Calling the Manager

Opening Scene

A skilled line cook handles the overwhelming majority of a service’s decisions independently, but knows precisely which specific situations genuinely warrant calling the manager over — a serious customer complaint, an unusual substitution request, anything with consequences beyond the cook’s own authority to decide alone. That judgment about when to escalate, rather than push through independently, is a genuine mark of real skill, not a limitation. AI agents need this exact same calibrated judgment.

In Plain English

Human-in-the-loop design deliberately builds explicit checkpoints into an agentic workflow where a human reviews or approves a decision before the agent proceeds, particularly for consequential, irreversible, or genuinely uncertain actions. This isn’t a failure of autonomy — it’s a deliberate, calibrated design choice, recognizing that full autonomy is appropriate for some decisions and genuinely inappropriate for others, connecting directly to the constraint handling covered in Article 10 and the error handling covered in Article 12.

The Old Way

Before human-in-the-loop design was recognized as a deliberate, calibrated practice, early agentic systems often defaulted to one extreme or the other:

• Some early agentic systems were designed for full autonomy on every decision, without any deliberate consideration of which specific decisions genuinely warranted human review.
• Other systems required human approval for literally every single step, effectively eliminating the genuine efficiency benefits agentic autonomy was meant to provide.
• There wasn’t yet a well-established framework for deliberately calibrating exactly which decisions needed human review and which didn’t.

Recognizing human-in-the-loop design as a deliberate, calibrated spectrum — not an all-or-nothing choice — represents real, accumulated practical wisdom from early agentic deployments.

What’s Changing (and Why AI Is the Reason)

1. Well-designed agentic systems increasingly calibrate human review specifically to a decision’s consequence and reversibility — full autonomy for low-stakes, easily reversible actions, and mandatory human approval for high-stakes, irreversible ones.
2. This connects directly to this content library’s dedicated series on AI governance and responsible AI, since calibrated human oversight is a core, practical principle in deploying AI systems responsibly.
3. As agentic systems take on increasingly consequential real-world authority, deliberate, well-calibrated human-in-the-loop design has become a genuine trust and safety requirement, not an optional refinement.

The Metaphor, Fully Extended

The Kitchen	Human-in-the-Loop Concept
Handling most decisions independently, without needing approval for everything	An agent operating autonomously for low-stakes, routine decisions
Knowing precisely which situations genuinely warrant calling the manager	Knowing precisely which decisions genuinely warrant human review
A cook’s calibrated judgment about escalation as a mark of real skill	An agent’s calibrated design for escalation as a mark of real system maturity
A manager who’s brought in for genuinely consequential decisions	A human reviewer brought in for genuinely consequential agent decisions

Quick Recap

• Human-in-the-loop design builds explicit checkpoints where a human reviews consequential agent decisions before they proceed.
• This is a deliberate, calibrated design choice, not a failure of autonomy or an all-or-nothing decision.
• Well-designed systems calibrate review requirements specifically to a decision’s consequence and reversibility.
• This connects directly to broader AI governance and responsible AI practices.

Where This Fits in the Series

Article 14 covered knowing when to bring in a human. Article 15 covers a related but distinct choice: planning everything out ahead of time versus improvising step by step.