Start with the work, not the model
An agent is useful when it can own a bounded workflow: inspect context, choose a permitted action, use tools, check the result, and escalate when the situation falls outside its rules. It is not automatically the right answer because a task contains text or requires judgment.
Before building, describe the current process in plain language. Identify the trigger, inputs, decisions, systems touched, expected result, reviewer, and failure path. That description is more valuable than a feature list.
The Rule of R
A strong agent candidate usually satisfies all three conditions. If one is missing, a chatbot, a small automation, or a documented manual process may be the better investment.
- Repetitive: the work occurs often enough that improvement compounds.
- Rule-based: inputs, policies, outputs, and exceptions can be described consistently.
- Return: the saved time, reduced delay, or improved control is worth more than the build and operating cost.
Choose the smallest system that works
If a person only needs help finding information, use search or a grounded chatbot. If the same fields always move between two systems, use deterministic automation. Add an agent when the workflow requires interpreting variable context, choosing between permitted tools, and explaining the decision.
Prove one workflow with real but controlled data. Add evaluation examples, an approval gate, and an audit trail before expanding scope.