MMOOZZGG.com

AI against corruption

Network graph analysis: how AI sees webs of companies, people and money

7 min read4 sources

Network graph analysis puts people, companies, accounts and contracts into one network and shows who is linked to whom and in how many steps. A table answers “how many”; a graph answers “who with whom”. In Colombia, a review of 8,678,464 public contracts found 54,155 cases where one representative sits in several companies, and 8,749 companies with one phone number.

On this page · 6 sections

In short

Network graph analysis shows networks: who manages the suppliers, who shares a phone number, where the money from a contract went. This is how ICIJ opened links from 11.5 million Panama Papers files to journalists, and how a Colombian platform found 8,749 companies with one phone number in public contracts. A link in a graph is a risk sign, not proof.

What a link graph is

A graph is made of nodes and links. A node is an object: a person, a company, an account, a contract, a phone number, an address. A link is a relation between them: manages, signed, transferred money, took part in a tender.

The strength of a graph is in paths. The request “show everyone within two steps of this person” takes a chain of complex queries in ordinary tables, but one line in a graph database. ICIJ’s head of data gives this example when describing how journalists worked through the Panama Papers.

A graph does not replace tables and rules. Red flags count the signs of a single purchase, and the graph adds what shows up only in combination: a shared director of two bidders, a shared phone number, money moving in a circle.

AI does two jobs here. First, it builds the graph from different registries and matches records where the same person is written in different ways. Then it looks for risk patterns in the graph: closed loops of transfers, shared nodes between competitors, a short path from the buyer to the supplier.

Case modelA conditional example
signedbidderbidderheadssame numbersame numbertransfertransferCustomerContractSupplier ASupplier BDirectorShared phoneAccount

Every link leads to a document: a contract, a bank statement, a registry entry.

Panama Papers: a scale no one can read

The Panama Papers are 11.5 million files, 2.6 TB in size, about 214,488 offshore entities. More than 370 journalists from 76 countries worked through them together.

No one can read such a volume by eye, so ICIJ loaded the links between people and companies into the Neo4j graph database and opened it to journalists through Linkurious. A journalist clicked on a point, it opened into linked nodes, and search worked by name. Experienced journalists and programmers wrote complex queries against the database.

First, the documents had to be made readable. The team ran text recognition with Tesseract on 30–40 temporary servers and, according to ICIJ, recognized and indexed about 99% of the data. Most of the files were emails and database files.

For an oversight body, what matters here is not the scale but the method. Documents turn into nodes and links, the links go into the graph, and a person asks questions of the graph instead of rereading the archive.

A graph shows a link, but not whether it is legal. An offshore entity in a graph is not a violation in itself: the conclusion is made by an auditor, an investigator or a court.

Nodes and links in procurement

A sample case: the graph of one procurement case. Nodes: “Buyer”, “Supplier A”, “Supplier B”, “Director”, “Shared phone”, “Account”, “Contract”. The links are labeled: “signed”, “manages”, “same number”, “transfer”, “bidder”.

How to read this diagram. The buyer signed a contract. Supplier A and Supplier B are linked to the contract as bidders, so on the surface they compete. The director manages both, and both use one number, the shared phone. The money from the contract goes by transfer to an account. Each link is legal on its own; together they form a risk sign.

Each link in such a graph comes from a document, not from a guess. The tender record gives the bidders, the company registry gives the directors, the bids give phone numbers and addresses, and treasury and bank data give the transfers.

The Colombian platform Oceano shows how many such links real data holds. Among 8,678,464 public contracts from 2014–2019, it found 54,155 cases of one representative in several companies, 8,749 companies with one phone number and 45 contracts with people who had died.

Each such finding can have a legal explanation: a holding has one director for several subsidiaries, a business center has one phone for all tenants. So the link is marked as a risk sign, and the auditor checks whether the linked companies met in the same tenders. How such links are checked is covered on the page about affiliation.

Networks around one buyer

A graph around one buyer is built like this: the buyer in the center, around it all of its suppliers over several years, then their directors, founders, addresses and phone numbers. The picture shows at once whether the suppliers are spread across the market or lead back to a few people.

Risk signs in such a network:

1. Several suppliers of one buyer with a shared director or founder.

2. Competitors in a tender with a shared phone number or address.

3. A link between a supplier and an employee of the buyer through shared founders or a former job.

The third sign needs data on officials and owners. In its Arachne+ system, the European Commission uses data on company owners, sanctions lists and politically exposed persons and counts 78 risk indicators for EU budget spending; the earlier version had 107.

According to the European Commission, Arachne+ takes data on company owners and structure from the Orbis database, and sanctions lists and data on politically exposed persons from World Compliance; data on beneficial owners is used separately. The service launched on 27 April 2026.

An important detail of the same system: Arachne+ makes no automatic decisions. It supports decisions based on a risk score, and people draw the conclusions.

The path of money through a chain of companies

In a graph, the path of money shows up as a chain of transfers: from the buyer to the supplier, from the supplier to the next companies. The auditor wants to know where the chain ends and whether it comes back to people linked to the buyer or to the losing bidders.

Risk signs on the money path: the money reaches a person linked to the buyer; transfers go in a circle between companies with shared owners; the chain includes a company set up shortly before the contract, with no staff or assets.

Such a graph needs payment data: treasury statements on the contracts and, where there is legal access, the suppliers’ bank transactions. Without them, the graph shows only formal links, not the movement of money.

The money path connects a formal link with a benefit. A shared director of two bidders stays a sign, while a transfer from the winner to a company linked to that director is already a reason for a detailed check.

Similar chains are searched for in tax schemes, where goods and money pass through a series of companies and the tax is lost at one of the links. How this looks in the data is covered on the page about VAT schemes.

How a graph helps the auditor

A person can’t hold a thousand links in mind, but a machine collects them in minutes and updates them when new data arrives. The auditor gets not a table of hundreds of rows but a diagram of the case that shows the main points.

What the auditor does with a signal from the graph:

1. Looks at which link triggered the signal: a shared director, a shared phone, a transfer.

2. Checks the source of each link: a registry extract, a tender record, a payment document.

3. Rules out legal explanations: a holding, a shared business center, a registry error.

4. Decides whether a check is needed and passes the materials on if the signs are confirmed.

The graph of one case, which updates with each new document, is described on the page about the case model. It also shows how a signal from the graph becomes a case for the auditor.

A graph flags a risk but does not prove a violation. A shared phone may belong to an accounting firm that serves dozens of clients. The conclusion about a violation is made by an auditor, an investigator or a court, and Arachne+ writes this directly into the description of the system: it makes no automatic decisions.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. ICIJ, 3 April 2016 and 25 April 2016, 3 April 2016 (icij.org)
  2. El Espectador, 25 October 2019 (elespectador.com)
  3. European Commission, Anti-Fraud Knowledge Centre (checked 16 October 2026), 16 October 2026 (antifraud-knowledge-centre.ec.europa.eu)
  4. ICIJ, Wrangling 2.6TB of data: The people and the technology behind the Panama Papers, 25 April 2016 (icij.org)

Found a mistake? Tell us

Publications

For government bodies

We build such systems for government bodies and state companies, from data analysis to a pilot on one scheme. The data stays on the organization's own servers.

Contact us