Start with repeated friction.
A team member copying the same details every morning is an obvious candidate. A less visible candidate is the manager who spends an hour reconstructing status from messages. Both consume attention, but they may need different solutions.
Ask people which tasks they repeat, which information they re-enter, and which exceptions bring the work to a stop. Observe examples rather than relying only on estimates. The aim is a short list of problems with a clear frequency and consequence.
Remove unnecessary steps first.
Before connecting tools, ask why the step exists. An approval might be a useful control, an outdated habit, or a workaround for unreliable information. Automating it without understanding the purpose can make the wrong process harder to change.
Try simplifying the rule, clarifying ownership, or eliminating duplicate collection. Automation is most useful when it carries a better process, not when it preserves avoidable work at higher speed.
Evaluate more than the minutes saved.
Consider frequency, variability, consequence, data quality, and recovery. A frequent task with clear rules and reliable inputs is different from a rare, high-consequence decision with incomplete information.
- Can the rule be explained in plain English?
- Are the inputs available and reliable?
- Who handles exceptions and failed actions?
- Will the next team have capacity for the faster output?
- How will the business measure the result?
Choose a bounded first intervention.
A good first scope might connect a completed intake to an assigned task and a visible status. It should have a clear beginning, a definition of done, and a recovery path. Start where an improvement will be useful even if no additional automation follows.
Establish a baseline for administrative time, waiting, or rework. After rollout, compare like-for-like work and ask the team what changed. Do not count time as saved if it simply moved into checking, correcting, or maintaining the automation.
AI is not required for every automation. Predictable rules usually deserve a predictable implementation. Use AI when handling unstructured information creates enough value to justify review and failure handling.