1. Start with a repeated frustration
Choose a workflow that happens often enough to inspect. “Help with admin” is too broad. “Sort incoming customer enquiries and prepare a reply for review” has a clear beginning, an output and an owner.
2. Use five practical filters
Frequency: does it happen every week? Rules: can someone explain what a good result looks like? Access: are the inputs available in a supported tool? Review: who catches an exception? Value: is there enough recurring effort or delay to justify implementation? A workflow does not need to be perfect, but it needs a boundary.
3. Write down the exceptions
Collect ten ordinary examples and five awkward ones. Include a missing attachment, an ambiguous request, an unexpected sender, an incorrect number and a message that needs the owner. A system that only handles the easiest examples will create more supervision than it removes.
4. Choose one definition of done
A useful definition might be: every in-scope enquiry receives a category, a named owner and a draft that makes no unverified promises. It does not include sending that draft, inventing a price or resolving every customer issue. Keep the first outcome narrow enough to verify.
5. Make a baseline before changing anything
For one working week, record the number of items, handling time, response delay and rework. This is your starting point. Use the same measures during a pilot, including the time spent checking the AI. If the problem is too small, improve the checklist or existing software first.
A practical exercise: bring one ordinary example, one exception and a rough weekly time estimate to your first conversation. Please remove client names and sensitive data.