The CRM holds a new client address, the contract uses the old one, and accounting has a third. Declaring the CRM the source of all truth is insufficient. First identify the kind of information, who may validate changes and which systems need the update.

Assign a source to each field

A communication contact, a contracting party’s details and a delivery address can have different owners and validation rules. Displaying them together does not make them interchangeable. A shared record needs the value, source, effective date and relevant client role.

For an illustrative example, a manager receives a new contact phone number. This does not justify changing the legal entity’s details. A requested change to those details follows its own agreed validation process.

Separate a proposed change from confirmation

ActionResultControl
Propose a changeRequest with a sourceInput completeness
Validate the valueAccepted or rejected requestDefined authority
Update the sourceNew record versionChange history
Send to consumersAccepted updates or exceptionsBusiness-result validation

This does not require four people for every edit. Control depends on the consequences of an error. The system should still distinguish an unverified proposal from a value approved for use.

Deduplication must preserve history

Similar names do not prove that records belong to the same client. Before merging, check identifiers and linked contracts, enquiries and orders. Decide which record remains authoritative, how links move and how an incorrect merge can be repaired.

Also consider concurrent changes. Two teams may update a record at once. “Last write wins” is not appropriate for every field. Conflicts in consequential fields need a visible review path.

AI can propose a change, not grant authority

An assistant may extract details from a document or suggest similar records. Its confidence does not establish field correctness. Results require a source, access checks and an agreed decision about further use.

Start with a bounded field set and evaluation cases: similar names, outdated documents, missing values and different access rights. Assess the need for a separate MDM solution after this work. Rules and integration may serve a small process; a more complex one may need a dedicated architecture.

  • List key fields and their consumers.
  • Assign each field a source and owner.
  • Agree change validation and conflict review.
  • Test repeated updates, duplicates and recovery.

Questions and answers

Must all data be in one system?

You need aligned values and accountability. The architecture may use several systems with explicit field sources and exchange rules.

Where should deduplication begin?

With match criteria and checks on linked records. Bulk merging without recovery rules can create harder problems.