Repeatability
High
Data normalization tasks follow consistent patterns: trim whitespace, lowercase emails, reformat phone numbers with regex, and apply fuzzy matching via pg_trgm or Levenshtein. The structure is the same every time, making this highly automatable.
Ambiguity Tolerance
Medium
The formatting rules are well-specified, but fuzzy matching thresholds (what similarity score counts as a 'likely duplicate') require a judgment call. The agent can produce a reasonable default and document it, but a human should validate the threshold before running.
Data & Tool Availability
Medium
The agent needs the actual table schema, column names, and ideally sample data to write accurate SQL — without these, it must make assumptions that may not hold. If the user provides the schema, this becomes straightforward; if not, the script will need manual adjustment.
Error Cost
Medium
The script itself is low-risk to generate, but if run directly on production data without review, a bad UPDATE or DELETE could corrupt 500K records irreversibly. The task wisely scopes output to a migration script and merge report for manual review, which keeps error cost manageable.
Human Judgment Required
Low
Formatting normalization is rule-based and requires no taste or intuition. Duplicate detection thresholds are the one area needing human sign-off, but the agent can surface those decisions explicitly rather than hide them.