Repeatability
High
The logic is structurally identical for every record pair: normalize strings, compute similarity scores on name and domain, apply a threshold, flag borderline cases. No unique judgment is needed per row.
Ambiguity Tolerance
High
Success criteria are concrete: one master record per company, confidence scores attached, and a flagged review set. The user has already defined what 'done' looks like, including the manual review boundary.
Data & Tool Availability
High
The input is a local CSV with well-defined fields. Standard libraries (rapidfuzz, pandas, recordlinkage) cover all matching needs. No external APIs, credentials, or live data are required.
Error Cost
Low
The output feeds a mail-merge, not a financial transaction. False merges are annoying but recoverable, and the human review gate on the 80 closest calls catches the most dangerous edge cases before anything is sent.
Human Judgment Required
Low
The task is algorithmic: string normalization, domain parsing, and similarity scoring. The user has correctly reserved the genuinely ambiguous cases for human eyes, so the agent only handles the clear-cut work.