Entity resolution flow
01

Normalize: name, address, unit, code

02

Candidate: deterministic and similarity rules

03

Decide: match, no match, review

04

Golden record: source and precedence

05

Govern: version, split, downstream notice

False merges can be worse than missed matches

Merging different customers mixes transactions, access, and privacy; a missed match mainly duplicates counts. Set thresholds and review by risk.

Define business identity before evidence. A same name, similar address, or product title is insufficient alone.

Normalize while preserving originals

Standardize case, spaces, regions, units, suffixes, and phone formats while retaining source values. Validate supposedly deterministic IDs.

Version normalization rules so historical decisions remain explainable.

Separate candidates from decisions

Use blocking keys and similarity to retrieve candidates, then decide with multiple fields. Output score, evidence, and conflicts without overwriting sources.

Keep human review for high-risk entities and unresolved cases.

Track golden-record provenance

CRM may own name, recent verification may own address, and master catalog may own specification. Record source, time, and precedence per field.

Merge decisions must support split and rollback and trigger dependent recomputation.

Accept with real dirty data

Test same-name people, renames, old addresses, shared phones, bundles, multiple barcodes, and bad codes. Measure false merge, missed match, review volume, and rollback.

BuildTable can be evaluated for AI-friendly master data; algorithms, review workflow, golden records, and writeback need confirmation.

Public references

Build an AI-ready data foundation

Download BuildTable or talk with us about your data modeling scenario.

Download BuildTable