Cleaning a messy nonprofit spreadsheet is exactly the kind of grunt work AI handles well.

The operations are structurally identical every time: normalize strings, parse dates, resolve location aliases, flag or merge duplicates. This is a well-worn data-cleaning pattern with no unique judgment required per row.

Most success criteria are crisp — consistent date format, canonical location names, no exact duplicates — but fuzzy deduplication (same pair with slightly different names) requires a defined matching threshold, and the right threshold isn't fully specified in the task.

The source data is a Google Sheet, which is trivially accessible via API or export. A code agent can read, transform, and write back the cleaned file without needing any external context or credentials beyond sheet access.

Incorrectly merging two distinct mentor–mentee pairs or dropping a valid record could corrupt grant reporting data, which has real downstream consequences. However, the original sheet is preserved and the output is reviewable before submission, making errors recoverable.

The vast majority of decisions — date parsing, location normalization, exact duplicate removal — are rule-based. Only edge cases in fuzzy name matching benefit from human review, and those can be flagged by the agent rather than silently resolved.