Start with the smallest change that fixes the problem: removing a step, using fixed rules or adding one AI step. An agent fits when the next step depends on what’s found along the way.
Isn’t an agent the more advanced choice?
More autonomy is a design choice, not a goal. It brings more to test and more to look after, so it should earn its place.
1. What actually needs to improve?
Is the repeated work necessary? A clearer form, fewer approvals or one shared record might remove it altogether.
Practical check: Name the burden before choosing the technology.
2. When the route is known, use rules.
Example. Send an internal reminder two days before a recorded due date.
Define which date counts, who receives it, and when to stop or cancel.
Check for missing dates, changed dates and duplicate reminders.
Practical check: A fixed workflow may be enough.
3. Where interpretation helps, add one AI step.
Example. Pull a requested date out of an email, and ask a person when it’s unclear.
The rest of the process can stay fixed. Only the step that needs to read language uses AI.
4. When an agent fits.
Example. Investigate an exception using different permitted sources, where the next useful step depends on what is found.
Keep the required approvals and limits explicit, whatever chooses the steps.
5. Compare the options on the same cases.
Run the current process and the proposed one on the same representative cases. Compare accepted results, exceptions, total effort and running cost. Keep the simpler option unless the added flexibility earns its place.
Practical check: Predictable doesn’t automatically mean correct.
Read more: AI Assistant, Workflow Automation or AI Agent: What’s the Difference? · Where to Start with AI: Pick a Problem, Not a Tool
Examples are illustrative. This is general guidance, not advice for your situation.