Wednesday, 19 January, 2022

The tax authorities switch to artificial intelligence to track down fraudsters

This argument is flourishing among the “yellow vests”: the average French are burdened with taxes while the rich – the multinationals, the great fortunes – always manage to bypass them. Resentment supported by the revelations of large-scale tax evasion: HSBC, UBS, Panama or Paradise Papers lists, Football Leaks… And also by some resounding cases of famous cheaters, including that of the former Budget Minister Jérôme Cahuzac.

The illegal avoidance of tax would be between 80 and 100 billion euros per year in France, according to Solidaires finances publics, the leading union of tax officials. To compare with the 18 billion reported by tax audits in 2017, the shortfall is colossal. All the more so as the results were inflated by the providential contribution of the Service for processing corrective declarations (STDR). Opened in mid-2013 and closed at the end of 2017, this cell for tax repenters frightened by the end of banking secrecy in Switzerland made it possible to easily collect 8.3 billion euros.

But the windfall is drying up (a last billion came in in 2018). However, Bercy must open this year a counter on the same model for companies, which would entrust their dubious assemblies in exchange for reduced penalties. “These regularizations on the sly are an admission of weakness which makes the tax audit lose its repressive and dissuasive mission”, tackle Véronique Pascalides, at the CGT. But that brings in the money, while since the peak of 2015 tax audits have paid less and less. “The yield remains higher by about 10% than the low water level of ten years ago”, defends Bastien Llorca, assistant to Maïté Gabet, the director of tax control. As for the 2015 record, it was due to “exceptional business”. Two adjustments had been notified, 1.4 billion to Vivendi and 1.1 billion to Google … which were ultimately never cashed. The first was abandoned, the second canceled by the administrative court (the tax authorities appealed). On average, Bercy does not recover more than 50% of the sums claimed, or 9.4 billion in 2017.

3,100 job cuts

In the balance sheet, “the results remain behind the stated ambitions”, soberly indicated the Court of Auditors in June 2018. In any case, they do not reflect the effort of the General Directorate of Public Finance (DGFiP) to move upmarket in tax stalking. However, over the course of the scandals, governments have voted in ten years no less than a hundred anti-fraud measures … but few provide new concrete investigative ammunition to agents. And, at the same time, the DGFiP has seen its troops shrink. Bercy’s commitment to “make tax audit staff safe” only concerns the 4,500 auditors responsible for external audits. It was not applied to other departments, those responsible for off-site checks and, upstream, those who look for anomalies and schedule files to be checked by their colleagues.

Since 2010, more than 3,100 jobs have been cut, out of around 12,000 jobs. Suddenly, the number of checks drops. “The coverage of the tax audit is muted, we lose contact with the field, deplores Anne Guyot-Welke, to Solidaires. As we have to make numbers, we no longer have the means to drive out the small free-riders of everyday life, which thrive. But, at Bercy, we note that the reductions in staff and the number of controls have not affected performance. “To do better, we must above all target better, judge Bastien Llorca. Today, almost one in four audits results in only a small turnaround. This is our weak point. “

This is where the tax 2.0 comes in. The DGFiP thinks it has a lethal weapon: its artificial intelligence software capable of browsing, searching and sorting in a gigabyte of data to extract suspicious profiles. A large-scale industrial project: a budget of 20 million euros has been released over five years to modernize IT and develop this tool. The software was designed by a data mining unit called Mission Queries and Valuation (MRV), made up of 22 high-level computer scientists. The team began in 2015 by gathering files for 5 million taxable companies, previously accessible on disparate applications, then adding available public data (filing of patents and trademarks, trade register, etc.). In 2017, the information of the 37 million taxpaying households was aggregated there (tax, banking, savings, real estate, Urssaf data, CAF, Social security, etc.). And, for a year, the algorithms have been at work.

Promising results

“We have a double added value,” explains cell head Philippe Schall. First, thanks to the multiple cross-references of data, we can allow any controller to increase the relevance of his requests ”: if he seeks to identify e-commerce sites which evade VAT, he will define 4 or 5 criteria. ‘alert and will suddenly obtain a much richer risk analysis report than if he himself had had to make tedious cross-checks. “Next, our algorithms analyze the characteristics of all the companies checked since 2001, to model fraud indicators and therefore automatically identify suspected cases”: the unit can thus send agents in a region the list of all companies in their area detected as having potentially concealed part of their turnover, based on a robot profile of those who have been pinned on this subject for twenty years.

From these explorations, the MRV cell sends thousands of files every quarter to the local services, pointing out what made the machine tick. “The first were rather mediocre, criticizes Christophe Bonhomme-Lhéritier, at the CFDT. The learning curve is arduous. The result depends on the raw material. However, due to a lack of personnel, the quality of the databases tends to deteriorate. “

The unions are especially afraid that the reign of algorithms is a pretext to accelerate job cuts. “On the contrary, this tool should make it possible to promote the work of research agents”, argues Philippe Schall, who encourages agents to make it their own. The first results are promising. In 2018, out of 24,000 MRV proposals, 14,000 triggered off-site checks, yielding 100 million, and 6,900 triggered external checks, generating 235 million euros in adjustments. “Nearly half of our proposals resulted in verification, against 10% to 20% with traditional analysis methods,” he continues.

And the superpowers of MRV software will be increased tenfold by the influx of data from abroad. For a year, after the United States, 49 countries (including those of the European Union) have practiced the automatic exchange of information on the accounts of their citizens, and 53 more since the beginning of the year (Switzerland, Panama…). Likewise, companies with a turnover of more than 750 million euros are now forced to send the tax authorities a report by country of their activity, workforce, profit and tax paid. “We have taken the time to make this information reliable, but we are starting to integrate it and using it opens up new perspectives,” said the geeks boss of the DGFiP with a smile. The objective: to trigger 50% of controls based on MRV proposals by 2022, against 15% today. With the hope of reducing by 25% the controls without adjustment.

The eye on social networks

The next challenge for these digital sleuths: automated social media spying. Controllers already regularly consult the Facebook, Instagram or Twitter accounts of potential cheaters, on the lookout for information on their lifestyle or their real domicile. Company reassessments were notified on the basis of CVs of senior executives on LinkedIn living in Paris while the company claimed to have no establishment in France. And enterprising inspectors in Marmande (Lot-et-Garonne) have even spotted hundreds of swimming pools of individuals not declared for property tax, thanks to aerial views of Google Maps!

But the Minister of Public Accounts, Gérald Darmanin, wants his services to go further, for the process to be industrialized there too. The MRV has just sent a request for approval to the National Commission for Informatics and Freedoms (Cnil) to experiment with the computer search of information posted on social networks…. “It’s about compiling only public posts. But the subject is sensitive in terms of respect for private life and constitutes a real technical challenge to extract the relevant information from it, ”recognizes Philippe Schall. Notice to fraudsters, tax 2.0 has arrived.


DGFIP Remote tax audit Tax fraud