Repeatability
High
The extraction schema is fixed — same seven columns every time — and the duplicate-detection rule is deterministic. The structure is highly repeatable even if individual receipt quality varies.
Ambiguity Tolerance
Medium
Column definitions are mostly crisp, but 'category' requires judgment calls (e.g., a SaaS tool bought for a client trip — travel or software?), and 'project code' may not appear on the receipt at all, requiring external context the agent likely lacks.
Data & Tool Availability
Medium
The agent needs access to all 340 files and a capable OCR pipeline (e.g., GPT-4 Vision, AWS Textract, or similar). Blurry or rotated images will degrade extraction quality, and project code mapping requires a reference the user must supply.
Error Cost
Medium
Wrong amounts or dates in an expense CSV used for tax filing or client billing carry real financial and compliance risk, but the output is a reviewable CSV — errors are catchable before they cause irreversible harm if the user audits the output.
Human Judgment Required
Medium
Category assignment for edge cases and project code inference genuinely need human context. Blurry receipts where OCR fails entirely also require a human to manually read the source document.