How does the tax administration use Artificial Intelligence? The answer lies in the 2025 activity report of Public Finances, which places AI among the year’s emblematic actions.
Moreover, according to Amélie Verdier, the Director General of Public Finances, “the use of artificial intelligence is at the heart of our strategy. 56% of controls concerning professionals and more than 54% of those targeting individuals are scheduled thanks to these technologies.”
The official message is that an administration is mastering its transformation. The DGFiP (Direction générale des Finances publiques) asserts that it has “laid the foundations for a prudent use of artificial intelligence to provide better service to users and to lighten the workload for agents.”
Because the DGFiP is not a newcomer in this field. It has been developing systems based on artificial intelligence for more than a decade, long before the term became ubiquitous in public debate. Yet, in 2025 the Direction de la Transformation Numérique (DTNUM) equipped itself with a formal AI roadmap and strategy.
A strategy AI formalized in 2025
Two structuring projects illustrate the administration’s early lead in this domain.
The first is the CFVR (Ciblage de la Fraude et Valorisation des Requêtes), an AI system dedicated to the fight against tax fraud. With an estimated budget of €34.5 million, it has enabled more precise targeting of tax audits. The 2025 report even notes a 2.8% increase in rights and penalties notified, and mentions a “strengthened fight against patrimonial and declarative fraud by individuals” as well as “enhanced monitoring of crypto-assets.”
The second is Foncier Innovant, a program for automated detection of undeclared constructions (pools, extensions, …) based on aerial imagery. Its deployment, revealed by the press before any official communication, had been a topic of heated discussion with an approximate cost of €33 million.
The DGFiP also uses AI in the arena of local public governance, with a predictive alert system that analyzes the financial data of municipalities to anticipate their difficulties, classifying them by comparable profiles. This device, cited by the OECD in its 2025 report on governance through AI, combines a scoring algorithm with clustering techniques to identify weak signals as early as the fourth year of data.
The Generative AI assistant: the flagship tool of 2025
The pivot of 2025 is the move to generative AI with the launch of an internal assistant structured into three modules. First, a general chat, allowing agents to ask questions or obtain summaries of texts. Then, DocuFiP is a semantic search engine within the internal tax doctrine.
Finally, “Plain Language” (or “Speak Clearly”), aimed at reformulating technical responses from agents into accessible language for users. Tested since summer 2025 with 500 agents, it is expected to be offered to all agents by the end of 2026.
The justification put forward by management for developing these tools in-house rather than resorting to market solutions is explicit: to avoid agents using “commercial AIs like ChatGPT,” with the data leakage risks that this implies. A digital sovereignty argument that is tempered by the fact that the second phase of experimentation ultimately relied on an open-source model developed by OpenAI.
The internal assessment of the experimentation, whose data were made accessible, paints a more nuanced picture than the institutional discourse.
Des résultats officiels en demi-teinte
On the quality of the chat responses, the average score given by agents who tested the tool stood at 5.3 out of 10. Among them, 66% feel that the assistant responds “partially” to their expectations. Many noted that the replies were too general, occasionally inaccurate, and that a simple search via a traditional search engine was often more effective. The use of the tool also declined rapidly during the trial, signalling a difficult uptake.
Regarding the “Plain Language” module, the official assessment itself concedes that the tool “does not truly save time on drafting messages, as proofreading takes a lot of time.” Agents must systematically check the reformulations produced, which can contain errors about articles of the General Tax Code, imprecise formulations, or even anglicisms.
Facing the initial poor results of the chat, DTNUM initially steered communications toward a perceived lack of mastery of the “prompt” by agents, before acknowledging the tool’s limits and changing the model in the second phase.
