Stop duplicates before they enter the CRM, not after

Create a workflow 'Real-time CRM dedup' triggered immediately on account_created OR user_created events. Goal: catch duplicates at creation moment rather than letting them propagate, then needing cleanup later.

Configure find-records step that searches for existing records similar to the newly-created one. Match criteria: same domain (for accounts), same email (for users), fuzzy name match (Levenshtein distance < 3), same primary contact. Returns: list of candidate matches with similarity scores.

Branch step: did the search find any candidate matches? If NO matches — record is unique, proceed to standard onboarding flow (handoff to auto-enrich-new-accounts). If YES matches found — continue to AI confidence scoring.

Configure AI step that compares the new record to each candidate match and produces a confidence score (0-100) per pair. Inputs: all available fields, recent activity, contextual clues (e.g. same source UTM suggests same person). Considers nuances (e.g. 'sales@acme.com' and 'john@acme.com' are different people at same company, not duplicates). Output: ranked match candidates with confidence + reasoning.

Multi-split based on top candidate's confidence score: HIGH (>90) — auto-merge into existing record; MEDIUM (60-90) — flag for review in dedup queue, do NOT auto-merge; LOW (<60) — likely not a duplicate, proceed with new record but tag as 'review-candidate' for periodic re-evaluation.

On high-confidence branch: merge the new record into the existing one. Keep the existing record as survivor (preserves history), copy any new fields from the new record, mark the new record as 'merged into [existing_id]'. The new record reference still works for inbound webhooks but redirects to the survivor.

On medium-confidence branch: create a RevOps task with both records side-by-side, AI's reasoning for the suspicion, and merge/separate decision buttons. SLA: review within 5 business days — uncertain duplicates accumulate fast if not handled.

Validate and publish. Monitor: auto-merge volume per week (proves the workflow is catching real duplicates), false-positive rate (sample audit of auto-merges — were any wrong?), review-queue depth (high = too many medium-confidence cases, suggests AI prompt tuning). Together with the scheduled-data-quality-audit, this is the dedup defense-in-depth pattern.