Entity Resolution: Techniques, Tools & Enterprise Use Cases

Updated October 2026

Choosing an enterprise entity resolution tool starts with the deployment model, not the vendor. Decide whether you need an MDM platform with stewardship, an embedded engine for real-time matching, a pay-per-record cloud service, or an open-source library your engineers run. Then test two or three finalists on your own labeled records, measuring precision, recall, explainability, and cost per rerun.

How should an enterprise choose an entity resolution tool?

Start from where the resolved entity will be used and who will fix its mistakes. A golden customer record that syncs to ERP and CRM needs survivorship and stewardship, which points to an MDM platform. A sanctions check during onboarding needs a sub-second answer, which points to an embedded engine or a real-time API.

This guide covers the enterprise buying decision. For the techniques themselves, such as blocking, Fellegi-Sunter scoring, and LLM matching, see what entity resolution is and how it works. For golden records, ID stability, and stewardship design after matching, see identity resolution and entity mastering.

Five questions narrow the field quickly. Where does the data live today? How fresh must the answer be? Who reviews uncertain matches? How much must each match be explained to an auditor? And who will own the match rules five years from now?

Which questions decide the deployment model?

Latency, review workflow, and data location usually decide it before any feature list does. Each answer removes whole categories of tools, which keeps the proof of concept small.

Freshness and incremental matching

Batch matching is cheaper and easier to govern, but many enterprises also need answers mid-workflow. Check incremental support carefully. AWS Entity Resolution offers automatic incremental processing only for rule-based matching, while manual bulk runs reprocess all data.

Stewardship and explainability

Someone has to resolve the uncertain middle band of pairs. MDM platforms ship review queues and match analyzers for business users. Libraries return scores and leave the review screen to you, which works when the reviewers are engineers.

Which entity resolution tools are enterprises evaluating in 2026?

The shortlist has shifted, mostly through acquisitions and new AI features. SAP now owns Reltio, Salesforce owns Informatica, and most vendors now pitch resolved entities as context for AI agents. Each tool below gets one documented strength and one limit.

Reltio

SAP announced on March 27, 2026 that it would acquire Reltio and make it a core capability of SAP Business Data Cloud. Reltio says the deal closed on May 7, 2026 and that it remains available standalone. Its entity resolution page describes pretrained ML matching, fuzzy rules, automatic unmerge, and persistent IDs. The limit for non-SAP shops is roadmap risk, since investment now follows SAP's priorities.

Informatica MDM

Informatica is now branded Informatica from Salesforce, after Salesforce completed the acquisition on November 18, 2025. Its MDM page promotes self-tuning AI match and merge with rule recommendations and explainability. It fits enterprises that already run Informatica integration and governance. The tradeoff is adopting a broad suite when matching may be the only gap.

Tamr

Tamr calls itself an AI-native MDM platform, with entity resolution that draws on built-in external data, an enterprise knowledge graph, and MCP connectivity for LLMs. Its pricing page says fees scale with golden records produced, users are unlimited, and implementation is included. Actual prices need a quote, though Tamr offers a guided proof of concept on your data.

Senzing

Senzing is an embedded SDK that resolves records in real time without per-source model training. Its pricing page lists 58,560 dollars a year for 10 million data source records and says it has run in production near 100 billion records. Senzing also offers an MCP server that answers mapping and SDK questions, though it does not call a live engine. You build the stewardship screens and survivorship logic yourself.

Quantexa

Quantexa rebuilds entities from raw data at the time of each request, with match confidence set per use case. That suits fraud, AML, and investigations where different teams accept different risk. The page claims 99 percent resolution accuracy but gives no method or baseline, so treat it as a vendor figure.

AWS Entity Resolution

AWS Entity Resolution offers rule-based matching, a preconfigured ML model aimed at consumer records, and matching through licensed data providers. Inputs come from AWS Glue and results land in S3. Pricing is 0.25 dollars per 1,000 records for rules and ML, with no free tier and every processed record billed.

Google Entity Reconciliation API

Google's Entity Reconciliation API clusters records from BigQuery and assigns a stable Machine ID to each entity. It natively supports up to 10 BigQuery tables. It is still in Preview, so support terms are limited and it is a poor fit for regulated production use today.

Splink and Zingg

Splink is an MIT-licensed Python library that released version 5.0.0 on September 28, 2026. It runs on DuckDB, Spark, and PostgreSQL, and the project says Spark handles over 100 million records. Zingg learns blocking and similarity models on Spark and reads and writes Snowflake, under AGPL 3.0. Both keep match logic in code you own, but neither ships stewardship tooling.

Neo4j Graph Data Science

Neo4j describes resolving entities with Cypher rules, Node Similarity, Weakly Connected Components, and link prediction over shared identifier nodes. This works well when fraud rings or supplier networks share phones, addresses, or bank details. Neo4j itself notes that Node Similarity is computationally intensive, so most teams pre-score records first.

How do you run an entity resolution proof of concept?

Run every finalist on the same labeled sample of your own data. Vendor demos use clean records, and your records are not clean. A two to four week bake-off on a fixed sample tells you more than any analyst report.

1. Pick a hard sample. Take a few hundred thousand records from at least three sources, including the messiest one.

2. Label pairs. Have people mark a few thousand candidate pairs as match or non-match. Splink's evaluation guide treats clerical labels as the basis for measuring quality.

3. Score at several thresholds. Compare precision and recall across thresholds rather than at one vendor-chosen setting.

4. Check clusters. Look for oversized clusters and chains where one bad link merges two real entities.

5. Test change. Add new records, edit an address, and force an unmerge. See what each tool does to existing IDs.

6. Ask for explanations. Pick ten merges at random and ask each tool to show why they matched.

Do not pick a winner on one number. Splink's edge metrics guide warns that F-score ignores how well non-matches are predicted. It also notes that accuracy can look high on imbalanced data, which describes almost every matching job.

Weigh false merges more heavily

A missed match leaves a duplicate. A false merge mixes two customers' histories, which can leak data or misstate risk. Set your threshold by the cost of each error in your domain, not by the vendor default.

How do you estimate entity resolution cost?

Model cost per rerun, not cost per license. Matching is never a one-time job. Sources change daily, rules get tuned, and every tuning cycle may mean another full run.

The AWS figure comes from its pricing page, and the Senzing figure from Senzing pricing. AWS ML matching lacks automatic incremental processing, so frequent full reruns multiply that per-run number. Open-source tools shift spend from licenses to people, which suits teams that already run Spark or DuckDB.

Where does entity resolution sit in an enterprise data architecture?

It sits between raw source data and every system that needs one ID per real thing. Matching produces candidate links, mastering turns them into golden records, and an ID registry serves stable IDs to BI, search, and AI agents. Skipping the registry means every team rebuilds matching on its own.

Resolved entities also feed knowledge graphs and semantic layers. A graph with duplicate nodes answers confidently and wrongly, which is why knowledge graph platforms and MDM tools both ship matching features. For how MDM and a broader semantic model divide the work, see semantic data unification vs MDM.

Who should own match rules and resolved IDs?

The company should own them, whichever tool does the matching. Match rules, thresholds, labeled pairs, and stable IDs are institutional knowledge about customers, suppliers, and products. If they live only inside a vendor's schema, switching tools means relearning years of decisions.

Ask every vendor how you export rules, labels, match history, and the ID crosswalk in open formats. Keep the labeled sample from your proof of concept, since it lets you test any future tool against the same truth.

Galaxy is an applied lab that builds this layer with teams on the stack they already run, using internal tooling to stand up entity resolution, ontologies, and semantic models quickly. It builds in the open with Filament and plans to open source more of that tooling.

Frequently asked questions

What is the best entity resolution tool for an enterprise?

No single tool wins everywhere. MDM platforms such as Reltio, Informatica, and Tamr fit teams that need golden records and stewardship. Senzing fits real-time matching inside your own apps. AWS Entity Resolution fits batch jobs on AWS. Splink and Zingg fit teams with data engineers who want code they control. Pick by deployment model first, then vendor.

How much do entity resolution tools cost?

Pricing models differ more than prices. AWS Entity Resolution charges 0.25 dollars per 1,000 records for rule and ML matching. Senzing lists 58,560 dollars a year for 10 million data source records. Tamr prices by golden records produced, and Reltio and Informatica quote privately. Splink and Zingg are free to use but need engineering time.

Did SAP buy Reltio?

Yes. SAP announced an agreement to acquire Reltio on March 27, 2026, and Reltio says the deal closed on May 7, 2026. SAP plans to make Reltio a core part of SAP Business Data Cloud. Both companies say Reltio stays available as a standalone product, so non-SAP customers can still buy it.

How do you run an entity resolution proof of concept?

Pull a sample of a few hundred thousand real records from at least three sources. Have people label a few thousand candidate pairs as match or non-match. Run every shortlisted tool on the same sample and compare precision, recall, and cluster quality at several thresholds. Then test explainability, incremental updates, and unmerge behavior.

Which metrics should I use to compare entity resolution tools?

Use precision and recall on labeled pairs, and do not rely on one score. Splink's documentation warns that F-score ignores how well non-matches are predicted and that accuracy looks good on imbalanced data. Add cluster checks such as oversized clusters, and weigh the business cost of a false merge against a missed match.

Should entity resolution run in batch or real time?

Run batch when the output feeds analytics, reporting, or a nightly golden record. Run real time when a decision happens during a workflow, such as onboarding, sanctions screening, or an agent lookup. Many enterprises need both. Check whether a tool supports incremental matching, since AWS offers automatic incremental processing only for rule-based matching.

Is open-source entity resolution good enough for enterprises?

Often yes, if you have engineers to run it. Splink is MIT licensed, released version 5.0.0 in September 2026, and runs on DuckDB, Spark, and PostgreSQL. Zingg learns its own blocking on Spark under AGPL 3.0. What you give up is stewardship screens, survivorship tooling, and vendor support, which you must build or buy separately.

What is the difference between an entity resolution tool and an MDM platform?

An entity resolution tool decides which records are the same thing. An MDM platform wraps that decision with survivorship rules, stewardship queues, hierarchy management, and sync back to source systems. Reltio, Informatica, and Tamr are MDM platforms with matching inside. Senzing, Splink, and AWS Entity Resolution focus on the matching itself.

Can a graph database do entity resolution?

It can handle part of it. Neo4j describes using Cypher for simple shared identifier rules, Node Similarity and Weakly Connected Components for larger sets, and link prediction for harder cases. Neo4j notes Node Similarity is computationally intensive. Most teams still clean and score records before loading resolved entities into a graph.

Who should own match rules and resolved entity IDs?

The company should own them. Match rules, thresholds, labels, and stable IDs encode what the business knows about its customers and suppliers. Keep them in open formats on infrastructure you control, with export paths from any vendor. That lets you audit decisions, switch tools later, and reuse the same entities in analytics and AI agents.