One reviewer is a different control than two
A single human-in-the-loop step catches a different class of error than dual control does. One reviewer checking an AI-drafted action can catch a wrong fact, a bad fit with policy, or an obviously poor recommendation. What a lone reviewer cannot provide is separation between the party who initiated or shaped a decision and the party who authorizes it.
That distinction matters because the two failure modes are different. A single-reviewer workflow fails when the reviewer is rushed, under-informed, or simply agrees too often with a confident-sounding draft. A missing-segregation workflow fails when the same identity — human or automated — can both propose and approve, so there is no independent check on that identity's judgment or incentives at all.
Where the two-person pattern already exists
This is not a new problem invented by AI. Segregation of duties is a standard financial and IT control: the person who initiates a transaction should not also be the one who authorizes it. Cloud providers encode the same idea for infrastructure changes. AWS's Well-Architected guidance describes the two-person rule as a safety mechanism that requires two authorized people to perform tasks considered important, specifically to add protection against accidental or malicious action by someone holding critical permissions.
An AI workflow that drafts a refund, approves a credit line change, or triggers a contract amendment is doing the same kind of consequential write-back that dual control was built for. The fact that a model produced the draft, rather than a person, does not remove the need for an independent second check — if anything, it adds a new source of error that a lone reviewer may not be positioned to catch.
Decide the shape of approval by consequence, not habit
Most teams default to a single approval gate because it is the first thing they build and it is good enough for low-stakes actions. The decision about how many independent approvers a workflow needs should instead follow from the consequence and reversibility of the action, the same logic NIST's AI Risk Management Framework playbook applies when it calls for organizations to establish policies that differentiate human roles and responsibilities for AI oversight rather than leaving oversight undefined.
A practical way to sort workflow decisions: reversible, low-value actions can run with no review or a single spot-check; consequential but recoverable actions get one accountable reviewer with full evidence; irreversible, high-value, or regulated actions get two independent approvers, where neither one is the person or system that generated the original recommendation.
- No review: fully reversible, low-value, high-confidence outputs
- Single review: consequential but recoverable actions with a clear owner
- Dual control: irreversible, high-value, or regulated write-backs
What breaks a two-person control in practice
Dual control fails quietly in a few predictable ways. The most common is a shared service identity: if the same API credential both drafts the action and executes the approval, the second signature is cosmetic. Another is approval fatigue — if the second approver sees the same decision packet as the first with no new information, they will tend to agree rather than independently verify.
The fix is structural, not motivational. Give the second approver a genuinely different view of the case — a different summary, a flag on what the first approver changed or missed, or a sampling rule that routes some fraction of single-approved cases to a second reviewer after the fact. Log both approvals, the identity behind each, and the time between them, so an audit can tell a real second check from a formality.