Repeatability
High
The task runs on a fixed weekly cadence with the same four sources, same target fields, and same output format every time. Structural variation is low and bounded — new edge cases are rare and handleable with rule updates.
Ambiguity Tolerance
High
Success criteria are crisp: six named fields, YYYY-MM-DD dates, deduplicated rows, SKU conflicts flagged, one CSV out. There is almost no subjective judgment about whether the output is 'done.'
Data & Tool Availability
High
The agent needs only file access (email attachments or a shared folder) and standard libraries (pandas, openpyxl, fuzzy matching). No external APIs or live credentials are required beyond file ingestion.
Error Cost
Medium
Inventory errors can propagate into purchasing or fulfillment decisions, but the output is a flagged CSV reviewed by a human before action — not an auto-executed transaction. Mistakes are visible and correctable before downstream harm occurs.
Human Judgment Required
Low
Date normalization and column mapping are fully rule-based. SKU typo resolution is the one gray area, but the spec wisely calls for flagging conflicts rather than silently resolving them, keeping humans in the loop on ambiguous cases.