Describe the rule, the columns and the expected result using invented rows, including awkward ones. Building an automation is different from processing live records.
Can’t I just remove the names and upload it?
Invented rows are not real records with the names removed.
1. Check the calculation and the connections.
Logic. Test known totals, duplicates, dates, currency and missing values.
Behaviour. Review what the code reads, writes, deletes or sends, and which outside services it contacts.
Access. Use the approved environment and the minimum permissions. Don’t disable privacy controls to make an error go away.
Practical check: A local spreadsheet can still contain queries that contact other systems.
2. Debug without private records.
Before sharing an error. Remove bank details, credentials, private paths and customer or employee information from logs and screenshots.
Reproduce it safely. Create a fictional case showing the same failure, and compare the result to a known answer.
Practical check: A generated formula or script still needs checking before it touches real data.
3. Hand it over.
Document the inputs, the checks and the owner. The finished automation may run without AI.
Practical check: Approval covers one tool, task and data set. It is not a blanket licence.
Read more: Using AI on HR Spreadsheets Without Sharing Employee Data · Using AI for Spend Analysis Without Sharing Supplier Data
This is general guidance, not security or accounting advice.