LLMs, GenAI, RAG, agents, evaluation · insight-1563
Reserve human approval for the few high-impact actions, because constant prompts get rubber-stamped
- Evidence type
- practitioner heuristic
- Difficulty
- medium
- Roles
- Applied AI, ML engineer
- Industries
- saas, banking
Summary: situation, action, reason, boundary
Situation
A team makes an agent ask for approval on every tool call to be safe, and reviewers click approve out of habit.
Action
Classify actions by impact, let read-only and easily reversible ones run inside a restricted sandbox, and ask a person only for irreversible, costly or external-facing actions, showing the exact change.
Reason
Anthropic reports that frequent approval requests slow work and make people stop paying attention, and that sandboxing cut its permission prompts by 84 percent in internal use, a vendor-reported figure.
Boundary
Agents with many tool calls per task and human reviewers.
Symptom or risk
Approval fatigue means the rare dangerous action is approved as casually as the harmless ones, so the control gives false comfort.
Exceptions
In a new high-risk domain, start strict and relax with evidence.
Trade-offs
Tiering needs an agreed risk classification and a trustworthy sandbox.
Pitfalls
Letting low-risk actions chain into a high-risk outcome with no check.
Lessons and concepts
Claims and their sources
Anthropic states that frequent approval requests cause approval fatigue, and reports that sandboxing reduced permission prompts by 84 percent in its internal usage (vendor-reported).