AI-assisted, human-approved
AI data cleaning you can inspect before it runs
A useful cleaning assistant should reduce manual work without turning your spreadsheet into a black box. This workflow profiles the file with deterministic code, asks AI for suggestions within a fixed rule language, and waits for your approval before execution.
By the TidyX team · Updated
Is AI useful for data cleaning?
Yes, for deciding what to fix, as long as a person approves each change and deterministic code does the work. A model can read a summary of the file, suggest candidate rules such as trimming stray spaces, treating N/A as missing or parsing a date format, and explain why each might help.
It needs guardrails where the data actually changes. A model's description of a change is not a measurement of it, and asking again can return a different plan, so counts, row changes and repeat runs should come from code that ran on the whole file. In this workflow, AI takes part in one of five steps.
- Profile: deterministic code counts rows, columns, missing values and duplicates and infers column types. No AI is involved.
- Propose: AI reads the profile and the context you supply, then suggests rules from a fixed list, each with a rationale.
- Review: you accept, skip or edit each rule; high-risk rules need an extra acknowledgement.
- Run: a deterministic engine applies the confirmed rules and writes a new output file.
- Verify: the report records actual changes, warnings, hashes and the engine version.
What the AI does—and what it never does
The model proposes a structured plan from the data profile. It can describe why a rule may help, identify the columns it targets, and estimate impact. It does not receive permission to write arbitrary code, directly edit your source file, or silently run a destructive transformation.
Sample sharing is optional and off by default. The model receives the profile plus selected fields from the context and dictionary you supply, but no sample rows or per-column sample values unless you opt in. Opting in adds masked per-column examples and at most five masked rows, leaving out columns marked sensitive.
- Fixed, validated cleaning operations
- Explicit review for every suggested rule
- Separate deterministic execution after confirmation
- A new output file instead of overwriting the original
A practical review-first workflow
Start with a CSV or XLSX file. The profiler reports row and column counts, inferred types, missing values, duplicates, and other quality signals. That profile is useful on its own and gives the plan generator bounded context about the shape of the dataset.
After the plan is generated, choose Review all for a compact pass or Step by step for focused decisions. Both views edit the same plan. You can accept, skip, or adjust supported parameters, and high-risk changes require an additional acknowledgement before confirmation.
Results designed for verification
The execution engine applies the confirmed rules in order and writes a new cleaned file. The report shows actual row and column changes, warnings, rule outcomes, input and output hashes, the frozen rule snapshot hash, and the engine version recorded by the run.
New accounts receive 60 one-time signup credits, no card required. AI planning and each cleaning run have separate quotes to approve; uploading, profiling, rule review and existing downloads do not consume credits. Paid Trial, Monthly and Yearly plans are described on the pricing page.