Purpose
Bring in digital agents as formal operating units that carry out tasks, orchestrate processes, and support decisions within governed limits. They expand the company’s capacity without increasing human intervention or cost in the same proportion.
Includes
Execution agents, decision agents, Human-in-the-loop and Human-on-the-loop, traceability, and autonomy limits.
What It Enables
- Autonomous task execution
- Intelligent process orchestration
- AI-assisted decisions
- Growth without increasing headcount in the same proportion
What Changes in Operations
- part of execution is delegated to agents;
- routine decisions can be resolved automatically;
- agents can coordinate processes;
- the human-system relationship evolves toward human-AI collaboration.
Deliverables
- Agent catalog
- Responsibilities and limits for each agent, with safeguards
- Traceability of their actions
- Human oversight policies
Indicators
- % of tasks executed by agents
- Reduction in human intervention
- Operational savings from less manual work
- Ability to grow without increasing headcount in the same proportion
Risks If Poorly Managed
- Automation without control or an owner
- Opaque decisions without traceability
- Dependence on unsupervised models
- Operational or regulatory risks
- Poorly designed automation with no economic impact
Signs of Maturity
- Agents have clearly defined functions.
- Human oversight is defined.
- Automated decisions are traceable.
- The company knows when to delegate and when to supervise.
- Growth does not require increasing headcount in the same proportion.
Typical Quick Wins
- Agents for operational reports and personalized action plans
- Agents for internal support or administrative assistance
- Agents for analyzing operational information
- Agents for repetitive administrative tasks
Relationships
Agents operate on processes that have already been simplified (Lean) and automated, and they need Architecture and Governance to operate safely and under control.
In Practice
- It is the newest and least understood dimension. An agent is not a tool someone uses: it is a digital worker that executes, decides within limits, and produces consequences.
- The risk is real: Gartner (2025), with a survey of more than 3,400 organizations, projected that more than 40% of agent projects will be canceled before the end of 2027 due to rising costs, unclear value, or insufficient controls [1].
- What an agent needs: defined functions, explicit autonomy limits, a record of what it decides, and human oversight.
An AI agent without limits is not autonomy: it is a decision without an owner.
— Mauricio Peschard
Levels of Autonomy
| Level | What the Agent Does | Examples | Scheme |
|---|---|---|---|
| 1. Intelligent assistance | Supports the human; does not take actions. | Information analysis, recommendations, report preparation | Human-in-command |
| 2. Supervised execution | Executes tasks, but requires human approval. | AI-generated responses that are reviewed, assisted approvals | Human-in-the-loop |
| 3. Supervised operation | Makes operational decisions autonomously, with human oversight. | Automatic case classification, process routing, simple decisions | Human-on-the-loop |
| 4. Governed autonomy | Operates with high autonomy within defined limits, without human intervention in each decision. | Agents that coordinate processes or optimize operations; document classification, ticket assignment, routing (once the agent is trusted) | Human-out-of-the-loop (with governance) |
The Level Is Set by Trust, Not by the Task
- Each level is tied to a human role: in-command, in-the-loop, on-the-loop, out-of-the-loop.
- The same task can start at Level 3, with human oversight, and move to Level 4 once the agent has proven reliable or when the task is very straightforward.
- Level 4 applies when four conditions are met: low risk, a repetitive or straightforward decision, automatic controls, and earned trust in the agent.
- “With governance” means that nobody steps in on each decision, but the limits, logging, and auditing of dimension 6 (Governance) remain active.
The Schemes, Explained
Human-in-the-loop (HITL). The human is a mandatory part of the decision: AI proposes and does not execute without approval. Flow: AI → recommendation → the human reviews and decides → execution. Examples: AI-suggested credit decisions, customer responses reviewed before sending, assisted medical diagnosis. The human always takes part before the action.
Human-on-the-loop (HOTL). The agent operates on its own and a human supervises the system, stepping in only if problems are detected. Flow: AI → decision → execution, with the human monitoring. Examples: fraud detection that blocks transactions, supervised algorithmic trading, support agents that respond on their own but can be stopped. The human supervises; they do not take part in each decision.
Human-out-of-the-loop (with governance), Level 4. AI decides without human intervention. Used only when risk is low, the decision is repetitive, and there are automatic controls. Examples: document classification, automatic ticket assignment, process routing.
References
- Gartner. (2025, June 25). Predicts: Over 40% of agentic AI projects will be canceled by end of 2027 [Press release].