Repeatability
High
Every invoice targets the same five fields, and the extraction logic is structurally identical across documents. Variation in layout and quality is handled by OCR preprocessing, not by changing the task definition.
Ambiguity Tolerance
High
The output schema is fully specified — five named columns with clear semantics. Success is measurable: each row either has the correct values or it doesn't, and human verification provides a defined quality gate.
Data & Tool Availability
High
The input files are a local folder of PDFs and images, accessible to any OCR pipeline. Mature tools (AWS Textract, Google Document AI, Azure Form Recognizer, or open-source alternatives) handle rotation correction and mixed-quality scans out of the box.
Error Cost
Low
The task explicitly includes human verification before the CSV is used, so extraction errors are caught before they propagate. Downstream risk is low because no irreversible action is taken on the raw AI output.
Human Judgment Required
Medium
Most invoices are routine, but handwritten annotations, smudged dates, and ambiguous line-item groupings genuinely require human review. The task already accounts for this, making the human role targeted rather than pervasive.