- Two systems hold different versions of the same customer or product.
- Meetings spend time arguing about whose figure is right.
- Somebody re-keys or reconciles the same information as part of their job.
What this phase buys you
Numbers stop being negotiable. The things you found mean one thing across systems, so an automated answer can be trusted.
Three entities, a paragraph each saying what counts and what does not, one system of record, and one named owner. Three pages at most.
Not this. Not a master data programme and not a governance function. Both are how this turns into two years of meetings.
The three entities that matter most have one authoritative definition and one named owner each.
Reconciliation becomes somebody's permanent job, and any automated output inherits the disagreement and gets blamed for it.
What you actually do
Four moves. None of them needs a platform, and none of them is glamorous. That is the point: this is the work that gets skipped, which is why so many companies are stuck a phase earlier than they think.
Pick three entities, not thirty
Choose the three things the business argues about most. Customer, product, and whatever the third one is in your world.
Write one definition each
A paragraph per entity saying what counts and what does not, with the edge cases that actually come up. Plain language, not a data model.
Name the system of record
For each entity, one system wins. Everything else is a copy, and copies do not get to disagree.
Make the disagreement visible
Count how often the systems differ, and publish the number. What gets counted stops being an argument and starts being a backlog.
The usual mistake
Launching a master data programme. You need three definitions, not a governance function. Programmes of that size die before they decide anything.
- Systems disagree about the same thing. Two systems hold different versions of the same customer, product or number, and no definition is authoritative. The AI inherits the disagreement and gets blamed for it.