"Humans Above the Loop" discusses three oversight levels for AI work (HITL, HOTL, HATL) emphasizing clear direction and autonomy for better results.
Clear direction, high autonomy - and reviewing results together.
What It Is
How humans oversee AI work can be described in three levels - often abbreviated as HITL, HOTL, and HATL:
- Human in the Loop (HITL): Human approves every single step. Nothing happens without sign-off.
- Human on the Loop (HOTL): AI works independently, human monitors and intervenes when needed.
- Human above the Loop (HATL): Human sets direction and goals. Intervention only on escalation or strategic decisions.
The framework is described by McKinsey in "The Agentic Organization" (September 2025). The design principle: default is agent execution, humans are selectively reintroduced.
Why Direction Works Better Than Rules
AI agents deliver better results when given clear direction and a high degree of autonomy - rather than rigid rules and deterministic step-by-step instructions. The pattern is familiar from working with human teams: people who understand context and have room to maneuver deliver better than those executing instructions.
Teams that still manually review every step create new bottlenecks in their work system and end up slower than before.
What "Above" Really Means
"Above the Loop" does not mean less control. It means shifting attention. Away from checking individual outputs, toward organizational learning.
Concretely: humans and AI agents review examples of their work together. Examining results, questioning processes, improving practices. Without this shared learning, agents optimize locally on their own, and nobody notices when quality drifts or unwanted patterns emerge.
How To Spot It
- Every AI output is manually reviewed, even when 95% are flawless - the team is stuck in HITL
- A manager spends more time approving AI results than doing strategic work
- AI agents receive such tight constraints that their results are worse than necessary
- Conversely: AI agents run completely unsupervised, with nobody reviewing results
What To Do (FL3 + FL2 - Strategy and Coordination)
- Choose your loop level deliberately: For each process, decide HITL, HOTL, or HATL. The answer depends on risk, not on habit
- Direction over rules: Agents need context and goals, not deterministic runbooks
- Build in learning: Regularly review work together - sessions where human and AI output are examined side by side
- Define escalation paths: HATL means clearly defined WHEN a human steps in
- Connect with Autonomy Gradient: The loop levels correspond to agent autonomy levels - both must be designed together
The Trap
Two extremes: controlling everything (HITL for all) and creating new bottlenecks. Or letting everything run and losing control. The art is giving direction, allowing autonomy - and consistently learning from the results together.