1) Generate sensible candidate pairs (afternoon)
Start with simple, high‑precision matching rules you can run in an afternoon: exact email matches, normalised phone numbers, same company + similar name, or identical VAT/company numbers. Run these as scheduled batch jobs (daily or weekly) rather than real‑time to keep scope small.
If you use HubSpot, Salesforce, Marketo or Pardot the rule idea is the same: produce a short candidate list rather than a single canonical action. Export candidate pairs to a sheet or tag them with a lightweight property in your CRM so the AI and reviewers work from the same queue. For practical HubSpot patterns see [hubspot-crm-data-cleanup.html].
2) Score suggestions with AI and capture a short rationale (same day → week)
Send each candidate pair and a small redacted context packet (last activity, owner, key fields) to your chosen AI model to produce a confidence score (0–100) and one or two lines of rationale (e.g. “emails match; phone differs by country code”). Keep the prompt focused and limit the fields included so the model can't hallucinate unrelated facts.
Record the score, rationale and a small provenance set as properties on the CRM pair or as columns in your review sheet: suggested_by=ai, confidence_score, rationale, snapshot_ids. Never let the model write the merge directly — write suggestions only. Gate any automated merge behind a human approval step.
3) Human review, audit trail and safe escalation (week → ongoing)
Place suggestions in a visible review queue (HubSpot review properties, a shared sheet, or a simple Trello/Asana board). Reviewers see the score, rationale and a short snapshot of the two records plus the suggested merge action. Use a simple policy: auto‑apply merges only above a high confidence threshold and only for non‑sensitive records; route uncertain pairs to a human reviewer. Add a per‑record 'do_not_automate' flag for VIPs or sensitive accounts.
Keep an audit trail and rollback path: before any merge, export a minimal snapshot (IDs + key fields + timestamps) and store it with the review record; include reviewed_by, review_time and review_decision. Measure precision by sampling reviewed merges (approved and rejected) and calculate false positive rate; increase the auto‑merge threshold slowly as measured precision improves. Simple operational controls — a reviewer rota, short SLA for queue items and a one‑click revert from the snapshot — make this suitable for small teams on the South Coast or in Fareham and Hampshire.
If you want a short hand to get started, Optira can help set the scoring prompts, review queue fields and rollback snapshots so your CRM optimisation and workflow automation runs safely and predictably.