Cleaning 3,500 messy real-estate records is a textbook job for a data agent.

Address normalization and deduplication follow deterministic rules (USPS standards, fuzzy matching on address + date) that apply uniformly across all 3,500 rows. The same logic runs identically on every record, making this highly automatable.

Success criteria are crisp: no duplicate address/sale-date pairs, standardized abbreviations, and clean keys that join to the CRM. An agent can validate its own output against these rules without subjective judgment.

The input is a self-contained spreadsheet with well-defined columns. Standard libraries (pandas, dedupe, usaddress) and USPS address normalization APIs provide everything needed; no live external access or special permissions are required.

A bad merge or missed duplicate could corrupt CRM records or skew the comparison tool's comps, which is a real downstream problem. However, the original source data is preserved and errors are detectable and reversible before the cleaned file is ingested.

The vast majority of decisions are rule-based. The only genuine judgment calls are rare ambiguous cases (e.g., is '123 Main St Unit 4' the same listing as '123 Main Street #4'?), which can be flagged for a quick human review rather than blocking full automation.