Workflow or AI agent? Choosing the simplest reliable approach

Start with the work, not the label.

Before deciding whether a process needs an agent, write down what happens from the first input to the final outcome. Who makes each decision? What information do they use? Which steps always happen, and which change from case to case?

This often reveals several different problems inside one workflow. A document may need interpretation, an approval may follow an explicit rule, and an update may simply need to reach the right system. The same technology does not need to handle all three.

The useful question is which approach can do each part reliably, at an appropriate cost, with a clear path when something goes wrong.

Use rules for decisions you can state.

If a decision can be expressed clearly, conventional software is a strong starting point. Required fields, permissions, calculations, deadline checks, and straightforward routing usually do not benefit from asking a model to decide.

For example, an invoice with no supplier identifier can be sent for review by an explicit rule. There is no need for an agent to deliberate about whether the identifier is missing. The rule can be tested, explained, and changed when the business requirement changes.

Use a workflow for a known sequence.

A workflow coordinates the steps: receive the request, validate the information, ask for approval, record the decision, and notify the next owner. It can include branches and exceptions without letting a model invent the process.

An AI step can sit inside that workflow. A model might extract candidate information from a document, while ordinary code checks the result and a person reviews uncertain fields. The overall sequence stays explicit.

Our illustrative document-intake example follows this pattern. Interpretation is useful, but it does not need to control every action.

Use an agent when choosing the path adds value.

An agent becomes useful when the next step depends on context and the system needs to choose among approved actions. A request might require finding a document, checking another source, and asking a clarifying question before it can make a recommendation.

That flexibility brings more decisions to evaluate. Define which sources the agent may use, which tools it may call, what it may change, and when it must stop or ask a person. A useful agent has a bounded job; access to every tool is not a goal.

• What useful choice does the agent make that a fixed workflow cannot handle well?
• How will the result and the actions be checked?
• Which actions require approval before they happen?
• Who can resolve a failure or an uncertain case?

Test the smallest meaningful version.

Take representative cases, including awkward inputs and exceptions. Compare a rules-first approach with the AI-assisted version where interpretation is needed. Evaluate handling time, correctness, cost, and the burden left for a reviewer.

Do not judge only by whether the happy-path demonstration looks impressive. A system that saves a few steps but creates a large verification task may not improve the operation.

The result can be a combination: software for calculations and permissions, a workflow for coordination, a model for interpretation, and people for consequential judgment. Intelligence belongs where it helps; predictability belongs where the work depends on it.