MMOOZZGG.com

AI against corruption

Fraud investigation model: how AI builds a full map for each investigation

8 min read3 sources

A case model, or fraud investigation model, joins into one map everything a machine found in tenders, payments and registries: who signed the contract, who runs the suppliers, where the money went and when. AI builds such a map faster than a person and holds thousands of links at once. The reviewer reads the finished picture and decides what to check.

On this page · 6 sections

In short

A fraud investigation model brings people, companies, accounts, contracts and dates into one graph where every link is labeled and leads to a document. AI builds such a model for each case and updates it when new data arrives. ICIJ journalists analyzed the links from 11.5 million Panama Papers files in a graph database.

What a case model is

A case model is built as a graph: nodes and links between them. The nodes here are the buyer, suppliers, directors, phone numbers, addresses, accounts and contracts. The links are labeled: “signed”, “runs”, “same number”, “transfer”, “bidder”.

Every link has a source and a date: a registry extract, a tender record, a payment order. So the model can be checked: any node opens down to the document it rests on.

The model is built around one review question, for example a specific contract or supplier. From it the machine goes out two or three steps of links and stops, so the map stays readable.

The model does not replace the review files; it links them. Instead of a folder with a hundred documents, the reviewer sees one map and opens only the documents behind suspicious links.

The model draws no conclusions about guilt. It shows who is linked to whom and how the money moved, and highlights risk signs separately. Arachne+, the European Commission’s system for finding fraud, conflicts of interest and irregularities with the EU budget, works the same way: it makes no automatic decisions.

A risk sign on the model means risk, not proof of a violation. The reviewer, investigator or court decides, and the model gives them a signal to check and the collected documents.

Case modelA conditional example
signedbidderbidderheadssame numbersame numbertransfertransferCustomerContractSupplier ASupplier BDirectorShared phoneAccount

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

What it is built from: documents, payments, registries

The model is built from the same data the control function already works with, only brought together:

1. Tenders: notices, records, bids and final prices.

2. Contracts, amendments, acceptance certificates and delivery dates.

3. Payments: who sent how much to whom and when, with the payment purpose.

4. Company registries: directors, founders, addresses, phone numbers, registration dates.

5. Information about the buyer’s officials and their links, if the reviewer has lawful access to it.

Arachne+ in the EU relies on a similar set: data on company owners, sanctions lists and public officials. From it the system computes 78 risk indicators; the earlier version had 107.

Some data comes not as a table but as a document. A language model extracts from a scanned contract the parties, the amount, the dates and the signers, and turns them into nodes and links. A person checks a sample of such extractions before trusting them.

Each record gets a source and date mark when it is loaded. Without it, a link in the model cannot be shown to the reviewer as an argument; it can only be checked again.

Personal data in the model is processed under the personal data law: only reviewers with the right to this information get access, and every opening of a case is logged.

The link graph: who is connected to whom, and through whom

A made-up example, as in the diagram above. The Buyer signed a Contract with Supplier A. Supplier A and Supplier B both bid in the tender for this contract, and both have the same number: a Shared Phone. One Director runs both companies. The Contract was paid to Supplier A’s Account, and a transfer went out from this account.

Each link alone means nothing: companies may share an accountant with one phone, and a director may run two firms. Together they form a risk sign: the tender may have had not competitors but two companies of one person. This is a signal to check, not a conclusion.

Every link in the graph has a weight: how reliable it is and how unusual. A shared legal address in a business center with a thousand companies means little; a shared mobile phone for two bidders in the same tender means much more.

Real data has many such matches. In 8,678,464 public contracts from 2014–2019, the Colombian platform Océano flagged 8,749 companies with the same phone number and 54,155 cases of one representative in several companies.

The graph also answers the question “through whom”. If there is no direct link, the machine looks for a path of two or three steps, for example through a shared founder or address. A person follows such chains by hand for one case; the machine follows them for all participants at once.

Methods for finding such paths are covered on the graph analysis page, and typical signs of links between buyer and supplier on the affiliation page.

Timeline: money and decisions by date

The second axis of the model is time. Every link and every payment sits on a date scale, and the order of events often says more than the events themselves.

What the machine marks on the scale: a company was registered a few weeks before the tender; the director changed right after the contract was signed; the advance went out before the acceptance certificate was signed; a transfer from the supplier’s account went out on the day the contract was paid.

The machine also notices rhythm. If payments to one supplier come in series before the end of a quarter or year, the scale shows it at once, while a table hides it.

Some signs show only when dates come from different sources. Océano in Colombia found 45 contracts with dead people: to do this, it matched the contract date with the date of death from another registry.

The reviewer needs the timeline too. In a case they see who made each decision and when, which documents to request first, and what happened after an earlier review, if there was one.

Why a person cannot keep a thousand links in mind

A person can reliably keep a few links in memory. A case with a dozen companies, their directors, phones, accounts and contracts has hundreds of links, and a review of all tenders of one agency has thousands.

The second difficulty is links between cases. A supplier in today’s tender may have appeared in a review three years ago; the graph keeps past cases and shows such an overlap at once.

The International Consortium of Investigative Journalists (ICIJ) gave an example of the scale. The Panama Papers held 11.5 million files (2.6 TB) on 214,488 offshore entities, and more than 370 journalists from 76 countries analyzed them together.

Nobody can read such a mass of files from start to end. ICIJ loaded the links between people and companies into the Neo4j graph database and opened them through Linkurious, so journalists searched for links instead of paging through files.

A machine builds the model of one case in minutes if the data is already in one place: this is an estimate for the method, not a project timeline. Most of the time goes earlier, into collecting and reconciling data.

The gain is not only speed but completeness. An analyst follows the chain they noticed; the graph shows all chains between the parties to a contract, including ones nobody thought of.

How the model updates in real time

The model is not built once. When a new tender, payment or registry extract arrives, the machine adds nodes and links and recalculates the risk signs for the whole case.

A made-up example: a month after the contract, Supplier B changed its director, and the new director is an employee of the buyer. A new link appears in the model, the risk score rises, and the case moves up in the reviewer’s queue.

Every change is saved with a date, so you can see what the model knew at the moment of any decision. This matters when the reviewer’s conclusions are later disputed.

Updates work the other way too. If a supplier gave an explanation and documents, the reviewer marks this in the model, and the link stops raising the score.

The decision still belongs to a person. The European Commission’s Arachne+ computes indicators from fresh data on owners and officials but makes no automatic decisions; the case model works the same way.

How new data enters the system every day and how it forms a queue of signals is covered on the real-time monitoring page.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. European Commission, Anti-Fraud Knowledge Centre (checked 16 October 2026), 16 October 2026 (antifraud-knowledge-centre.ec.europa.eu)
  2. El Espectador, 25 October 2019 (elespectador.com)
  3. ICIJ, 3 April 2016 and 25 April 2016, 3 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