Skip to content

AI Glossary · Concepts

What is Entity Resolution?

Entity Resolution

Entity resolution links records that refer to the same real-world person, company or object when their names and identifiers differ.

Also known as: record linkage, entity matching, fuzzy matching

· Updated · Chain of Thought

AI AgentsContext Management

  • 1 episode
  • 2 sources

Matching records is a data problem with or without an LLM. Splink’s official documentation describes probabilistic record linkage that deduplicates and links records without unique identifiers (Splink documentation). An exact name match is one signal, not a universal proof of identity. Two records may share a name and represent different companies, or use different names for the same company.

For example, a support tool might contain “Airbyte Inc.” while a CRM contains “Airbyte.” An agent could propose a match using domain names, addresses and other metadata. Keep the evidence and uncertainty with the proposed match. Test both likely matches and difficult non-matches, then require review before a low-confidence match changes billing, access or another consequential record.

Check which pairs reach the matcher

Splink uses blocking to choose which record pairs to compare at all: it produces every record comparison that satisfies at least one blocking rule, and only those pairs are scored. Looser rules capture more true matches at the cost of more comparisons. A rule requiring the same first name and surname, Splink’s documentation notes, “would miss a true match where there was a typo in the name” (Splink, What are blocking rules?). A pair excluded at this step never reaches the model, so lowering the acceptance threshold later cannot recover it.

Suppose the source record is “Northwind Labs” with domain northwind.example and registry ID R17, and registry IDs are verified and unique:

CandidateEvidenceDecision
A: Northwind LabSame domain, registry ID R17Match
B: Northwind LabsDifferent domain, registry ID R88Different entity
C: Northwind LabsNo domain or registry IDUnresolved

An exact-name blocking rule admits B and C and drops A, the one real match. The identical name in B does not outweigh a different registry ID, and C needs more evidence before anyone decides. Test candidate generation against known matches, and record the blocking rules, comparison fields and acceptance policy separately so a missed match can be traced to the step that lost it.

Describing how his agent correlates records across support and sales tools, Michel Tricot says, “it’s able to resolve the organization in zendesk into what do they look like on gong” (episode 52, 23:05).

That is his on-show example of a workflow, not evidence that every matching problem is solved. For builders, distinguish finding a candidate from accepting it as the same entity. Upstream, named-entity recognition finds mentions such as company names in text; entity resolution decides which mentions and records refer to the same thing. Knowledge graph covers relationships among entities, context management covers what an agent sees, and what is context in AI agents? explains how metadata supports a task.

Sources

  • Splink documentation — Primary source for the explanation and distinctions in this entry.
  • Splink: What are blocking rules? — Explains how blocking rules choose which record pairs are compared, and the tradeoff between comparison volume and capturing true matches. Checked November 2, 2026.

From the conversation