Merging two messy account databases is exactly the kind of data grunt work AI handles well.

The structure is consistent: two files with known fields, a normalization step, fuzzy matching logic, and a consolidated output. This pattern is the same every time it runs, making it highly automatable.

The output format (deduplicated CSV with source column, plus a review list) is clear, but thresholds for what counts as a 'definite match' versus a 'potential duplicate' require tunable judgment. The agent needs explicit matching rules or will make arbitrary cutoff decisions.

Both files are static exports (Excel and CSV) that can be handed directly to the agent — no live API access or special permissions required. Standard Python libraries (pandas, fuzzywuzzy/rapidfuzz, phonenumbers) cover all normalization and matching needs.

False negatives (missed duplicates) could mean duplicate outreach to the same account, which is embarrassing but recoverable. False positives (wrongly merged records) could corrupt CRM data, but the manual review list acts as a meaningful safety net for high-confidence matches.

Borderline cases — subsidiaries with different names, rebranded companies, contacts who moved between firms — genuinely need human review. The task wisely scopes this in by asking for a flagged list, so the agent handles the bulk work and humans handle the edge cases.