Execution controls

Approval Fatigue & Human Oversight

Does the approval step help a person make a real decision?

Reference notes by Min · Reviewed October 15, 2026

Working governance definitionAlso: Meaningful human oversight · Human review quality · Approval UX2 primary references

Definition

Approval fatigue describes a review condition in which repeated, low-information requests encourage people to approve without sufficient attention. In agent workflows, meaningful oversight requires relevant information, decision authority, adequate capacity and an effective way to reject or change the action.

A review requirement can fail even when its software works exactly as designed. If the reviewer receives dozens of indistinguishable prompts, cannot inspect the destination or has no time to understand the consequences, the control may add delay without providing reliable oversight.

What it covers

Decision content
Show the action, destination, affected records and consequences.
Reviewer authority
Route to someone empowered to approve this type of action.
Workload
Consider queue volume, interruptions and escalation coverage.
Effective control
Make rejection, correction and cancellation affect execution.

Design the review around the decision, not the agent’s entire conversation. Show material facts from trusted records and let the reviewer inspect supporting evidence. Keep the proposal distinct from the model’s explanation of why it seems reasonable; a persuasive explanation can still describe a mistaken action.

In practice

Practical design patterns by Min.

  • Use a concise action preview with a clear difference from the current state.
  • Explain why review is required and what remains uncertain.
  • Separate routine permitted work from exceptions that need judgment.
  • Review mistaken approvals and reviewer feedback alongside speed metrics.

Define which action classes need review, who handles them and what happens when the queue exceeds capacity. Reducing unnecessary interruptions may improve attention, but do not remove a consequential checkpoint solely to improve completion rates. Use sampled review exercises and incident analysis to identify confusing proposals.

Research notes

Source-grounded details and editorial applications. Follow each link for the original document.

Interaction design is part of oversight

Amershi and colleagues propose and evaluate guidelines for human–AI interaction. The research is relevant to making system behavior understandable and correctable; it does not validate a particular agent approval threshold.

Source: Amershi et al.: Guidelines for Human-AI Interaction (CHI 2019) ↗

Show what is being authorized

OWASP’s transaction guidance emphasizes presenting significant transaction information and protecting it from modification. Applying that principle to recipient lists and attachments helps make an agent’s proposed communication reviewable.

Source: OWASP: Transaction Authorization Cheat Sheet ↗

Working example

Reviewing an agent’s customer email

An agent proposes sending a response that includes an attachment and a new external recipient.

  1. Show the final recipient list, attachment and message rather than a generic confirmation.
  2. Highlight the added recipient and allow the reviewer to remove it.
  3. Bind authorization to the reviewed version and request a new decision if significant content changes.
Illustrative scenario, not a reported deployment.

Why the controls matter

In the email scenario, the relevant change is the new recipient’s access to the attachment. A summary saying “send the customer an update” hides that consequence. After the person edits the proposal, the system must execute the edited version rather than allow the agent to reconstruct different content later.

Common pitfalls

  • Treating an approval click as evidence of informed judgment.
  • Measuring review quality only by how quickly the queue clears.
  • Automatically executing a sensitive proposal when a reviewer does not respond.

Very rapid approvals can be a signal to investigate, but they do not by themselves prove negligence. Some requests are straightforward. Examine the task mix, information shown and error patterns before changing reviewer expectations or introducing additional friction.

Questions for operational review

Use these questions to decide what evidence your team needs. They are editorial prompts, not a compliance checklist.

  • Can the reviewer see the consequential change without reading a full transcript?
  • Does rejection reliably stop the action?
  • What evidence shows review catches mistakes under realistic workload?

References

Primary documents reviewed for this entry. Examples and implementation patterns are editorial synthesis.

  1. Amershi et al.: Guidelines for Human-AI Interaction (CHI 2019) ↗Accessed October 15, 2026
  2. OWASP: Transaction Authorization Cheat Sheet ↗Accessed October 15, 2026