Slight disappointment on entering the digital data analysis cell of the tax authorities. Described as the rear base of the tax “Big Brother”, from which the tax authorities would track us even on social networks, the fraud targeting unit and valuation of requests (CFVR) is nothing like a futuristic room, with technologies ultra-sophisticated.
This room of a few tens of square meters, on the sixth floor of the Sully building, at the Ministry of Economy and Finance, is an office space like any other, in which a dozen people are busy on lines of illegible codes. Admittedly, the team, made up of tax experts, but also computer scientists, statisticians and data scientists, is undoubtedly younger than in the rest of the house. But the austere atmosphere of the DGFIP (Direction Générale des Finances Publiques) is preserved in every way.
Limit the holes in the racket
Created in 2014, but operational since 2017, this cell is nevertheless at the heart of a small technological revolution. Its objective: to analyze, using statistical and mathematical methods, the data available to the tax authorities to detect anomalies in the tax declarations of companies and individuals.
Initially, the project essentially consisted of merging and making usable these terabytes of banking, heritage, social, or even notarial data to which the tax authorities have access, but in a totally compartmentalized way. A project that is all the more considerable as year after year, the gold mine is enriched, with in 2018 the integration of the foreign bank accounts of French residents, in 2019 the information provided by the digital economy platforms on their customers, and since 2020 the recovery by the tax authorities of certain data on social networks, on an experimental basis.
What limit the holes in the racket. “Before 2017, you could not declare the IFI and make a donation of 4 million to one of your children. Today, this type of anomaly no longer occurs,” testifies Arnaud Tailfer, tax lawyer at Arkwood, who no longer counts the number of clients being pinned for having undervalued a property compared to the prices of notaries.
Work on existing fraud schemes
But beyond these classic frauds, requiring simple algorithmic modeling, what can the tax authorities do with this ocean of data? For example, is artificial intelligence able to identify new fraud patterns? “For the moment, it is mainly a question of working on existing schemes, such as VAT fraud, which was the first to be modelled, explains Gilles Clabecq at the head of the analysis unit. To do this, we establish profiles of suspect populations through correlations. We can also randomly look for atypical behavior, such as a profit that increases without an increase in turnover, for example. »
The advantage of these models: as we integrate the characteristics of fraudsters into the machine, targeting improves. Example with the research tax credit (CIR): in year N, the robot screened 65 variables (turnover, sectors of activity, number of jobs in research and development, etc.) from 1, 2 million companies having benefited from the CIR, some of which have been caught for fraud. The following year, the algorithm produced a population of 1,600 companies deemed to be at risk. Among them, many restaurants, and even a few theaters, a priori reluctant to do R & D.
A battery of innovative tools and technologies
Today, out of the 230 fraud risks identified in companies, around thirty have been modeled by the unit, with very conclusive results. “The proportion of files involved in control on the basis of programming by data analysis is one in three against one in ten previously”, says Gilles Clabecq.
The tax authorities have also equipped themselves with a battery of tools and technologies, such as “innovative land”, which tracks swimming pools and undeclared constructions using aerial views from Google, or the Galaxie software, which graphically represents all the financial and capital links existing between companies and natural persons.
A kind of genealogical tree on which are also all the old tax audits. “For agents, who had to retrieve information from separate databases, and draw diagrams by hand, it is a very valuable tool,” testifies Stéphane Créange, the deputy director of tax audit.
Efficiency contested by the unions
But beyond this indisputable time saving, has artificial intelligence really improved the efficiency of the tax authorities? In recent years, many unions have denounced the forced digitization of their administration, which has been accompanied by the elimination of 1,500 to 2,000 jobs each year for ten years at the DGFIP.
Last year, Bercy was nevertheless proud to announce that the analysis of digital data had represented 45% of the programming. “Except that this only represents 1.2 billion euros, or only 9% of revenue from tax audits, which, moreover, have not progressed with artificial intelligence”, denounces Sabine Portella, at Solidaires-Finances publiques.
This difference is partly explained for reasons of temporality: the frauds targeted in 2021 will only generate revenue one or more years later. And at this stage, artificial intelligence above all saves time on small files, which do not constitute the bulk of fraud.
“We must not forget that nearly 40% of revenue from tax audits comes from the fraud of large companies, where the problem is above all that of proof”, explains a former consultant from Bercy on these issues.
To catch these big fish, there is no secret: it would take more investigations and therefore more agents. In 2023, 850 jobs should be cut again at the DGFIP, with an exceptional envelope of 450 million euros to continue the digitization of the tax authorities.
Weak but growing revenue
In 2021, the tax authorities collected €10.7 billion revenue from tax audits. A figure up from 2020 (7.8 billion euros), a year which had suffered from the pandemic, but slightly down from 2019 (11 billion euros), which had been a record year.
Of this amount, 1.2 billion euros come from controls resulting from data analysis. A figure up 51% over one year.
For 2022, these revenues already represented 1.4 billion euros at the end of August, of which 10 million euros are to be put on the account of innovative land which has made it possible to detect 20,356 undeclared swimming pools since the start of the year.