Start with an ordinary working day
Ask someone to walk through the task while performing it. Note where information comes from, which systems it passes through and what triggers the next step. Look for copying, waiting and repeated checks.
Consider an illustrative example: an approved order is entered again in another system. The first trial could transfer the approved information without covering the entire order process.
Identify what is actually a rule
Write down the decisions. If someone can explain them precisely, they may become validation or an automated action. If they require professional judgement, keep that part with the person.
List exceptions too: missing information, duplicates, unexpected formats or a system that does not respond. A solution is only useful if these cases can be handled.
Compare a few candidate tasks
Consider frequency, time per task, error risk, data access and the consequences of failure. Use observations from real work rather than a rough guess.
For illustration, 20 weekly repetitions with five minutes of manual entry amount to 100 minutes. That is a starting point, not a savings promise: review, exceptions and operation still need to be counted.
Bound a trial you can evaluate
Choose one input, one clear action and a visible result. Agree how to measure quality, time spent and exceptions. Include difficult examples as well as straightforward ones.
Keep review available before allowing actions to run without supervision. Expand only when the result is understandable and stable.
Agree who owns the workflow
Identify someone who knows the process and establish who responds to errors. Describe access, dependencies and how changes to rules will be implemented.
After the demonstration, ask: can we understand and operate this when everyday work changes? Documentation and ownership make automation part of the working process.