Tax Fraud and AI: A Technology Now Challenged by Complex Fraud

Since 2014, artificial intelligence has deeply permeated tax audits, making the General Directorate of Public Finances (DGFIP) one of the pioneering administrations in this area: detection of undeclared constructions or developments from aerial photographs, collection of data on social networks, etc.

And above all, more efficient programming of tax audits, by selecting files using this technology, before the agents' checks.

"Before, we had to look at between 10 and 15 files to find an anomaly. Thanks to AI, we are now at one file in three," estimates Gilles Clabecq, head of the department responsible for implementing AI in tax audits.

Mainly made up of contract engineers, the department's workforce increased from five to 32 people between 2013 and 2025.

In this "kitchen", one of the models developed internally makes it possible to identify tax anomalies in a file in relation to a database of audit results from the last five years.

Another groups companies according to homogeneous categories and criteria and identifies suspicious "behavioral disruptions", such as a reduction in revenue.

Every quarter, several tens of thousands of files identified by AI are sent to the control services, and it is up to the auditors to assess their relevance, explains Gilles Clabecq.

Of the approximately 15 billion euros to be collected in 2024, AI contributed 2,5 billion euros, according to the DGFIP.

"Transparency"

But on the ground, agents are not unanimous on the quality of the controls proposed by the AI, assures Solidaires Finances publiques, the leading union at the DGFIP.

The organization deplores a "lack of transparency" of the models and a "forced march" deployment of AI which is not neutral on employment.

According to Benjamin Gandouin, from Solidaires Finances publiques, funding to develop AI is "always conditioned by forecast savings that are made on the backs of our agents".

According to a parliamentary information report in June 2024, "tax audit staff numbers are marked by a drop, going from 12.303,7 full-time equivalent jobs in 2016 to 10.427 in 2022".

But the DGFIP refutes "any correlation between job cuts and the development of AI", and puts forward other arguments, including organizational reforms, the elimination of the housing tax on primary residences and the abolition of the contribution to public broadcasting.

Complex fraud

Solidaires Finances Publiques also points out blind spots in technology.

"The argument was that AI would help solve complex fraud, particularly international fraud, but the models do not allow it to be discovered," underlines Damien Robinet, national secretary of the union.

For this type of fraud, characterized by sophisticated practices to evade tax, often of international dimension, the lack of sufficient data volume can hamper the development of new models.

But the DGFIP is particularly counting on the generalization of electronic invoicing in 2026.

"The added value of AI will really explode from there (...) With billions of invoices, this is an unprecedented volume and granularity and this is data (...) very reliable", underlines Gilles Clabecq. Enough to boost new models and more easily detect certain frauds, in particular those involving fictitious companies, devoid of activity, and created with the aim of collecting public money.

This is the case of VAT fraud, which involves several companies established in at least two EU Member States, and consists of obtaining the deduction or reimbursement of VAT relating to an intra-Community supply of goods when this VAT has not been paid to the tax administration concerned.

This so-called "carousel" fraud costs the European Union almost 50 billion euros per year, according to the latest available estimates from Europol.

"When a temporary company files a request for a VAT credit refund and it is found [from the invoice] that it is linked to a supplier known to be unfavorable to the DGFIP, we will be able to say 'we do not refund'," anticipates Gilles Clabecq.