AI agent

An AI agent is a workflow component that can take actions — call APIs, query data, make decisions — without a human approving each step. What distinguishes an agent from a chatbot is the action-taking capability, not the model or the interface.

Reviewed by Transformation.ai editorial · Markdown version · Stage: Pilot

What an AI agent is

Three properties separate an agent from a model that generates text. First, action-taking: the agent can call APIs, update records, trigger the next step — not just produce a suggestion a human then acts on. Second, autonomy within defined bounds: the agent operates without a human approving every step, though what it's allowed to do is defined by the workflow it runs in. Third, a defined goal: the actions serve something specific, not general conversation.

Together, the three property test matters because it separates "AI agent" from "chatbot with tools." A chatbot that can call an API when you click a button is not an agent — the action is human-initiated. An agent that reaches into a workflow, makes the call, and only surfaces when something goes wrong is.

The autonomy spectrum

Agents are not binary. Every agent runs at some autonomy tier — a named level of decision authority. The five tiers, from agentic back to human-driven:

• L0 Suggest — the agent proposes; a human acts. "Agent" by the strict three-property test requires action-taking, so a true L0 system is a proposal system, not an agent. In practice the term is used loosely; what matters is the caps.
• L1 Approve — the agent acts; every output requires human approval before it takes effect.
• L2 Exception — the agent acts autonomously on a defined set of cases; humans review only exceptions.
• L3 Self-improve — the agent tunes itself within an evaluation ceiling the human sets.
• L4 Self-extend — the agent proposes new workflows entirely.

"Agent" in enterprise conversation usually means L1 or L2. L3 and L4 agents are still rare because the governance substrate — evaluation ceilings, decision logs, kill switches — isn't yet common. The workflow map is what a team builds before deciding which tier an agent will run at.

What an AI agent is not

• Not a chatbot. A chatbot generates text in response to a user message. An agent acts. The interface is not the distinction.
• Not a copilot. A copilot acts only when a human is in the loop at every decision. That's closer to L0. A copilot can be useful; it's not what L2+ enterprise discussions mean by "agent."
• Not fully autonomous. Every enterprise agent runs within some autonomy tier, and every tier has constraints — approval gates, exception lines, evaluation ceilings, or kill switches. "Fully autonomous agent" is not a tier this framework describes, because enterprise workflows don't operate without bounds.
• Not a model. The model is the reasoning engine. The agent is the workflow component wrapped around it — the tool calls, the state, the constraints, and the action surface.