Agentic workflow
A process where the agent picks the steps, not just executes them. The dividing line that changes how the workflow must be governed.
An agentic workflow is a process where one or more AI agents handle the steps — including the planning of which steps to take. The agent decides the sequence, picks the tools, and adapts when something goes wrong. The human sets the goal and reviews the outcome.
Contrast with an automated workflow, where a person or a BPM platform has pre-defined the steps and the system executes them in order. RPA is automated. An AI agent that triages a support ticket, gathers context across three systems, and either resolves or escalates with an annotated summary is agentic.
Why it matters now
The distinction matters because the two require fundamentally different governance. Automated workflows are validated once and trusted to repeat. Agentic workflows must be observed continuously, because the agent will sometimes invent steps you didn’t anticipate — useful ones and broken ones alike.
You can’t write a single SOP for an agentic workflow. You write a goal, a set of allowed tools, and the boundary conditions for “stop and ask a human.” Then you watch what the agent actually does, and you keep tightening the boundary as failure modes surface.
The misuse
“We agentified the workflow” almost always means “we wrapped an LLM around an existing automated workflow and called the LLM an agent.” If the steps and the tool order are pre-defined in code, the LLM is doing classification or text generation — not agency. The system is automated with AI components, not agentic.
The diagnostic question: if the next step were removed from the prompt, would the system still complete the task by figuring out what’s missing? If yes, agentic. If no, you’ve built a slightly fancier RPA bot, and the governance posture should reflect that.