The French government is expanding the use of artificial intelligence in public administration, with plans to provide 1 million public employees with AI assistants. This initiative is funded by a 655 million euro plan aimed at integrating AI into daily administrative tasks. The move comes as officials recognize the importance of ensuring AI decisions are traceable, explainable, and under human oversight. A survey accompanying the Our AI plan found that 89% of public agents already use AI tools, with 55% relying on unverified public tools. This widespread use has raised concerns about the need for proper governance to manage the risks of uncontrolled AI deployment. Authorities such as the National Cybersecurity Agency (ANSSI), the National Commission on Informatics and Liberty (CNIL), and the Council on AI and Digital have pointed out the difficulties in assigning responsibility for decisions made by autonomous AI systems. As these systems become more self-directed, it becomes harder to determine who is accountable for a decision, and the source of a malfunction can be obscured by complex and unclear processing steps. This has led to worries about critical issues, such as the incorrect denial of benefits or the spread of errors through administrative processes. The path outlined by the authorities begins with AI assistants that currently assist with writing and summarizing information and moves toward a future where AI may play a more central role in decision-making. This raises the question of who is considered the author of administrative decisions. Solutions exist that focus on the structure of the AI systems rather than waiting for future technological breakthroughs. An autonomous AI agent can be designed in two ways: one that makes up its logic during execution, creating a gap in responsibility, and another where the decision-making logic is clearly defined and validated beforehand. For processes like handling aid applications, eligibility criteria, required documents, and thresholds are established and approved during the design phase. The AI agent gathers the necessary information, applies the pre-approved rules, and refers ambiguous cases to a human employee. Each decision can then be traced back to a specific rule and the person who validated it. Autonomy in this context refers to the sequence of actions taken, not the underlying reasoning behind the decision. The workflow remains the top priority, ensuring that each administrative action is repeatable, justified, and linked to a documented intention, and thus to an identifiable author. With the new European AI regulation coming into effect, traceability is no longer a luxury but a legal requirement. French law aligns with this, as the Constitutional Council has accepted administrative decisions based on algorithms but explicitly ruled out those based on systems that change the rules they apply without oversight and validation.