Why a Human Is the One Who Gives Final Approval
We covered earlier that the AI in the executor role only needs to handle the work itself. After that comes the auditor, a separate role from the executor, whose job is to surface any problems.
So who makes the final call to move forward or stop?
That is the last piece of the three roles in the separation of powers (here meaning: dividing execution / audit / approval between separate agents): the approver. In this structure, the approver role is held by a human.
What the Approver Does
The approver's job is to make the decision: "we can proceed."
After the executor finishes the work and the auditor completes the check, the approver is the one who says "go" and takes ownership of that decision.
There is a reason this role belongs to a human.
AI can execute. AI can audit. But the capacity to accept responsibility for a decision — to say "this is on me" — is something only a human can do right now. The approver is a human not because AI cannot be trusted without human oversight, but because the design needs a clear place where accountability sits. Assigning a human to this role makes that location visible in the structure.
Why the Final "Go" Belongs to a Human
No matter how precisely execution and audit are designed, the act of moving forward still requires someone to own that decision.
This is where irreversible actions (operations that cannot be undone once taken) become the central concern.
Publishing an article, deleting a file, sending something to an external service — these are operations that cannot simply be "undone" after the fact. Because they cannot be reversed, there must be a checkpoint before execution: a moment where someone confirms, "do we actually want to proceed?"
The human approver stands at that checkpoint.
Because the executor and auditor are AI, it is possible to create a documented fact that a human reviewed the matter before any irreversible action was taken. That fact becomes part of the record. When something needs to be explained later — "why was that decision made?" — the record is the answer.
On the other side of this: it would not be practical to require human review for every small, reversible operation. Stopping the flow for every minor task that can be easily undone would create unnecessary friction in the overall process.
So human review is reserved for what is irreversible and significant. Reversible, low-impact work is left to the AI side. That distinction is what allows the approver's attention to stay focused on the situations where it genuinely matters. This is a key part of how the structure works in practice.
What the Approver Does — and Does Not Do
The approver's core action is simply: decide whether to proceed or stop.
The approver does not dive into implementation details. Introducing a design change at the approval stage — "maybe we should have structured it this way instead" — breaks the process. Approval is the act of deciding whether to proceed given everything that has happened up to that point. If there is a need to revisit the design, that conversation needs to happen before approval, not during it.
There is another principle we keep in mind. The approver should not work alone.
When a decision is irreversible and carries real weight, a single perspective is more likely to miss something. Having a second person review the same material often surfaces problems that were not visible before. The principle of "do not decide alone" is especially effective for high-stakes decisions. We will go deeper on this in another chapter, but in this structure, irreversible actions include a two-person rule (a second reviewer is required before the action proceeds).
On top of that, the reasoning behind each approval decision is recorded. Writing down "why we gave the go" is what connects approval to the document-first principle (the practice of leaving the basis for decisions as written text). Approval is not a feeling or a moment in the flow — it needs to exist as something that can be explained and referred back to. When a similar decision comes up later, having a record of the previous one means the past can inform the present.
The Three Roles Work Together
The executor (AI) completes the work. The auditor (AI) surfaces any problems. The approver (human) decides whether to move forward. Because these three roles belong to separate parties, each one can focus entirely on its own job.
The goal of this structure is a state where it is always clear: when someone does something, who was watching, and who is accountable.
When all three roles sit in the same place, execution, audit, and approval all happen in the same space. If something goes wrong, it is hard to trace where the problem originated. Separating the three makes it visible which layer a problem came from.
In practice, the benefit of this structure is not most felt when something goes wrong. It shows up quietly during ordinary days when nothing is happening. Each task accumulates as a record. Who decided what stays in writing. After two weeks, after a month, trust in the structure builds gradually.
That is the full picture of the three roles in the separation of powers.