Fluree AI is here — the hosted Fluree platform, built on FlureeDB. Get started free
Entity resolution, explained.
The working guide to entity resolution — what it is, how matching and merging actually work, why it decides whether AI answers can be trusted, and what to look for when you evaluate entity resolution software.
What is entity resolution?
Entity resolution is the process of determining when records in different systems describe the same real-world entity — the same customer, supplier, or product — and merging them into one accurate, unified profile.
The problem it solves is everywhere. Customer in Salesforce, client in the ERP, and account in billing may all describe the same organization, but no system understands that. Entity resolution software compares attributes across sources, scores how likely two entries describe the same thing, and consolidates confident pairs while preserving the evidence behind every decision.
Done well, the output is a golden record — a single governed profile per entity, with lineage back to every source. That’s the job of Fluree Sense, the entity resolution engine of the Fluree platform.
How does entity resolution work?
Entity resolution compares candidate records across data sources, scores how likely each pair describes the same real-world entity, and merges confident matches into a single profile — with lineage preserved back to every source.
Two techniques carry the load. Deterministic rules handle exact identifiers — tax IDs, email addresses, account numbers — while machine learning models, trained with input from your subject-matter experts, handle the messy rest: nicknames, abbreviations, transposed fields, and format drift. Before any comparison happens, values are normalized to shared vocabularies so sources can actually agree on what they’re saying.
The best implementations keep humans in the loop. In Fluree Sense, high-confidence matches promote automatically while ambiguous ones route to data stewards for review — and continuous change data capture re-resolves entities as source systems change.
Why does entity resolution matter for AI?
Because AI systems consume identity literally — when the same customer exists five times under five names, a model returns five conflicting answers with equal confidence.
Dashboards tolerated duplicates because analysts reconciled them by hand. Models and agents interpret nothing. In Fluree’s April 2024 study, GraphRAG for GenAI Accuracy, the same questions answered over raw relational data started near 20% zero-shot accuracy — while semantically modeled knowledge graphs started at 60–65% and reached 90–99% once enriched.
Same models, same questions — resolved, semantically modeled data was the difference. That’s why entity resolution is the first step of most AI-ready data programs.
What are the common challenges of entity resolution?
Real enterprise data is messier than any demo dataset. These are the four places where projects stall.
The hardest parts are inconsistent data quality across sources, the absence of shared keys, ambiguous cases that need human judgment, and keeping resolved entities current as systems change.
Challenge 01
No shared keys across systems
The same organization is account 1188 in the CRM and vendor 4421 in the ERP. Resolution has to work on evidence — names, addresses, relationships — because identifiers that span systems usually don’t exist.
Challenge 02
Inconsistent formats and vocabularies
“USA,” “United States,” and “U.S.” defeat naive comparison. Values need normalizing to controlled vocabularies and canonical identifiers before matching, so sources can actually agree.
Challenge 03
Ambiguity that needs human judgment
No model gets every match right. High-confidence decisions should promote automatically while ambiguous ones route to data stewards for adjudication — the workflow Fluree Sense builds in — before they touch governed data.
Challenge 04
Decay without synchronization
A one-time cleanup starts rotting the day it finishes. Continuous change data capture re-resolves entities as sources change, so accuracy holds over time.
What’s the difference between entity resolution, identity resolution, and deduplication?
Three overlapping terms, three different jobs. Here’s how to keep them straight.
Deduplication removes duplicate rows within a single dataset; identity resolution links the identifiers one person uses across channels, mostly for marketing; entity resolution matches and merges profiles for any entity type — people, organizations, products — across many systems.
Deduplication is the narrowest of the three: it finds exact and near-exact copies inside one table and collapses them. It doesn’t reconcile conflicting attributes, and it doesn’t work across sources.
Identity resolution is a marketing-centric specialization — stitching cookies, device IDs, and email addresses into one person for targeting and measurement.
Entity resolution is the general, enterprise-grade case: every entity type across every source, conflicting attributes reconciled, provenance preserved for audit — producing the governed master data the narrower techniques can’t.
How do you choose entity resolution software?
The category is crowded and every demo looks polished. Evaluate against how you’ll actually deploy — your sources, your governance, your timeline.
Choose based on six things: connectors for your actual sources, machine learning with human review, provenance you can audit, output your models can reason over, continuous synchronization, and time to first result.
Connectors for the systems you actually run
Onboarding a new data source should be configuration, not a services engagement.
Machine learning plus rules, with stewards in the loop
Pure rules break on messy inputs; pure ML can’t explain itself. You want both, plus built-in adjudication workflows.
Provenance and lineage by default
Every merged profile should trace back to its sources — that’s what makes results defensible in an audit.
Output your models can reason over
Flat mastered tables lose relationships. Resolved entities in a connected, semantic format keep context intact for retrieval and agents.
Continuous sync, not one-time cleanup
Ask how updates flow after go-live. Batch re-runs mean the golden copy is always slightly stale.
Weeks to value, not quarters
Ask for a demo on a sample of your data, not the vendor’s. Time to first pipeline is the honest benchmark.
One entity, one answer — everywhere.
See entity resolution run on a sample of your own data. Fluree Sense matches, merges, and keeps golden records continuously in sync — in weeks, not quarters.