Refactoring a legacy ETL pipeline is squarely in AI's wheelhouse — with a human sign-off before it touches production.

ETL refactoring follows a well-worn pattern: parse the existing logic, map it to idiomatic Python, add error handling and logging layers. The structural work is highly repeatable and doesn't require novel judgment each time.

The performance target (40 GB in under 2 hours) and feature requirements (row validation, rejection handling, logging, benchmarking) are concrete. However, the fixed-width field schema, rejection policy semantics, and acceptable data quality thresholds are not specified and will require clarification or inference from the source script.

The agent needs the 800-line Perl script, the fixed-width field layout specification, sample input files, and PostgreSQL schema details. If these are provided, the agent can proceed; without them, it can only produce a structural scaffold that requires significant human gap-filling.

A buggy ETL pipeline silently corrupting or dropping rows in a 40 GB production dataset is a serious but recoverable problem — the original data files still exist. The risk is real but bounded, and a proper test harness against known-good output mitigates it substantially before deployment.

Decisions about what constitutes a 'malformed row,' how aggressively to reject versus coerce data, and whether the refactored output is semantically equivalent to the original all require a domain-knowledgeable human to validate. The coding mechanics are automatable; the correctness sign-off is not.