As of 13 August 2026, AI can only partly clean and deduplicate your business database.

What to actually do

1. Hand it to a person

   The route this page recommends

   A person who owns the outcome does this end to end, worth it when the failure is dear.

2. Use a tool built for this

   Second choice

3. Do it yourself

   The distant third

   A chat interface, power-user skill, and roughly 1 hour until you can act on the result.

   How to actually do it

   1. Open your database administration tool and create a dated backup or restorable snapshot before exporting anything.
   2. Export a representative sample with stable record IDs, the relevant identity fields, relationship fields and audit or consent fields, and remove data that the model does not need.
   3. Write down the database system and version, the fields that must not be overwritten, and the business rule for deciding when two records are the same.
   4. Paste the prompt and the schema, duplicate policy and redacted sample into an approved AI tool, then answer its questions about missing schema or field meanings.
   5. Save the AI's report-only duplicate list and run its proposed checks against a copy of the database, comparing row counts, unique IDs, linked records and protected fields with the original.
   6. Ask a colleague who knows the customers or companies to approve ambiguous groups, then apply only approved changes through a dry run and keep the original IDs and restore point.

   Prompt

   Act as a data-cleaning engineer. I will provide a database export, its column definitions and our duplicate rules.
   
   Do not change or delete anything. First:
   1. Describe the data types, missing values, inconsistent formats and suspicious values you find.
   2. Propose normalisation rules for names, addresses, telephone numbers, email addresses, company numbers and dates, but do not invent missing values.
   3. Identify likely duplicate pairs or groups and give each a confidence level with the exact fields that support the match.
   4. Keep records separate when the evidence is weak, and flag ambiguous cases for human review.
   5. Choose a surviving record only when the rules explicitly allow it. Otherwise recommend which fields should be reviewed.
   6. Produce a safe, reversible plan for a dry run, including a proposed output table with original IDs, group IDs, match reasons, confidence, recommended action and fields that would change.
   7. If code or SQL would help, write it for [database system and version], but make it report-only by default. Do not include DELETE, UPDATE or irreversible merge commands unless I explicitly request them after reviewing the report.
   8. State what I must check before applying any changes, including row counts, unique IDs, consent or suppression fields, linked records and a restore plan.
   
   Use only the supplied data and rules. Do not guess that two records are the same merely because their names look similar. Ask concise questions if the schema, database system or duplicate policy is missing.
   
   Business context: [what the database contains]
   Duplicate policy: [what makes two records the same]
   Fields that must never be overwritten: [list]
   Database system: [system and version]
   Data export and schema:
   [paste a redacted sample or attach the export here]

   Open it prefilled in ChatGPT or Claude, or copy it into Gemini, which takes no prefill link.