Practical AI · 3-minute read
Your customer database contains Dave, David and D. Martinez. One might be a duplicate. One might be the purchasing manager’s replacement. One might be a perfectly innocent Dave who would like everyone to stop sending him pump parts.
AI can help find records worth checking. The valuable output is a review queue with reasons, not a freshly merged database nobody can explain.
Define what one record represents
West Mesa Supply, a fictional oilfield supplier, is preparing a customer newsletter. Its practice file contains invented companies and contacts. Some rows represent a business, others a branch, and others a person. Combining them because they share an address would erase useful distinctions.
Before cleaning anything, decide what each record represents and which fields matter for this task. A company name, contact name and customer account number do different jobs. Keep a stable record ID so every suggestion can be traced back to its source.
The UK Government Data Quality Framework distinguishes accuracy, completeness, uniqueness, consistency, timeliness and validity. A filled-in field can still be wrong. A correctly formatted email address does not prove that it belongs to the right customer. [1]
Separate easy formatting from uncertain identity
A consistent date format is a different decision from merging two contacts. Use existing database rules for routine validation where they work. AI can help explain unusual entries, group possible matches and organize questions for the record owner.
Review this fictional or approved customer sample using these field definitions and matching rules. Return record IDs, suspected issue, supporting fields, conflicting evidence and the question a reviewer must resolve. Distinguish formatting issues from possible duplicate identities. Do not invent missing values, infer contact permission, merge records, change subscription status or enrich the list from outside sources. Mark uncertain matches ‘Needs review.’
Two people can share a name. A shared office address or generic inbox can serve several contacts. An older record may contain an important restriction that the newer one lacks. Preserve those conflicts for review instead of rewarding whichever row looks tidiest.
Test the cleaned sample before the campaign
Keep an access-controlled original and a change log. Have the customer-record owner confirm proposed identity matches using authoritative business records. Use an approved tool and the minimum necessary fields; practice with fictional data first.
Check known duplicates and known separate people to see whether your method gets both right. Review names with accents, alternate spellings and different naming conventions. Do not assume unfamiliar names are errors.
Before using the result, verify that campaign exclusions and unsubscribe records still work. Cleaning a list does not create permission to contact anyone. Confirm ambiguous contact preferences through your established process.
Your turn: Which field in your customer list looks complete but cannot currently be trusted?
Source [1]: UK Government Data Quality Framework. Supports the data-quality dimensions. The supplier, matching workflow and prompt are original guidance; campaign permission decisions require the business’s established process.
