When to use AI agents, and when a workflow is enough

"Agent" has become one of the most used words in enterprise AI. In practice, many problems described as agentic are better served by a well designed workflow with one or two AI steps inside it. Choosing the right level of autonomy is one of the most important architecture decisions in an AI project.

Three levels of autonomy

It helps to think in three levels.

Deterministic workflows. Every step and branch is defined in advance. AI may not be involved at all. These are predictable, cheap, and easy to audit.

AI assisted workflows. The sequence is still fixed, but individual steps use AI: extracting fields from a document, classifying a request, drafting a reply. The model makes local decisions inside a structure you control.

Agents. The model decides which steps to take, which tools to call, and when it is done. This handles open ended tasks, at the cost of predictability.

A simple test

Ask two questions about the task.

1. Can you write down the steps in advance? If the path is known and varies only in a few predictable ways, use a workflow.
2. Does the task require choosing between many tools based on what is discovered along the way? If yes, an agent may be justified.

Most back office processes, such as invoice handling, onboarding, and report generation, pass the first test. Research, troubleshooting, and multi system investigation are more likely to need agentic behavior.

Why the distinction matters

Agents introduce costs that are easy to underestimate:

• Variable latency and cost, because the number of model calls depends on the path taken.
• Harder testing, because the same input can produce different sequences of actions.
• Larger blast radius, because an agent with write access can change real systems.

None of these are reasons to avoid agents. They are reasons to use them where their flexibility pays for itself.

Design patterns that make agents dependable

When an agent is the right choice, a few patterns make it production worthy.

• Explicit graphs. Frameworks such as LangGraph model an agent as a graph of nodes and edges with typed state. You can see where the agent can go, add checkpoints, and resume after interruptions.
• Narrow tools. Give each tool a single clear purpose with validated inputs. A tool that can "run any SQL" is a risk. A tool that "returns open orders for a customer" is not.
• Human approval for consequential actions. Sending external communication, changing financial records, or making decisions about people should pause for a person to approve.
• Budgets and stop conditions. Limit steps, time, and spend per task, and define what the agent should do when it reaches a limit.
• Tracing. Record every model call, tool call, and state transition so failures can be diagnosed.

Human oversight is a design feature

In recruitment, for example, an AI system can parse CVs, compare them to requirements, and draft an evaluation. It should not reject a candidate on its own. Building the review step into the graph, with the evidence a recruiter needs to make the call, produces a better system and a more defensible one.

The practical path

Start with the simplest structure that solves the problem. Add AI steps where variation defeats rules. Introduce agentic loops only for the parts of the task that genuinely need them, with approvals around anything consequential. This keeps systems understandable while still capturing what modern models can do.

Plenware designs agentic AI systems and intelligent automation with this progression in mind.