Article overview
- An AI spreadsheet-cleanup workflow should preserve product identifiers as identifiers, even when they contain only digits. Define protected columns before changing data types, trimming values or asking an assistant to deduplicate rows.
- This is especially relevant when a spreadsheet moves between a supplier, an ecommerce catalogue and an operations tool. Each application can apply its own formatting and import assumptions.
- Identify product codes, external record IDs, postal codes and other reference fields. Keep their original representation where the destination requires it.
- Microsoft's guidance on leading zeros and large numbers explains why import types matter in Excel.
Putting this workflow into practice? Explore AI Workflow Automation for implementation scope, controls, and next steps.
An AI spreadsheet-cleanup workflow should preserve product identifiers as identifiers, even when they contain only digits. Define protected columns before changing data types, trimming values or asking an assistant to deduplicate rows. A clean-looking workbook can be unusable if a code no longer identifies the same item.
This is especially relevant when a spreadsheet moves between a supplier, an ecommerce catalogue and an operations tool. Each application can apply its own formatting and import assumptions. Test the final exported values, not only what the current cell appears to display.
Mark fields that must not be calculated
Identify product codes, external record IDs, postal codes and other reference fields. Keep their original representation where the destination requires it. A numeric-looking code is not necessarily a quantity, and two similar-looking codes are not necessarily duplicates.
Microsoft's guidance on leading zeros and large numbers explains why import types matter in Excel. Choose the supported text-import controls for the actual version rather than assuming a later display format restores information already lost.
Write a transformation brief the reviewer can check
Specify which columns can be cleaned, which can be standardised and which must be preserved exactly. For example, removing accidental spaces from a descriptive category may be acceptable while changing an external item code requires a controlled mapping.
Tell the assistant how to represent uncertainty. If a value is unreadable or a code conflicts with the master, return a review item with its source row. Do not ask the model to infer a missing digit from nearby records and silently replace it.
Keep the original and transformation evidence
Work from an approved copy and retain the source reference. Store the transformation instructions, output version and review decisions together. This makes a later failed import investigable without requiring someone to remember which prompt was used.
Restrict the data supplied to the assistant to what the task requires and what the organisation permits it to process. Product-data work can still contain private commercial fields. Sharing permission should be resolved before uploading a workbook, not assumed from the ability to open it.
Worked example of two codes becoming one
Illustrative scenario: a source catalogue contains codes 00123 and 123 for different products. A cleanup converts both to the number 123 and then treats the second row as a duplicate.
| Check | Expected result | Reason |
|---|---|---|
| Source-to-output identifier comparison | Both original codes retained | They identify different source records |
| Master lookup | Each resolves to its approved item | Formatting alone does not establish identity |
| Duplicate review | No automatic merge from numeric equality | The loss occurred during transformation |
Keep a fixture containing both codes, a long identifier and a value that resembles a date. Reopen the exported file through the intended destination's import path. A screenshot of the workbook is insufficient evidence if the exported value changes.
Validate independently of the model
Compare protected fields exactly, reconcile record counts and check the keys that link to downstream systems. Validate quantities and totals with ordinary deterministic checks where possible. A valid table structure does not prove the contents remain correct.
The AI extraction review guide explains the distinction between well-formed output and verified values. Apply the same principle to catalogue cleanup: model confidence should not substitute for a source comparison.
Prepare the handover and rerun rules
Record which file is approved for import and what should happen when the source is revised. A later run must not silently overwrite a manually approved mapping. Preserve the reason for corrections and use difficult rows as regression examples.
If the cleaned data feeds management reporting, follow the Excel-to-presentation workflow only after the input is reconciled. Explore AI workflow automation, or describe the workbook and destination to scope the transformation and review checks.
Frequently asked questions
Can formatting a column later restore a damaged identifier?
Not reliably. A display format may make a value look similar without recovering the original source information. Compare against the retained source and apply a reviewed correction when the original representation matters.
Should an AI assistant decide which product codes are duplicates?
It can assist a review, but the merge rule must come from the business's identity model. Keep the source references and require evidence before combining records that downstream systems treat as different products.
AI Workflow Automation
I design workflows that use Claude, DeepSeek, OpenAI, spreadsheets, slides, docs, and business data to save real team hours.