Giving an AI agent more tools to clean messy data doesn't necessarily make it better at cleaning messy data.
Researchers tested seven configurations of LLM-based data-cleaning agents against financial, clinical, and environmental-monitoring datasets, running 126 total trials using synthetic corruption. The setups ranged from basic LLM prompting up to a "full conservative" version that added executable checks, evidence retrieval, source ranking, citation alignment, and reversible repair scripts. A plain deterministic profiling baseline, with no LLM involved, beat every AI configuration on raw error detection, scoring an F1 of 0.561 versus the best LLM setup's 0.421. The most cautious, evidence-heavy configuration made zero unsafe or unnecessary edits, but that config also skipped direct repairs entirely.
The real finding here isn't that AI failed, it's that there's no free lunch. Stacking on more capabilities like retrieval, citations, and conservative guardrails traded away detection power or repair activity even as it improved safety and traceability. No single configuration won on every measure the researchers tracked.
Anyone planning to point an autonomous agent at production data without a known-clean reference should read this as a caution label, not a feature list.