Human-in-the-loop is the standard answer to AI risk, and it is the right answer. It is also frequently implemented in a way that provides the appearance of control rather than control itself.
- ›The two failure modes of approval workflows
- ›Designing reviews people actually perform
- ›What must always be reviewed, and what should not be
- ›Knowing when to stop reviewing something
Failure mode one: rubber-stamping
When every AI output requires approval, reviewers develop a click reflex. After the fortieth identical proposal, nobody is reading. The queue still exists, the audit log still shows human approval, and no human judgment occurred.
This is worse than no review, because it manufactures false assurance and distributes accountability to someone who did not actually evaluate anything.
Failure mode two: the bottleneck
The opposite: so much requires approval that the queue grows faster than anyone drains it. Work stalls waiting on a manager who is in meetings, and the team routes around the system entirely.
Both failures have the same root cause: uniform review policy applied to non-uniform risk.
Design by consequence
Sort AI actions into three tiers and treat them differently.
Always review, individually. Irreversible actions, anything customer-visible, anything financial, anything legally significant. Small volume, real attention.
Review in batches. Reversible internal changes with moderate consequence. Present them grouped, at a set time, with the ability to accept many at once after scanning. Batching preserves attention better than a stream of interruptions.
Do not review; monitor. High-volume, low-stakes, fully reversible actions. Do not queue these. Let them execute, and review the pattern weekly rather than the instances. Undo exists for the exceptions.
Important: If an action is fully reversible and low stakes, individual approval is theater. Move it to monitoring and spend the reclaimed attention on the tier that matters.
Make review cheap
A reviewable proposal shows what would change (before and after), why the AI concluded it, and what happens if it is wrong. If a reviewer has to open three tabs to evaluate a proposal, they will stop evaluating.
Know when to graduate an action
The clearest signal: you have approved the same verb from the same agent, unchanged, for several weeks. That is not oversight; it is a queue. Delegate that specific verb with an explicit permission and keep the audit trail. See AI autonomy levels explained.
The reverse signal matters equally: when you start correcting proposals frequently, or when a mistake reaches a customer, move that action back up a tier. A system that only ratchets toward more autonomy has no feedback loop.
The reviewer capacity question
In an AI-native company, review capacity becomes the real constraint. Plan for it: name who reviews what, batch it into defined windows, and prune AI work whose output nobody consumes. See what makes a company AI-native.
FAQ
Who should approve AI actions?
Whoever would have done the work, or their manager. Approval by someone without domain context is the fastest route to rubber-stamping.
Does an audit trail replace review?
No, but it changes what review is for. Audit plus undo lets you review patterns after the fact instead of instances before, which is the right trade for low-stakes actions.
How do we prevent approval fatigue?
Fewer things in the queue. Every item you move to monitoring buys attention for the items that need it.
Brainis routes AI actions by risk class, batches approvals, and keeps a tamper-evident audit chain with undo. See how autonomy works.
Sharing insights on business operations, AI, and modern team management.
Run your company on Brainis
All 11 Operating Systems on every plan, from $29 a month. No per-seat pricing — you pay for AI capacity, not headcount.
See pricing