A workflow is the right answer more often than people think
If the process is “read this brief, create three drafts, run a checklist, send for approval,” write the steps down. A normal workflow is easier to test, audit, estimate, and repair than an agent that invents its own route through the task.
An agent earns its place when
- The next step depends on what the previous step discovered.
- Several tools or sources must be chosen dynamically.
- The task repeats but the path varies within clear boundaries.
- The system can recover from a failed tool call.
- A human can approve risky actions before they happen.
The risk test
| Question | If the answer is yes |
|---|---|
| Can the system send, buy, delete, publish, or change data? | Add approval and audit logs before autonomy |
| Can a wrong action harm a customer or budget? | Keep the action behind a human gate |
| Can the task be completed with a checklist? | Start with a workflow |
| Will the path vary in ways you cannot enumerate? | An agent may be useful, but define a strict boundary first |
The practical rollout
Start with a workflow that logs each step. Let an agent suggest the next action while a person approves it. Only remove an approval after repeated tests show that the risk is understood and the benefit is measurable.