corebanq Accounting applies a second-person approval step to postings and bills by default. That default is a company policy, and changing it is itself an approved change; we do not claim that every change always needs a second person. A maker cannot approve their own change, and that rule is enforced in the database. VAT returns go through independent second-person review before export, and payroll runs need second-person approval. Approvals sit alongside an append-only, hash-chained journal that can be verified, so the evidence of who did what is reviewable.

AI in Corebanq Accounting is off by default, and its proposals need a person's confirmation and carry no posting authority. We do not describe agents acting in the ledger. For how Finray builds its own software under the same discipline, see [how we build](/company/how-we-build/). Related platforms: [Corebanq Accounting](/platforms/corebanq-accounting/) and [Corebanq](/platforms/corebanq/).

## Frequently asked questions

### Is the four-eyes principle the same as segregation of duties?

They are related but not identical. Segregation of duties splits incompatible roles across people as a standing design. The four-eyes principle is a per-action rule: one person prepares, a different person approves that specific action. Four-eyes is one common way to implement segregation of duties.

### Does four-eyes mean two people look at every transaction?

No. It is commonly applied by risk, for example to payments above a limit, changes to master data, payroll runs and policy changes. Reviewing everything by hand does not scale and tends to produce rubber-stamping, so the scope is a policy decision that should itself be controlled.

### Can a software agent act as the second pair of eyes?

An AI agent can prepare, check or flag, which is useful. But the approval that carries authority should come from an accountable person with their own identity. A model that both proposes and approves its own output is a single point of failure, not a second reviewer.

### How does a system enforce four-eyes rather than just record it?

By making the rule structural: the approval is a separate state transition that the system rejects when the approver is the same identity as the maker, ideally checked at the data layer and not only in the user interface. A policy document alone is a procedure, not an enforced control.

### What are the weaknesses of four-eyes?

Collusion between two people defeats it, reviewers can approve without reading, and approval queues can become bottlenecks. Controls such as approval limits, rotation, review evidence and an append-only audit trail reduce these weaknesses but do not remove them.

## Sources

- [GAO, Standards for Internal Control in the Federal Government (Green Book), 2025 revision (GAO-25-107721), Principle 10, paras 10.21-10.23 (segregation of duties)](https://www.gao.gov/assets/gao-25-107721.pdf)
- [Regulation (EU) 2022/2554 (DORA), Article 9(4)(c) access control and 9(4)(e) change management](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32022R2554)
- [OWASP Top 10 for LLM Applications 2025, LLM06 Excessive Agency](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/)
- [FINMA news release on Circular 2023/1 Operational risks and resilience - banks](https://www.finma.ch/en/news/2022/12/20221213-mm-anh-rs-op-risks/)
