MMOOZZGG.com

AI against corruption

Missing trader fraud: how AI finds VAT and refund schemes in chains of deals

7 min read5 sources

Finding missing trader fraud and other VAT schemes starts not with one tax return but with a chain: who sold to whom, who paid the tax and who got a deduction or refund. AI builds this chain from invoices and payments and shows the gaps where the tax never reached the budget. In the EU, VAT cases make up 18% of investigations of the European Public Prosecutor’s Office (EPPO) and 53% of the damage.

On this page · 6 sections

In short

Missing trader fraud and other VAT schemes show in the data as a gap in the chain of deals: the seller did not pay the tax, while the buyer claimed a deduction or refund. AI joins invoices, payments and company links into a graph and flags such gaps. In the EU, VAT accounts for 53% of the damage in cases of the European Public Prosecutor’s Office, and a tax inspector decides on each signal.

How a VAT refund scheme works

Every link in the chain pays VAT, and a buyer who paid the tax to a supplier gets a deduction. The scheme rests on a gap: on paper the tax was paid, but it never reached the budget.

The European Commission describes a well-known variant, missing trader intra-community fraud. A company gets goods from another EU country without VAT, sells them to other companies with the tax added and disappears without paying the tax to the budget.

According to the European Commission’s report of 11 December 2025, losses from this scheme ranged from EUR 12.5 to 32.8 billion a year in 2010–2023. The total VAT gap in the EU in 2023 reached EUR 128 billion, or 9.5% of the tax due.

For control, what matters is not the mechanism but the trace in the data. Such a scheme almost always leaves one: a company that sold a lot but paid little tax, and buyers who claimed deductions on its invoices.

Where the scheme shows in the data: the seller’s return shows little tax on a large turnover; in the buyer’s return, a large deduction comes from one new supplier; in a refund claim, the amount grows faster than the company’s revenue.

The scheme does not need a border between countries. The same gap appears inside one country when the chain includes a company with no staff, no assets and no history.

Chains and gaps in the data

The tax service sees each link separately: the seller’s return, the buyer’s return, payments. AI joins them into a graph where companies are nodes and invoices and payments are links between them.

What the system flags in the graph:

1. A gap: the buyer claimed a deduction, but the seller’s return shows the tax as not charged or not paid.

2. A short life: the company was registered recently, quickly built a large turnover and stopped filing reports.

3. A circle: on paper the goods pass through several companies and return to the first one.

4. Shared signs: one address, one phone, one IP address when filing reports, one director for several links.

5. A profile mismatch: a company with no warehouse and no staff sells large batches of equipment.

Each sign alone proves nothing. A signal to check appears when the signs meet in one place in the graph. How this analysis works is covered on the graph analysis page.

What a person does next: the inspector opens the graph around the flagged company, asks the buyer for delivery documents and checks whether the delivery really happened: transport, warehouse, payment. If there was no delivery, the deduction is removed and the files go to investigators.

Journalists also use graphs to find links. The Panama Papers held 11.5 million files on 214,488 offshore entities; ICIJ loaded the links between people and companies into the Neo4j graph database, and more than 370 journalists from 76 countries analyzed them together.

EU: the VAT share of the damage

The European Public Prosecutor’s Office (EPPO) investigates crimes against the EU budget. At the end of 2024 it had 2,666 active investigations with damage of about EUR 24.8 billion.

VAT cases make up 18% of investigations but 53% of the damage. There are fewer of them than other cases, but each one is larger on average: their share of the damage is almost three times their share of cases.

In 2024, assets worth EUR 849 million were frozen in EPPO cases. A freeze is possible when investigators quickly see where the money went, and for that they need the whole chain of companies, not one link.

For a public body the lesson is simple: if there are few reviewers, analysis of VAT chains gives more damage per case than many other schemes. Chains often pass through several countries, so no single national service sees them whole.

How the European Union checks budget spending with other tools is covered on the European Union page.

Tax services and AI: OECD data

In the Anti-Corruption and Integrity Outlook 2026, the OECD for the first time assessed the integrity systems of 62 countries: 37 OECD members and 25 partner countries. In the chapter on fraud, the report gives data from the 2024 survey on tax technology.

29 of 38 OECD countries use AI in tax administration. Of them, 76% use it to detect tax evasion and fraud, and 69% to assess risk.

These figures do not say how many schemes were found or how much money came back. They show something else: finding fraud has become the main task of AI in tax services, not a side one.

Such a system works like fraud protection in a bank. The model gives each refund return a risk score, the inspector starts from the top of the list, and the result of the check goes back into the model as an example.

A VAT refund is easiest to check before payment. The money is still in the budget, and the system’s signal turns into a request for documents, not a long search for a company that disappeared.

The link with public procurement

VAT schemes and public procurement overlap. A supplier under a public contract buys goods through a chain of intermediaries, one link in it does not pay the tax, and the supplier gets a deduction.

Joining tax data with procurement data shows the buyer not only the supplier but also its counterparties. If a public contract supplier claims deductions on invoices from companies with signs of a gap, this is a signal to check the contract price too.

The link works the other way as well. A company with the signs of a shell firm in tax data that bids in a tender gets a flag in the procurement system before the winner is chosen.

Open Contracting Partnership gives an example in its 2024 guide. In Kazakhstan, the analytics company Datanomix and the civil society group Adildik Joly used red flags to look for unlawful VAT refunds to contractors: about $400,000 was returned to the budget, and a proposal to close the loophole went to the Ministry of Finance.

Links between the supplier, intermediaries and the buyer’s people are covered in more detail on the affiliation page.

What you need for the analysis

The minimum data set for finding gaps:

1. Invoices, or purchase and sales ledgers, of both sides of the deal.

2. VAT returns and payments to the budget.

3. The company registry: registration date, directors, founders, addresses.

4. Bank payments between companies, if the law gives the service access to them.

5. For the link with procurement: the register of contracts and tender participants.

The work goes in three steps. First, the data is joined by common company identifiers. Then the system builds the graph and computes risk signs. Finally, the inspector gets a short list with an explanation of which signs were triggered.

A flag means risk, not proof. An honest company can have a gap in its chain too: the supplier made a mistake in its return, filed it late or closed for reasons unrelated to tax. The inspector decides on an audit and extra tax, and a court settles disputes.

Tax secrecy sets the requirements for the system: the model is deployed on the agency’s servers, and only anonymized summaries go outside. According to the OECD, 69% of member countries whose tax service uses AI apply it to risk assessment, so this is now normal work, not an experiment.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. EPPO, Annual Report 2024, 3 March 2025 (eppo.europa.eu)
  2. OECD, 24 March 2026 (pp. 9, 24, 119), 24 March 2026 (oecd.org)
  3. ICIJ, 3 April 2016 and 25 April 2016, 3 April 2016 (icij.org)
  4. European Commission, Taxation and Customs Union, VAT Gap, 11 December 2025 (taxation-customs.ec.europa.eu)
  5. Open Contracting Partnership, Red flags in public procurement, 2024, December 2024 (open-contracting.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