Following the tax office hack, Mistral engineers are to set up in government departments.

It has been a tough summer for the Directorate General of Public Finance. The hacking of the tax authority exposed the tax data of 678,000 individuals and professionals, and in the process laid bare the state's cybersecurity shortcomings.The response takes the form of a new agreement with Mistral AI, revealed by Les Echos and confirmed by the office of David Amiel, the Minister for Action and Public Accounts, who is signing the partnership.

Six million euros and Mistral engineers in government departments

The agreement, concluded following a public procurement process and without exclusivity, represents an investment of 6 million euros over 2026–2027. A modest amount (budgetary constraints make it necessary, it is admitted at Bercy) for an ambition stated without detour: “to unlock powerful use cases in little time”. Mistral will second its own engineers to government agencies: the famous “forward deployed engineers” sent to the front line, whom AI startups rave about.

Three workstreams are taking priority over the coming weeks. Cybersecurity first, given the DGFiP hack. Justice next, with the handling of mass disputes to ease the burden on overcrowded courts, as legal proceedings multiply in the AI era. Fraud finally: Bercy wants to target fraudulent behaviour on social media and non-compliance in retail and online, while it already uses machine learning to track undeclared swimming pools (if yours is properly registered, everything is fine).

“The latest achievements of AI must prompt us to reflect: it is urgent to call on AI to protect ourselves from AI”, argues David Amiel in the columns of the business daily.Depending on the sensitivity of the data, the tools will be hosted on Mistral’s infrastructure, on Outscale’s sovereign cloud, or directly in government agencies’ data centers. The choice of deployed models has yet to be decided.

A national champion its own civil servants loathe

This contract extends an already well-established relationship between the government and Arthur Mensch’s young startup. In June, the assistant developed with Mistral was rolled out to two million public-sector employees, and the startup (valued at nearly €12 billion, with the government an indirect shareholder through Bpifrance) already works with the Ministry of the Armed Forces, France Travail and the health insurance system. It is hard to be more central to the government’s AI strategy.

The pilot evaluation (10,000 licenses across six ministries) was conducted with researchers from Inria, INSA Rennes and the CNRS. The result: 75% of employees consider the tool useful, but 57% believe that ChatGPT, Claude or Gemini better meet their needs, with lack of confidence in the reliability of responses ranking as the leading obstacle. A sovereign tool will only be convincing if it truly rivals the tools it claims to replace, and that point has not yet been reached.

Sovereignty itself, moreover, deserves to be put in quotation marks: the Chinese GLM-5.2 model has been available on Mistral’s infrastructure since this summer.And in government, not everyone is pushing in the same direction. “If European AI comes down to Mistral, then we’re done for”, said Economy Minister Roland Lescure during a trip to Silicon Valley, advocating a complete ecosystem rather than a single champion.

Before it can protect French citizens’ tax data, sovereign AI will therefore have to win a more intimate battle: convince government employees to stop opening ChatGPT in another tab.

👉🏻 Follow the latest tech news in real time: add 01net to your sources on Google and subscribe to our WhatsApp channel.

Source: Les Échos