A workflow does not need an agent just because it happens more than once. Automate the steps you understand well enough to catch when the automation is wrong.
Automate repetition that stays understandable
Good candidates tend to repeat, use reasonably stable inputs and have a clear outcome. Renaming downloaded files by date is predictable. Summarising each support request may contain ambiguity. The second job could still be automated, but it needs a way to show sources, flag uncertain cases and let a human correct the result.
Start by writing the workflow down as a sequence of steps. If you cannot explain where an exception goes, automation may make the confusion happen faster. Sometimes the best first automation is not a model; it is a script, shortcut, template or form that removes copy-and-paste.
Use a small scorecard
Give each factor a low, medium or high value: how often does the task occur, how consistent is the input, how much time does it take, what does an error cost, and how much context must a human apply? High frequency and stable inputs are encouraging. High error cost or deep tacit judgement are reasons for review and slower rollout.
Calculate the total time saved including the time spent checking, fixing exceptions, monitoring the automation and updating it. A tool that saves three minutes per run but needs daily repair is not actually saving three minutes. Make the comparison over a realistic month, not an ideal one-click demo.
| Factor | Automation fits better when… | Keep human control when… |
|---|---|---|
| Input | Formats are stable and available | Important context is missing |
| Decision | Rules or outcomes are easy to define | The right answer depends on judgement |
| Error cost | Mistakes are reversible and visible | Mistakes affect safety, money or rights |
| Frequency | The task repeats predictably | It is an occasional exception |
| Monitoring | A failure can be caught quickly | Failures can stay invisible |
Introduce automation with a brake
Run the automation in shadow mode first. Let it suggest what it would do while a human still completes the task. Compare suggestions with real decisions and label disagreement. This reveals edge cases before the script is given permission to act.
Next allow it to take reversible steps with a clear undo. Keep approvals before sending a message, changing a record, granting access or spending money. Set a stop condition for failures and log the inputs, decision and action. Logs should help you diagnose problems without retaining sensitive details you do not need.
Only expand permissions when error patterns are understood and the person responsible has a monitoring plan. New input sources or model versions can shift behaviour, so keep regression cases and a rollback path.
Know when AI belongs in the workflow
Deterministic automation works when the input and action can be described with reliable rules. AI is useful when the task needs interpretation across varied text, images or speech. For example, a fixed label-to-folder rule needs no language model. Grouping unusual support requests into emerging themes might benefit from one.
Even then, keep the model away from final actions it cannot justify. Ask it for candidates, require evidence from the input, then validate the output. The more consequential the action, the stronger the confirmation and audit trail should be.
A little automation that finishes a dull task is a win. A general agent that is hard to understand and impossible to stop is a maintenance obligation in a clever costume.
FOLLOW THE EVIDENCE
Sources & further reading
Process note: AI tools assisted with research and drafting. Sources are linked so you can inspect the evidence. Unless an essay says otherwise, it does not claim hands-on testing or independent reproduction.
