Repeatability
Medium
Native PDFs follow predictable structures, but scanned images vary in quality, orientation, and layout across 18 months and potentially multiple statement formats. The agent must handle structural variation without a single fixed template, which reduces repeatability.
Ambiguity Tolerance
Medium
The output schema is well-defined (date, description, amount, account type, deduplicated CSV), but edge cases like split transactions, fee reversals, or ambiguous duplicate detection rules are not specified and require implicit judgment calls.
Data & Tool Availability
High
The files are self-contained and can be handed directly to the agent. PDF parsing libraries, OCR engines, and CSV output tools are mature and readily available, so no external API access or live credentials are needed.
Error Cost
High
Incorrect amounts, missed transactions, or false deduplication in financial records can cause downstream accounting errors, tax miscalculations, or audit issues. These errors are reversible in principle but costly to find and fix after the fact.
Human Judgment Required
Medium
Most extraction is mechanical, but OCR ambiguities on scanned statements (e.g., '0' vs 'O', smudged digits) and deduplication edge cases benefit from human review. The agent cannot know when a near-duplicate is a legitimate repeated charge versus a true duplicate statement.