Ir al contenido

Human-in-the-loop

De Scrum Manager BoK
Esta página es una versión traducida de la página Human-in-the-loop. La traducción está completa al 100 %.
⏱ 3 min  ·  📅 2026
The terms Human-in-the-loop (HITL) and Human-on-the-loop (HOTL) describe two models of human oversight in artificial intelligence systems that differ in the agent's degree of autonomy and in the type of intervention the person performs. Both describe the person's position in the decision cycle of an AI system.

Human-in-the-loop (HITL)

The human is in the loop: the system cannot complete a relevant action without explicit approval from a human. The AI proposes; the human decides.

  • The human validates or rejects each critical step before it is executed.
  • Greater control and lower risk; longer cycle time.

Example: an AI agent generates a pull request but does not merge it until a developer reviews and approves it.

Human-on-the-loop (HOTL)

The person is on the loop: the system operates autonomously, but the person monitors it and can intervene if they detect a problem.

  • The AI acts within predefined limits without approval at each step.
  • Greater speed and scalability; greater risk if the limits are not well defined.

Example: an agent automatically deploys to staging but alerts the team if coverage metrics drop below the threshold.

When to use each model

The choice depends on:

  • Reversibility of actions: irreversible actions (production deployments, sending communications, modifying customer data) require HITL.
  • Cost of error: the higher the cost, the closer to HITL the model should be.
  • Agent maturity: a new agent starts in HITL and can evolve toward HOTL as it gains proven reliability.

In the Reinforced DoD

The HITL/HOTL position for each type of agent action is part of the team's AI governance. The reinforced DoD for AI-enabled teams includes an explicit definition of which agent actions require prior human approval (HITL) and which are carried out under after-the-fact oversight (HOTL).

Common mistake

Starting with HOTL before understanding the agent's behavior. A new agent in a new domain should always start in HITL mode, building up a behavioral track record before supervision is relaxed. Granting autonomy to an agent without data on its reliability in the specific context is the root cause of most AI incidents in development teams.

Resources

🏦 IA aplicada al trabajo ágil [AI Applied to Agile Work, in Spanish]Skill Arena · Scrum Manager

See also

Want to advance in agility? You can search for course and exam calls, or go at your own pace by becoming a member of the Agile Club. This membership includes exclusive resources, e-learning classrooms, and access to Skill Arena: a space to practice and measure your agile skills at your own pace.