AI can do most of the heavy lifting here, but vendor deduplication will need a human eye.

The extraction schema is consistent (same 7 fields per vendor), but the source documents vary wildly in format, language, and structure. Scanned images add OCR variability, and deduplication logic must be applied case-by-case, making this partially but not fully repeatable.

The output format is well-defined (one CSV, 80 rows, specific fields, two flag types), which is favorable. However, success criteria for deduplication—deciding when two vendor names or files represent the same entity—are inherently ambiguous and require judgment the agent may not apply consistently.

The agent needs file access to 650 documents across a messy folder structure, OCR capability for scanned images, and document parsing for PDFs and Word files—all achievable with current tooling. The main gap is that the agent cannot independently resolve ambiguous vendor identity without external context the user holds.

A missed renewal date or wrong contract end date could cause a lapsed contract or missed negotiation window, which has real business cost. However, the output is a CSV that a human can review before acting on, making errors catchable before they cause damage.

Deduplication across vendor aliases, partial name matches, and split files requires contextual knowledge the agent lacks—e.g., knowing that 'Acme Creative LLC' and 'Acme Creative' are the same vendor. A human must validate the deduplication decisions and flag any extraction anomalies before the CSV is trusted.