Legacy mail, files and databases rarely agree. A trustworthy ontology preserves provenance and uncertainty before it tries to produce a clean customer or operational view.
Do not collapse a source claim into a business fact until the evidence supports it.
Begin with questions the business needs to answer
An ontology is not a catalogue of every field in every old system. Start with operational questions: who is this customer, what relationship exists, what happened, which record supports it and how confident are we?
Those questions reveal the durable concepts—party, account, case, communication, document, transaction, asset and event—and the relationships that connect them.
Keep source records immutable
Ingest the original record with its source, location, timestamp and content hash. Parse and normalise into a separate layer. This creates an evidence chain from each derived assertion back to what the organisation actually received or recorded.
- A source record says what one system claimed at one time.
- An assertion expresses a normalised claim with provenance and confidence.
- An entity groups assertions believed to refer to the same real-world thing.
- A resolution records how a conflict was reviewed and decided.
Make identity resolution explainable
Names, email addresses, phone numbers and company identifiers change. Use deterministic matches for strong identifiers, weighted candidates for weaker signals and a review queue for ambiguous cases.
Never let a similarity score silently merge two customer histories. Store why records were linked, which rule or reviewer made the decision and how to reverse it.
Publish a usable view without hiding uncertainty
Downstream CRM, search and agent tools need a simple view, but simplicity should be produced from traceable layers. Expose the selected value, its source and any unresolved alternatives when that uncertainty matters to the decision.
The acceptance test is not a perfectly tidy graph. It is whether a person or system can answer an operational question, see the supporting evidence and correct the result without damaging history.
