CRM Deduplication and Data Hygiene for Multi-Entity Groups

A one-off deduplication run buys about five months. In the CRMs we audit, duplicate rate returns to within a point of where it started inside two quarters unless intake changes at the same time. Forms, list imports and the API are where duplicates are manufactured, and cleaning without closing them is a subscription to cleaning. Matching on identifiers, not names: Fuzzy name matching on company records is where most deduplication projects fail, and it fails in both directions at once. Novák, spol. s r.o. and Novak s.r.o. are the same firm, and no string-similarity threshold catches that reliably. Meanwhile three plants of one industrial group share a name almost exactly and must never be merged, because each has its own ship-to address, its own buyer and its own open quotes. We match on registration identifiers first: IČO and DIČ resolved against ARES, EU VAT numbers against VIES, company numbers wherever the registry exposes them. Name similarity is a signal that sends a pair to review, never an instruction to merge. Survivorship, decided per field: A merge is a set of small decisions and they should not all default to the older record. We write survivorship field by field with a ranked source: billing address from the ERP, owner and lifecycle stage from the CRM, consent state from the most recent form submission regardless of record age, phone numbers unioned rather than overwritten. Every merge writes an audit entry naming the losing record ID, the rule applied and the timestamp, so a merge can still be explained a year later when somebody asks where an account went. The cases we deliberately do not merge: Contacts sharing an email domain, which is routine in recruitment where a candidate at one client is a hiring contact at another. Records differing only in a field neither source can arbitrate. Anything with an open deal on both sides: those go to a human queue showing both records side by side, because merging a live opportunity badly costs more than a duplicate ever will. Closing the intake: A duplicate check at form submission that resolves on the identifier rather than the email address. A staging step for list imports, so a marketing upload cannot write directly to production. And a written contract for whatever your ATS, webshop or ERP sends, with rejection on ambiguity instead of best-effort insertion. Not included: We do not buy or resell enrichment data, and we do not fill blank fields from third-party providers as part of hygiene work; that is a separate decision with its own consequences. We do not give legal advice on GDPR, though we will implement the consent and retention fields your counsel specifies. And we do not delete. Losing records are merged and archived, never removed, because the only thing worse than a duplicate is a record that vanished without a trace. Who this is not for: Anyone who needs a clean number before a board meeting. This work makes the account count honest, and honest usually means the first report shows fewer accounts than the old one did, which is a conversation better had in advance. Also not for teams unwilling to change how records enter the system. Without that, we would be selling the same run again next year, and we would rather decline than put it in the calendar.