Approval fatigue: why human-in-the-loop fails at scale

"Human in the loop" sounds responsible.

It reassures stakeholders. It looks good in a risk review. It gives product teams a simple answer when someone asks whether the AI can act on its own.

The trouble is that human-in-the-loop is not a design pattern. It is an intention. And intentions collapse quickly under workload.

If the human sees too many approvals, with too little context, at the wrong moments, the loop becomes theatre. The person is present, but no longer meaningfully reviewing. They click yes because yes is how the work moves on.

That is approval fatigue.

The loop is getting heavier

This problem matters more as AI shifts from answering to acting. Deloitte's 2026 State of AI report says workforce access to sanctioned AI tools grew from fewer than 40% to around 60% in one year. It also found that 85% of companies expect to customize autonomous agents to fit their business needs.

Gartner is more blunt about the risk: it predicts more than 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls.

That phrase - inadequate risk controls - is where UX enters the room.

Many teams interpret control as "add an approval button." But an approval button does not create control unless the human has enough information, authority, and attention to make a real decision.

Bad approval has three shapes

The first bad shape is approval without preview.

The AI says, "Ready to send?" The human sees a button, but not the exact email, recipient list, data used, assumptions made, or downstream effect. That is not approval. It is consent in the dark.

The second is approval without triage.

Every action gets a gate. Low-risk formatting changes, research steps, data lookups, customer-facing messages, and irreversible updates all interrupt the user equally. The system treats everything as important, so the user learns that nothing is.

The third is approval after commitment.

The AI acts, then the product offers an audit log or notification. The user technically knows what happened, but too late to steer it. This is post-hoc accountability dressed as oversight.

All three fail for the same reason: they put the human near the loop but not usefully inside it.

Oversight is a placement problem

The fix is not to approve everything. It is to place approval where it changes the outcome.

Good oversight distinguishes between steps that can run freely and steps that require human authority. Research can often run freely. Drafting can usually run freely. Ranking can run freely if the criteria are visible. But sending, deleting, publishing, rejecting, charging, notifying, or modifying a source of record usually needs a gate.

The gate should appear at the moment of consequence, not at every moment of activity.

That gate should also contain enough context for judgment:

• the plan so far
• the proposed action
• the exact preview of what will change
• the evidence or rationale
• the confidence or uncertainty
• the recovery option if approval proves wrong

This is where old product patterns still matter. Infrastructure teams have long used dry-run and plan-before-apply workflows because blind approval is reckless. The same logic belongs in AI products.

Approval quality beats approval quantity

A product with fewer approvals can be safer than a product with more approvals if those approvals are better placed.

The goal is not maximum human involvement. The goal is meaningful human control.

That means:

• Gate irreversible actions.
• Gate high-blast-radius actions.
• Gate actions that affect customers, money, rights, access, or public records.
• Let low-risk, reversible, preparatory work run.
• Show a dry-run preview before commitment.
• Make the AI's reasoning available on demand.
• Log the final action and the human decision.

This is how oversight survives real workload. It respects the user's attention as a limited operational resource.

The hidden danger: attention becomes the control

Poorly designed human-in-the-loop systems often depend on attention as the final safety layer. That is fragile.

People get tired. Work queues stack up. Senior users delegate. Teams normalize the happy path. Over time, the approval ritual becomes part of the workflow muscle memory.

If the product needs a human to detect every subtle error manually, the product has not designed control. It has outsourced the risk to attention.

Better AI oversight reduces the burden on attention by making risk visible. It highlights what changed, what matters, what is uncertain, and what cannot be undone. The human still decides, but the system does more work to make that decision real.

A practical test

Take one AI workflow and list every moment where the product asks for approval. For each one, ask four questions:

1. What happens if the human approves without reading?
2. Is the action reversible?
3. Does the user see the exact consequence before committing?
4. Would this gate still matter if the queue had 100 items in it?

If the answer to the last question is no, the approval is probably theatre.

The WFK position

Human-in-the-loop is not enough. The loop has to be designed.

The human needs the right gate, at the right moment, with the right preview, for the right level of consequence. Anything else is just a button that lets the system say a person was involved.

Oversight is not the opposite of automation. It is the architecture that lets automation become acceptable.