A large enterprise has customer records in a dozen systems with no shared identifier. Design the master data approach.

The first decision: which style

• Registry — a mapping of identifiers with no central store. Least invasive, and every consumer must still resolve.
• Consolidation — a central golden record built from sources, read-only. Good for analytics, does not fix the operational systems.
• Coexistence — golden record synchronised back to sources. Better, and requires every source to accept updates it did not originate.
• Centralised — one system of record that all others read. Cleanest and the largest change programme, and rarely achievable in a large estate without a forcing event.

Most enterprises should start with registry plus consolidation and move selectively toward coexistence for the attributes that genuinely need to be consistent operationally.

The matching problem

Probabilistic matching over names, addresses and contact details produces two error types with very different costs:

• False merges combine two real customers, which is a privacy incident and is hard to unpick.
• False splits leave duplicates, which is annoying and safe.

Tune deliberately toward false splits, with a stewardship queue for the uncertain band. The threshold is a business decision about which error is worse in this domain, not a model-tuning exercise.

What determines whether it lasts

• Survivorship rules stated per attribute — which source wins for address, for contact preference, for legal name — rather than a global source ranking that is wrong for half the fields.
• Stewardship capacity, because the uncertain band is permanent and unstaffed queues make the whole thing decorative.
• Feedback to sources. If duplicates are only fixed centrally, the sources keep producing them and the effort is permanent.
• Lineage on the golden record, so a disputed value can be traced to its origin.