The rules exist, but only in people’s heads
Filling a master-data record correctly follows real rules. The problem: those rules were never written down. They live in years of history and in a handful of experts — impossible to read, audit or hand over at scale.
Complex and contextual
A single field can follow hundreds of rules depending on country, customer type and other attributes. What is right in Market A is wrong in Market B.
Scattered and tacit
Knowledge sits in spreadsheets, emails and a few specialists’ memory. New team members simply do not know what to fill in, or with which value.
Migration risk
Without a reliable picture of how things really work today, the new system is designed on assumptions — leading to rework, dirty data and cutover delays.
Four master-data domains
The method was proven in the Customer domain and is being extended, one pillar at a time.
Customer
Who you sell to
Live in production
Vendor
Who you buy from
Pipeline built
Finance
Accounts and cost centres
On roadmap
Material
What you make and move
On roadmap
From raw history to knowledge anyone can ask for
Two complementary moves: an engine that discovers the rules from what was actually recorded, and an assistant that delivers them to whoever needs them, in plain language.
Discover · the rule-mining engine
Instead of interviewing experts, the engine reads the real registration history and reverse-engineers each rule, with a statistical confidence and how many real records back it. Tens of thousands of rules, recovered automatically.
Deliver · the knowledge graph and assistant
The rules become a connected graph an assistant can reason over. Ask a normal question and it answers in business or technical language, always grounded in real rules, never invented.
Discover, then deliver: the left half is data engineering; the right half is how every team finally gets to use the result.
See one rule come to life
Every rule has the same simple shape: a few inputs, like country and customer type, trigger a rule, and the rule determines the right value. Drag the nodes, hover one to focus it, scroll to zoom — this is literally how the assistant reasons.
- Every rule reverse-engineered from what was actually recorded
- Inputs, trigger and result — the same shape every time
- Multiply this by tens of thousands of recovered rules
Read it left to right: three inputs — Market A, a sold-to customer, a southern region — trigger one rule, and that rule determines the correct sales office. Now picture this multiplied by tens of thousands of rules.
Drive the rules yourself
No spreadsheets, no database queries, no need to know the system’s field names. Choose a country and a customer type, and watch the graph rewire and the fields resolve — with how confident it is and how many real records support it.
Synthetic and anonymised throughout — Market A, Entity A-01. The values show the shape of the rules, never a client’s configuration, and the record counts are illustrative.
Two answers unlock almost everything
After analysing every field, one pattern stood out: country plus customer type is the fingerprint of a record. Knowing just those two answers determines 9 in 10 setup fields.
~1,000 real country × customer-type combinations
A two-field fingerprint
Country and customer type alone determine 9 in 10 setup fields. A simple, intuitive starting point for any record.
Local by nature
Almost every value changes by country; only a few are truly global. The graph captures that local nuance automatically, across 80+ markets.
Evidence for every phase of the move
The project gives each migration phase the factual base it needs, replacing assumptions with evidence drawn from real records.
- Phase 1 · As-Is
Understand the legacy
A clear, audited picture of how records were actually filled today, by country and customer type, with confidence and coverage. Surfaces inconsistencies before they become migration problems.
- Phase 2 · Design
Design the target
The explicit rules feed the to-be blueprint: per-country defaults and field dependencies become visible and queryable instead of tribal knowledge.
- Phase 3 · Build and test
Validate and harmonise
The assistant answers fill-in questions in real time and can generate a full record setup in one call, accelerating data cleansing and harmonisation.
- Phase 4 · Cutover and run
A lasting copilot
After go-live it becomes a record-creation copilot: any user consults the rules without depending on a few experts, cutting human error and dirty data for good.
The project at a glance
Proven in production on the Customer domain, with the method ready to extend across the rest. Figures below are banded.
Every enterprise is sitting on rules it cannot read.
Decades of decisions are locked inside legacy systems and a few people’s memory. The pattern proven here — mine the rules from real history, connect them in a knowledge graph, and make them answerable in plain language — turns that institutional memory into a queryable asset. The system changes; the approach does not. It de-risks any major migration and outlives it.