Use fictional inputs to develop the logic. Test it, review it, then run it in the approved environment. Getting help to design a process and sending it employee records to process are different things.
Help building it, not uploading it.
In one practitioner discussion, someone concerned about employee data used AI to learn Excel and Power Query, then refreshed their reports from an export of the HR system. That is one reported experience, and it points to an opportunity worth exploring: use the assistance to develop the process without sending live records to the model.
1. Share the structure. Invent the examples.
Explain the input columns, the output you need and the rule you want to apply.
Create fictional data with invented IDs, dates and categories. Include known totals and a few edge cases.
Leave real records out. Removing the names isn’t enough. Build the examples from scratch.
Practical check: Also check file names, hidden sheets and error messages.
2. Ask for a method you can inspect.
Request a draft formula, query or script for the fictional example.
Understand it. Ask for each step, assumption and dependency, and any connection it would make.
Challenge it. Check the explanation against the code. A confident explanation is not a test.
Practical check: Use your approved tool, even for development work.
3. Know the answer before checking it.
The normal case. Compare the output with totals you have calculated independently.
Edge cases. Try missing IDs, duplicates, unexpected dates and changed column names.
Change control. Keep the tested version and the expected results. Retest when the inputs or the logic change.
Practical check: One successful sample doesn’t validate every case.
4. The finished workflow may not need AI.
Example. An approved query refreshes a report from the designated export folder.
Keep the separation. This design doesn’t require sending the employee rows to a model when it runs.
Verify the setup. Review the code, connections, storage and access before using real records.
Practical check: No AI call doesn’t mean no cloud services or no data risk.
5. Debug the process without exposing the data.
Reproduce safely. Recreate the error with fictional records before asking for help.
Inspect what you share. Remove credentials and private details from logs, screenshots and sample files.
Review before expanding. Connecting AI to live records needs its own review of data and access.
Practical check: Don’t bypass company restrictions to make a test work.
AI can help build it. It need not run it.
Choose one repeated spreadsheet step and create a fictional test set. Develop, verify and document the method, and keep the boundary around live data explicit.
Read more: When Has a Spreadsheet Stopped Being the Right Tool?