Human-in-the-Loop Workflows

Wherever AI output leads to a consequential action, and particularly where that action is irreversible or externally visible.

The design question is not whether to include a person but where. Proportionate review works: low-stakes work runs unattended, medium-stakes work gets lightweight approval, and irreversible or externally visible actions require structured review with documented rationale.

Requiring approval of everything creates a bottleneck teams route around, and a bypassed control provides worse assurance than a lighter control that survives contact with real workloads.

What the work includes

• Risk tiering with criteria covering reversibility, blast radius, and detectability
• Checkpoints placed before irreversible steps rather than at the start of a run
• Review interfaces that show reasoning, so the checkpoint is meaningful
• Measurement of how often review changes the outcome, to tune the tiers