Financials, UBOs and group structure your AI agents can cite.

Company data for AI agents, read from official government registries. Every value carries the register it came from, the filing reference and the time of the read, across 600M+ companies in 200+ countries.

model recall
Is Global Data Intelligence Limited active, who controls it, and what was its last reported turnover?
Global Data Intelligence Limited is an active UK company based in Northampton. It is a privately held business intelligence provider, majority owned by its founder, and its most recent accounts show turnover of around £4.2 million.
0 sourcesNo timestampNo filing referenceCannot be reproduced
Active on the UK register.1 Incorporated 28 January 2015, registered office Artisans' House, 7 Queensbridge, Northampton NN4 7BF.2 Two people with significant control are recorded, both natural persons.3 Turnover is not published: the accounts on file do not include a profit and loss account.4
1Company status · ActiveCompanies House, GBCS01 · 2026-03-06
2Incorporation · 28 Jan 2015Companies House, GBIN01 · 2015-01-28
3PSCs · 2 individualsPSC register, GBPSC · 2026-03-06
4Turnover · not publishedCompanies House, GBAA · 2026-01-31
3 values, 3 filings1 declined, reason givenReproducible on re-query
Every value above is on the public register. Check it yourself:Companies House · 09410808

Same question, same company: our own. The left answer reads well and cannot be checked. The right one declines the field the register does not publish.

the problem with recall

An agent that cannot show its source is a liability with good grammar.

General-purpose models produce a turnover figure for any company you name. It is fluent, plausible and unattached to a filing. Put that in a case file and the first question an examiner asks is the one the agent cannot answer.

This is company data for AI agents built so the agent never has to. Every record an enterprise AI agent retrieves is read from a named government register and returned with the reference a reviewer needs to open the original. 400+ registries, 200+ countries and territories, 600M+ company records, synced daily.

01

Entity resolution

Turn a name, domain, VAT number or registration number into one registry record. Similarly named companies are the first place an agent goes wrong, so the resolved registration number is returned with the match.

02

KYB verification

Legal name, legal form, registration number, incorporation date, registered office, current status and the officers on file. KYB checks read from the statutory register, not inferred from a website.

03

UBO and ownership verification

Shareholders, beneficial ownership data, parents and subsidiaries, hop by hop across jurisdictions. Each hop is a registry lookup and each holding percentage is returned with the filing that recorded it.

04

Read financials

Filed financial statements as structured fields: turnover, profit, assets, liabilities, employees, by period. Where a register publishes documents rather than data, the fields are extracted from the filed document and the document reference stays attached.

05

Cite

Every value in the response carries the register, the filing reference, the retrieval timestamp and a link to the original document where the register publishes one. The agent passes these through to the user or the case file.

06

Log

The response is structured, so the resolved entity, the fields used and the sources can be written to the decision log without parsing prose. Re-running the query reproduces the same values against the same timestamp.

MCP server, Regis, bulk feed and API. One dataset behind all four.

The same registry records sit behind every route. Find the row that sounds like your agent; the route on the right is the one to take.

Which route fits your agent
It runs in Claude, OpenAI, Microsoft Copilot or another MCP clientMCP serverConnectors for Claude, OpenAI and Microsoft Copilot. Connect once and the dataset appears as callable tools: resolve, verify, shareholders, UBOs, group structure, financials. No code in the agent itself.It is a pipeline that already knows the field it needsREST APIFixed schemas, one endpoint per dataset, an include_provenance flag on the request. Token in the header, no SDK.Its input is a question, not a field nameRegis APIOne string in. Entities resolved, records retrieved, a source on every value, a stated gap where the register has none.You index company data into your own retrieval or vector layerBulk feedScheduled, daily-synced extracts with provenance carried per record. Redistribution under licence.It onboards customers, merchants or sellers and needs KYB before approvalREST APIThe KYB verification endpoint returns identity, status and officers from the statutory register, so the approval decision carries a filing, not a guess.It walks ownership chains for AML or sanctions workRegis APIAsk for the full chain in one question. Each hop comes back against the register that recorded it, ready for the case file.It reads financial statements for underwriting decisionsREST APIFiled accounts as structured fields by period, with the document reference attached to the extraction.It verifies suppliers or bidders before a purchase order goes outMCP serverA procurement copilot calls the verification tools directly: registered, active, owned by whom it declares, each answer sourced.It monitors a portfolio for status, ownership or filing changesBulk feedDaily-synced extracts diffed on your side, so every change your agent flags traces to the register that published it.You want to check the data before any of thisData dictionaryEvery field, dataset by dataset, with its definition. No key needed.

Data provenance, walked end to end.

Provenance is only useful if a reviewer can follow it from the sentence the agent wrote to the document the register holds. This is the walk, using our own record.

1Agent answer

The sentence

What the user or the downstream system sees. Each value carries a marker back to the source list.

Incorporated2 28 January 2015. Registered office3 Artisans' House, 7 Queensbridge, Northampton NN4 7BF.
2Field in the response

The record

The structured entry the marker points at: field, value, register, filing reference, time of read.

Source record sources[2]
FieldIncorporation date
Value28 January 2015
RegisterCompanies House, GB
FilingIN01
Retrieved2026-09-01T08:14:02Z
3Filing at the register

The document

The statutory filing itself, at the government register, opened independently of us.

Companies House
GLOBAL DATA INTELLIGENCE LIMITED
Company number 09410808
Open at Companies House

When the register has no field

Some registers publish no financial statements, no shareholders or no beneficial owners for a given entity type. The response says so and names the register. A gap with a stated reason is a defensible entry in a case file. An estimate is not.

Freshness

Registers publish on their own schedules, from continuous to annual. Records are synced daily, and the retrieval timestamp on each value tells the agent, and the reviewer, which side of a register's update the read landed on.

Grounding an LLM and retrieving a record are two different things.

AI grounding adds context to a model that still answers from its weights. Regis is the retrieval-and-answer layer over this dataset. It runs on our own hardware, calls no third-party model API and holds no company facts in its weights. When a question asks for a field the register does not publish, it returns the gap and the reason. It does not fill it.

01
Why it matters for agents

An agent on a general-purpose model will produce a turnover figure for any company you name. Regis returns nothing it did not read from a register on that request, so the agent's output inherits the same constraint.

02
How it fits

Call POST /v1/regis/query with a question, or expose it through the MCP server as a tool for multi-step questions such as a full ownership chain. Structured endpoints remain available alongside it.

03
What comes back

The written answer, the resolved entities, the fields used and the source records, in JSON. Written straight into a case management system without parsing prose.

Every major provider now has an MCP server. The difference is where the data came from.

Dun & Bradstreet, Moody's and Sayari all ship agent access and all describe their data as sourced. What separates them is how many hands the filing passed through before it reached your agent, and whether it can still be traced back.

Global DatabaseDirect from registries
Dun & BradstreetLicensed network
Moody'sLicensed providers
SayariMixed collection
Where the data comes from
400+ government registries and tax authorities, connected to individually. No data partners, no web crawling, no self-reported records.
Largely licensed and curated. Tens of thousands of sources: a worldwide partner network, public records, court filings, web crawling and crowd-sourced or self-reported records.
Licensed. 170+ information providers plus Moody's own sources and regulatory sources and published annual reports, standardised into one database.
Mixed. Collects directly from registries, customs bureaus and regulatory agencies across 250+ jurisdictions, alongside trade and shipping records.
Steps between the filing and your agent
One. The register, then us, then you. We read the filing and return the value with its register and filing reference. No partner, reseller or modelled substitute in the chain.
Several. Filing to a Worldwide Network partner or web source, re-indexed into the Data Cloud under a D-U-N-S Number, modelled values added, then to you.
Several. Filing to one of 170+ information providers or Moody's own collection, standardised into Orbis, then to you.
Two. Filing collected and stored as the original document, passed through entity resolution into the graph, then to you. Source document linked.
Provenance depth
Register, filing reference, retrieval timestamp and document link on each value returned.
Record level. D-U-N-S is the anchor rather than the filing.
Blends multiple sources into one record. Filing-level attribution not stated publicly.
Links responses to primary documents with a confidence assertion.
Identifier model
Resolves on registration number, VAT number, name or domain. No proprietary identifier to adopt or licence.
D-U-N-S number, proprietary.
BvD ID, proprietary.
Sayari entity ID, proprietary.
MCP server and connectors
Yes. Remote MCP server, with connectors for Claude, OpenAI and Microsoft Copilot. mcp.globaldatabase.com
Connectors into Microsoft Copilot, OpenAI and IBM watsonx.
Published on Databricks Marketplace and Claude for Financial Services.
World Model MCP.
Redistribution rights
Available. Serve the data to your own customers, inside your product or through your agent, under a redistribution licence with the terms written into the contract.
Restricted. Onward distribution is limited by the terms of the partners and third-party sources the data is licensed from.
Restricted. Onward distribution is limited by the terms of the information providers the data is licensed from.
Restricted by contract.
AI training rights
Granted. The dataset can be licensed for use in model training and retrieval, with the scope set out in the contract.
Not offered. Model-training use is excluded by the terms of the partners and third-party sources the data is licensed from.
Not offered. Model-training use is excluded by the terms of the information providers the data is licensed from.
Not offered publicly.

Where this data layer is already doing the work.

AI agents for compliance, banking and financial services run the same six steps against one decision. Each workflow below links to the industry where it is most often deployed.

KYB API for onboarding agents

KYB automation for onboarding. Takes a company name from an application form, resolves it to the register, verifies status and officers, and writes a sourced decision to the case file.

Resolve → Verify → Cite → Log

Banking · Fintech · Marketplaces

Entity resolution software for agents

Deduplicates and matches company records across a CRM, a supplier master and an external list, on registration number rather than string similarity.

Resolve

Technology · Data teams

Claude MCP server for company data

The dataset exposed as MCP tools so an assistant or in-house agent framework calls it directly and receives sourced values.

All six steps as tools

AI platform teams

UBO verification for AML agents

Walks a shareholding chain across jurisdictions to the natural persons in control, each hop returned against the register that recorded it.

Traverse → Cite

Banking · Legal · Public sector

Company financials for underwriting agents

Filed accounts as structured fields by period, so an underwriting agent reads turnover, assets and liabilities from the statutory filing and cites the document.

Read financials → Cite

Insurance · Commercial lending

Supplier verification for procurement agents

Confirms a bidder or supplier is registered, active and owned by whom it declares, before a purchase order or contract award is issued.

Verify → Traverse → Log

Public sector · Manufacturing

Point your agent at a company you already know and check the filing it returns.

That is the fastest evaluation there is. A test key is issued from the platform, the MCP server connects in minutes, and bulk extracts are scoped with our team.

Common questions.

What company data can an AI agent retrieve through Global Database?

Identity and status, officers, shareholders, beneficial owners, group structure, financial statements and filings, read from 400+ government registries across 200+ countries and territories. Field availability varies by jurisdiction because registers publish different things.

What does "sourced" mean on this page?

Each value returned carries the name of the government register it was read from, the filing reference the register issued, the timestamp of the read, and a link to the original document where the register publishes one. A reviewer can open the filing independently of us.

What is AI grounding, and why is retrieval different?

AI grounding, or LLM grounding, supplies a model with context so its answer refers to that context rather than to training data alone. The model still composes the answer from its weights and can still produce a value that appears in no source. Retrieval, as Regis and the structured endpoints do it, returns the registry record itself with its provenance. Regis holds no company facts in its weights, so an answer without a retrieved record is an empty answer, not an invented one.

What is Know Your Agent, and how is KYBA different?

Know Your Agent (KYA) authenticates the software: which agent is this and is it authorised. Know Your Business Agent (KYBA) asks the question behind it: which legal entity is responsible for this agent, is that entity registered and active, who owns it and who may act for it. KYA is an identity problem. KYBA is a registry problem, and this dataset is what the verifying agent retrieves to answer it.

Is this a KYB software or KYB solution, or a data source for one?

A data source, delivered as API, MCP server, Regis and bulk feed. Teams building or buying KYB software use it as the registry layer underneath, so that every check the software runs resolves to a filing. It is not a workflow tool and does not replace case management.

Which delivery route should an agent use?

MCP if the agent runs in Claude or an MCP-compatible framework. REST if a pipeline already knows the field it needs. The Regis API if the input is a question rather than a field. Bulk if you index company data into your own retrieval layer. All four read the same records.

Does the MCP server work with Claude, OpenAI and Microsoft Copilot?

Yes. There are connectors for Claude, OpenAI and Microsoft Copilot, and because it is a standard MCP server it also connects to any other MCP client or in-house agent framework. Setup details are on the MCP server page.

What happens when a register does not publish a field?

The response states that the field is not published by that register for that entity type, and names the register. No estimate, modelled value or figure from a similar company is substituted.

How current is the data?

Records are synced daily. Registers publish on their own schedules, so each value carries a retrieval timestamp showing what the register held at the time of the read.

Can we redistribute the data, or train a model on it?

Both are available under licence. Standard access covers internal use. Redistribution, meaning serving the data to third parties inside your product or through your agent, is licensed separately. Model training and retrieval use is also licensed, with the scope written into the contract. Ask us for the terms before you build against them.

Does this help with EU AI Act logging obligations?

The response carries the resolved entity, the fields used, the register and the retrieval timestamp in structured form, which is the material an Article 12 event log needs to reconstruct a decision. Retention and the rest of the compliance framework remain the deployer's responsibility.

What security and compliance documentation is available?

ISO 27001 certification and UK ICO registration. Security documentation and the data processing agreement are available under NDA before contract.