Prepare an AI-assisted CRM duplicate review before merging records

Define what counts as the same entity

Ask the CRM owner to define matching rules for people, companies and locations separately. A parent company and a subsidiary may share branding while needing distinct records. Two contacts with the same name may be different people. Prepare a permitted export with record IDs and the minimum approved matching fields. Do not upload an entire contact database into an unapproved AI tool.

Document the current system’s identifiers before looking for similarities. HubSpot uses email for contact deduplication and domains for companies, with Record IDs or unique-value properties available for import matching. Those product behaviors do not resolve every real-world identity question.

Create a comparison packet, not a winner

For each candidate pair show both record IDs, exact matching evidence, conflicting fields, relationships, account owners and unresolved identity questions. Keep source timestamps so a reviewer can tell a recent correction from an old field. Classify the recommendation as investigate, distinct or supported candidate; avoid presenting a confidence score as verified identity.

In a fictional distributor, two records share a domain but represent different regional business units with separate service agreements. The correct next step is to ask the account owner about the record model. Automatically choosing the oldest record could erase a useful commercial distinction. Record what evidence would resolve the question.

Require an explicit merge decision

A CRM administrator should inspect current merge behavior, associations, field precedence and recoverability before any authorized action. Check subscription restrictions and retention requirements with the responsible owner. Never assume that a preferred email field authorizes outreach or that combining records should remove an opt-out. AI should have no merge or delete access for this preparation task.

After approved changes, verify the resulting record, relationship links and active work ownership. Maintain a decision register for rejected candidates as well, so the same legitimate pair is not repeatedly proposed. Measure reviewer corrections and false matches on a small sample. Reducing record count is not the objective; preserving accurate commercial context is.

Review checklist

• Define identity separately for people and organizations.
• Use stable record IDs in each candidate packet.
• Show conflicts and relationship differences.
• Review consent and merge consequences before action.
• Record rejected matches and verify approved changes.

Source context and limits

HubSpot documentation supports the identifier examples. The candidate review method is editorial guidance; merge behavior, permissions and recovery must be checked in the actual CRM before action.