Entity resolution is the work of deciding that "ACME Logistics Ltd." in your CRM, "Acme Logistics Limited" on an invoice and "ACME LOGISTICS HOLDINGS" in a KYC file are, or are not, the same legal entity. Inside one country and one alphabet it is hard. Across 200+ jurisdictions, with different identifiers, scripts, legal forms and registers, it decides whether your compliance checks, credit limits and reporting rest on the right company. This guide explains how company entity resolution works, why it breaks across borders, and how to build it on identifiers that can be checked, with dated, sourced figures throughout.
01What Is Entity Resolution?
Entity resolution (also called record linkage or entity matching) is the process of identifying which records, across one or many systems, refer to the same real-world entity, and linking them to a single identity. For companies, the real-world entity is a legal entity: a specific registered company in a specific jurisdiction, with a registration number issued by its register.
Entity resolution vs. deduplication
Deduplication removes repeated rows inside one dataset. Entity resolution goes further: it links records across datasets and anchors them to an external identity. Two CRM rows can be deduplicated against each other and still both point at the wrong company. Resolution against a register is what tells you which company that is.
Why the legal entity is the right anchor
Brands, trading names and websites change. A legal entity has a registration number, a jurisdiction, a status, officers and owners, all recorded by a public authority. That makes it the unit that contracts, sanctions, credit limits and regulatory reports actually attach to. Our guide to company registration number formats shows how those numbers differ by country.
02What Bad Entity Matching Actually Costs
Matching errors come in two kinds, and they cost different things. A false negative leaves two records for the same company unlinked: exposure is understated, a sanctions hit on one record never reaches the other, and the same supplier is onboarded twice. A false positive merges two different companies: one company's insolvency or adverse filing is attributed to the wrong counterparty, and decisions get made on someone else's data.
The manual alternative is slow. Fenergo's 2023 study of more than 1,100 executives at corporate and institutional banks found a corporate KYC review took 95 days on average and cost $2,598, as reported by Corporate Compliance Insights. A large share of that time goes on establishing exactly which entity is being reviewed. Data work in general carries the same weight: in Anaconda's 2022 State of Data Science survey, data preparation and cleansing took 38% of data scientists' time.
The register keeps moving
Entity resolution is not a one-off clean-up, because the set of companies it resolves against changes constantly. In the UK, Companies House recorded 801,864 incorporations and 726,735 dissolutions in the year to March 2025, against a total register of 5,427,787 companies, according to its official statistics. In the US, the Census Bureau counted 497,046 business applications in December 2025 alone, seasonally adjusted. Insolvencies add to the churn: Allianz Trade expects global business insolvencies to rise 5% in 2026 after 6% in 2025, according to its October 2026 Insolvency Report.
03Why Entity Matching Breaks Across Borders
Inside one country, names, addresses and identifiers follow one set of rules. Across 200+ jurisdictions, almost every field behaves differently.
| Problem | Example | What fixes it |
|---|---|---|
| Legal-form noise | Ltd / Limited, GmbH, S.à r.l., Sp. z o.o., K.K. | Normalise legal forms per jurisdiction, then compare the core name and the form separately |
| Non-Latin scripts | Japanese, Chinese, Cyrillic or Arabic registered names | Match on the original script and on transliterations; keep both |
| Name collisions | The same company name registered in several countries | Never match on name alone; require jurisdiction plus at least one identifier or address |
| Different identifier systems | UK company number, French SIREN, German HRB plus court, US state file number | Store identifier type and issuer with every value |
| Groups and branches | Parent, subsidiary and branch with near-identical names | Resolve each legal entity, then link them through ownership filings |
| Status changes | Dissolved company whose name is reused by a new one | Match against status and history, not a current-name snapshot |
Table 1. Why company matching fails across borders, and the fix for each problem.
Fuzzy name matching is necessary but not sufficient
Fuzzy name matching (edit distance, token similarity, phonetic codes) catches typos and word-order changes. It cannot tell two genuinely different companies with similar names apart, and across countries similar names are common. Name similarity is evidence, not a decision.
Non-Latin scripts and transliteration
A Japanese, Chinese, Russian or Arabic company is registered under its name in the original script. Your CRM probably holds a romanised version, and there is often more than one way to romanise the same name. A matcher that only compares Latin strings will miss real matches and create false ones. The safer approach is to keep the registered name in its original script, generate standardised transliterations, compare against both, and rely on the registration number or address to confirm.
04Identifiers: Which Ones to Trust
The fastest and safest way to resolve a company is to match on an identifier that a public authority issued. The problem is coverage: the best-known global identifier covers only a small share of companies.
The legal entity identifier (LEI) covers few companies
The LEI is a 20-character code issued under the Global LEI System. GLEIF reported that the active LEI population passed 3.1 million at the end of Q2 2026, with 92,000 organisations obtaining one during the quarter. That is fewer than the 5.43 million companies on the UK register alone. LEIs are excellent where they exist, mostly for entities active in financial markets, but most suppliers, customers and counterparties will never have one.
Registry numbers do the heavy lifting
Every company that exists legally has a number from the register that created it. The catch is that these numbers are only unique within their register, so a registration number must always be stored with its jurisdiction and issuer. GLEIF's Registration Authorities List catalogues almost 1,200 business registers and other registration sources in more than 230 jurisdictions (version 1.9, September 2026), a good sense of how many separate number systems a global matcher has to handle. Europe is adding a common wrapper: since 1 December 2025, French Kbis extracts carry the European Unique Identifier (EUID), built from the country code, the register code and the company's SIREN, according to Service-Public.fr.
| Field | Confidence | How to use it |
|---|---|---|
| Registry number + jurisdiction | Deterministic | Exact match resolves the entity |
| LEI / EUID | Deterministic | Exact match where present; maps to the registry number |
| VAT or tax ID | High | Validate format per country, then match; watch VAT groups and branches |
| Registered address | Supporting | Normalise and geocode; many companies can share one address |
| Officers and owners | Supporting | Strong tie-breaker when names collide |
| Domain, phone | Supporting | Useful for scoring; often points to a group, not one entity |
| Name only | Not enough | Never auto-merge on name alone |
Table 2. Matching fields by confidence. Validate VAT numbers with our VAT format checker.
05Deterministic vs. Probabilistic Matching
Every serious entity resolution system combines two approaches. Deterministic matching applies exact rules: if the registration number and jurisdiction agree, it is the same entity. Probabilistic matching scores how likely two records are to match from several partial signals, such as name similarity, address, officers and domain, and compares the score with thresholds.
| Deterministic | Probabilistic | |
|---|---|---|
| Works when | A trusted identifier is present on both records | Identifiers are missing or inconsistent |
| Precision | Very high, if the identifier is validated | Depends on thresholds and field quality |
| Recall | Low on messy data: no identifier, no match | Higher: finds matches rules miss |
| Explainability | Simple: "same company number" | Needs the agreeing attributes shown with the score |
| Main risk | Typos in identifiers, reused or reissued numbers | Confident false positives on similar names |
Table 3. Deterministic vs. probabilistic matching for company records.
The practical design is layered: try deterministic matches first, score the rest probabilistically, auto-merge only above a high threshold, and send the middle band to a human with the evidence attached. That review step is where ambiguous cases like parents and subsidiaries get resolved correctly.
How a match score is built
A probabilistic score is only as useful as its explanation. Each attribute is compared with the method that suits it: exact comparison for identifiers, token and edit-distance similarity for names after legal forms are stripped, normalised comparison for addresses, and set overlap for officers. Each comparison contributes evidence for or against the match. The output should show the reviewer which attributes agreed, which disagreed, and which candidate came second.
Setting thresholds by use case
There is no single correct threshold. A sanctions or KYB workflow should prefer sending a borderline case to review over missing a true match, so its auto-merge threshold sits high and its review band is wide. A marketing list can tolerate more automatic merges. Set thresholds per workflow, measure precision and recall on a labelled sample of your own records, and revisit them when the input data changes.
06Building a Scalable Entity Resolution Workflow
Five steps, each one inspectable
- Standardise: strip and normalise legal forms, transliterate non-Latin names while keeping the original, parse addresses, and validate identifier formats before using them.
- Block: comparing every record with every candidate is too slow at scale, so shortlist candidates first by country, identifier, name tokens or postcode.
- Compare and score: compare each attribute with the method suited to it, combine the results into a score, and keep the agreeing and disagreeing attributes alongside it.
- Decide: auto-merge above a high threshold, review the middle band, report no match below it. Store every review decision so the same ambiguity is not reviewed twice.
- Monitor: resolved entities change status, names, officers and owners. Watch the register and re-resolve when a material filing appears, with portfolio monitoring.
Resolve the entity, then link the group
Deduplication solves the row; linkage solves the roll-up. Once each record points at the right legal entity, ownership filings connect subsidiaries to parents, so exposure, spend or risk can be totalled at group level. Under OFAC's 50 Percent Rule, set out in guidance of 13 August 2014, an entity owned 50% or more in aggregate by blocked persons is itself blocked, which is why sanctions teams need resolved ownership, not just resolved names. See corporate linkage and UBO discovery.
07A Worked Example
A procurement team loads 40,000 supplier records from three ERPs. One supplier appears three times, as in the illustration at the top of this guide.
- Standardise: "Ltd." and "Limited" are normalised; the ERP record's VAT number is validated against the UK format.
- Block: candidates are shortlisted from UK registry profiles sharing the core name tokens "ACME LOGISTICS".
- Score: the CRM and ERP records agree with one registered company on address, domain and officer, and the ERP's VAT number links to it. Both are auto-merged.
- Review: the KYC record says "HOLDINGS" and supplies no number. The name is closer to the parent company, and the director appears on both. It goes to review, where an analyst confirms the onboarding was done on the parent, not the trading subsidiary, and corrects it.
- Link and monitor: the subsidiary and parent are linked through the shareholder filing, and both are monitored for status, officer and ownership changes.
Without the review band, the KYC file would have been silently merged into the wrong entity, or left as a fourth "supplier".
08How to Evaluate Entity Resolution Software and Providers
Engine only, or engine plus reference data
Entity resolution tools fall into two broad groups: matching engines you run on your own reference data, and providers that also supply the reference data, the registry records you resolve against. If you only buy the engine, the quality of the result is capped by the reference data you feed it.
| Area | Ask | Good answer |
|---|---|---|
| Reference data | What do you resolve against? | Records collected from official registries, with source and retrieval date per field |
| Identifiers | Which IDs come back? | Open registry numbers, LEIs and tax IDs you can join to your data, not only a proprietary key |
| Explainability | Why did this match? | Score plus agreeing attributes plus the runner-up candidate |
| Ambiguity | What happens in the middle band? | A review queue with evidence, and decisions retained across refreshes |
| Coverage | Which countries and scripts? | Per-country coverage figures, including non-Latin scripts |
| Freshness | How current is the reference? | Stated refresh cadence and change events pushed to you |
| Match rate | What will we get on our data? | A test on a sample of your own records, not a headline percentage |
Table 4. Questions to ask when evaluating entity resolution software or an entity resolution API.
Test on your own records
Headline match rates depend heavily on the input. A file full of registration numbers and postcodes resolves far better than a list of trading names. Before choosing a provider, send a sample of a few thousand of your own records, including your messiest ones, and compare auto-matches, review-band cases and no-matches. Check a random sample of auto-matches by hand: a high match rate with confident false positives is worse than a lower rate that sends ambiguity to review.
09How Global Database Resolves Entities
Global Database's master data management service resolves customer, supplier and prospect records against 600M+ registry profiles collected from 400+ official registries in 200+ countries. Records are normalised and blocked on name, address, domain, officer names and any identifier supplied. Each candidate is scored attribute by attribute, and the agreeing attributes and the runner-up are returned with the score. Resolution keys on public identifiers, registration numbers and LEIs. Above your auto-merge line records merge; below it they go to a review queue with evidence, and decisions are retained. Group structure comes from ownership filings, registries are read daily, and changes are pushed as events through the API, bulk feeds, CRM connectors for Salesforce, HubSpot and Microsoft Dynamics, or an MCP server. Resolved records can then feed KYB verification and screening.
10Frequently Asked Questions
What is entity resolution in simple terms?
Entity resolution is deciding which records refer to the same real-world thing and linking them to one identity. For companies, that means working out that differently spelled records in your CRM, ERP and KYC files belong to the same registered legal entity, identified by its registration number and jurisdiction.
What is the difference between entity resolution and deduplication?
Deduplication removes repeated rows within one dataset. Entity resolution links records across datasets and anchors them to an external identity, such as a company register entry. Two rows can be deduplicated against each other and still point at the wrong company; resolution against the register tells you which company it is.
Why does entity resolution matter for compliance and KYB?
Sanctions screening, KYB checks and credit decisions attach to a specific legal entity. If a record is matched to the wrong company, a check is run on the wrong entity; if two records for the same company stay unlinked, a hit on one never reaches the other. Under OFAC's 50 Percent Rule, ownership also has to be resolved, not just names.
What is a legal entity identifier (LEI)?
The LEI is a 20-character code issued under the Global LEI System and overseen by GLEIF. It is reliable where it exists, but coverage is limited: GLEIF reported just over 3.1 million active LEIs at the end of Q2 2026, fewer than the 5.43 million companies on the UK register alone. Most companies are resolved by their registry number instead.
What is the difference between deterministic and probabilistic matching?
Deterministic matching uses exact rules, for example the same registration number and jurisdiction means the same company. Probabilistic matching combines several partial signals, such as name similarity, address, officers and domain, into a score and compares it with thresholds. Good systems apply deterministic rules first and score the rest.
Is fuzzy name matching enough to match companies?
No. Fuzzy name matching catches typos and word-order changes, but it cannot tell apart different companies with similar names, which are common across countries. Name similarity should be one input to a score that also uses jurisdiction, identifiers, addresses and officers, and a name-only match should never be auto-merged.
How do you match company names in non-Latin scripts?
By keeping the registered name in its original script and matching on both the original and standardised transliterations, then confirming with identifiers or addresses. Transliteration alone is ambiguous, because the same name can be romanised in several ways.
Can entity resolution be automated for KYB?
Largely, yes. Records with a validated registry number, LEI or tax ID can be resolved automatically, and high-scoring probabilistic matches can be auto-merged. The middle band, such as parents and subsidiaries with near-identical names, should go to human review with the evidence attached, and those decisions should be stored.
How often should resolved entities be refreshed?
Continuously, if possible. In the UK alone, 801,864 companies were incorporated and 726,735 dissolved in the year to March 2025. Names, statuses, officers and owners change, so resolved records should be monitored against the register and re-resolved when a material filing appears.