The four-eyes principle is a control requiring that a defined action be handled by two people: one prepares or proposes it, and a second, independent person checks and approves it before it takes effect. It is common for payments, journal entries, system changes and credit decisions. Its purpose is to catch errors and make fraud require collusion.
Why does the four-eyes principle matter once agents act?
The obvious question is whether an agent counts as a pair of eyes. It can count as the first: an agent that prepares a payment and a person who approves it meets the form of four-eyes. It should not count as the second for any action where the control matters. An agent checking another agent, or checking a person's work and signing it off, has no independent judgement and no accountability, so the principle's purpose is lost.
There is a subtler risk. When an agent prepares well, the human approver can drift into approving by reflex. Four-eyes then becomes two eyes and a click. Keeping the second check real, rather than a formality, is the hard part, and it is part of agent governance is a bound, not a model.
How does four-eyes work in practice with agents?
| Arrangement | Meets four-eyes? | Why |
|---|---|---|
| Agent prepares, person approves | Yes, if the person reviews | The approver is independent and accountable |
| Person prepares, agent checks, person approves | Yes, and stronger | The agent adds a check; a person still decides |
| Person prepares, agent approves | No | The approver has no accountability |
| Agent prepares, another agent approves | No | Neither party can answer for the decision |
An agent is valuable as an extra check alongside the two people: it can test every item for anomalies before the human approver sees it, so the approver's attention goes where it is needed. Running that pattern is covered in how to run four-eyes checks on operational changes with an AI agent.
What does that look like in an example?
A fund administrator requires four-eyes on every change to investor bank details. Before agents, an operator entered the change and a supervisor approved it. Now an agent reads the investor's signed request, checks it against the register, compares the new details with known fraud patterns and drafts the change. A supervisor reviews the agent's checks and the original request, then approves. When the agent flags that the request's signature differs from the one on file, the supervisor rejects the change and the team contacts the investor through the registered address.
How is four-eyes different from segregation of duties?
Four-eyes applies two people to one action. Segregation of duties splits a whole process so that different stages sit with different people: one creates suppliers, another approves payments, a third reconciles the bank. Four-eyes can be one way to meet segregation at a single step, but a process can have four-eyes on every step and still fail segregation if the same two people do everything. With agents, the question for four-eyes is who approves each item; the question for segregation is whether any one party, person or agent, holds incompatible stages.
Does four-eyes help with accuracy duties?
It can. Under data protection law, personal data must be accurate, and an agent's output that updates a customer's record is personal data. A second, human check on changes to customer data is one way to meet that duty, particularly for changes the customer will rely on. The duty itself is discussed in GDPR accuracy principle and AI agent outputs.
Questions people ask
Can an AI agent be the second pair of eyes for low-risk actions?
For actions where the business does not need a person accountable, it can serve as an automated check. But that is no longer four-eyes in the control sense; it is an automated control and should be described as one. Where policy or regulation calls for four-eyes, the second party should be a person.
How do you stop four-eyes becoming a rubber stamp?
Show the approver the evidence, not just the conclusion. Vary what they see, sample their approvals, and track how often they edit or reject. An approver who never changes anything across hundreds of items may not be reviewing them.