Customer Master After the Merger: Matching Without a Common Key

Start with a tiered matching design, not a single algorithm. First normalize names (transliteration, case, punctuation), dates, and addresses; then run deterministic rules on high-confidence combinations such as normalized name + DOB + postal code, and exact member number against any cross-reference table. Because SSNs are masked or missing, do not design around them. Use probabilistic matching only for the residual, with blocking keys to control cost. Calibrate thresholds on a manually labeled sample of 500-1,000 pairs stratified by score band; measure precision and recall, and set the auto-link threshold where precision is at least 99% for KYC-relevant merges. Route the middle band to a review queue. Expect the 12% estimate to be unverifiable; report a confidence distribution rather than a single duplicate rate. Publish match keys and version them, because marketing and compliance need different tolerances. Re-run matching incrementally as new records arrive, and preserve original records with link IDs so merges are reversible.

Survivorship is a business decision and must be separated from matching. Assign accountability to the retail banking data owner, with the steward proposing rules and compliance vetoing KYC attributes. Use attribute-level survivorship, not record-level: legal name and DOB from the core after verification; contact preferences from the most recent documented consent, never the oldest record; risk rating from compliance systems; marketing attributes from the most recently active source. For conflicting KYC records, do not silently pick one: create an exception case and freeze automated cross-sell until resolved. Document source precedence, effective dates, and rule version, and log every override with user and reason. Because the MDM platform is a year away, implement the rules in a governed cross-reference table with change control, not hard-coded SQL. Give compliance read access to the merge audit trail. The key judgment is that over-merging two customers is worse than leaving a duplicate for review, because it can expose one person's data to another.

Measure the program on outcomes and error rates, not on match counts. Track duplicate rate by confidence band, false-positive rate from audit samples, manual review throughput and aging, downstream KYC duplicate cases, and the number of automated merges reversed. Add business metrics: reduction in duplicate mail cost, time to onboard, and cross-sell eligibility accuracy. Set a target such as a 90% reduction in high-confidence duplicates with a false-positive rate under 0.5%, and report confidence intervals because the 12% estimate is itself uncertain. Monitor drift when the new MDM platform arrives. The principal failure mode is a precision collapse under pressure to hit the deadline: teams lower thresholds, over-merge, and create a regulatory incident. Guardrail: any threshold change requires owner and compliance sign-off, and a sample is re-reviewed monthly.