

Artificial intelligence (AI) is no longer merely a research topic: it is disrupting the economy, transforming professions and being widely integrated into public services. In response to this revolution, the legal framework is taking shape, notably with the entry into force of the European regulation (AI Act) and the strict requirements of French administrative law.
NOVLAW Avocats advises both AI developers and the organisations that deploy AI (the State, local authorities, hospitals, private companies) on the legal security of their artificial intelligence systems (AIS).
Integrating AI into your processes requires a multidimensional audit:
The AI Act imposes obligations graduated according to the level of risk the system poses to fundamental rights.
Whether you are an AI "provider" or "deployer", NOVLAW Avocats helps you bring your tools into compliance, in particular through risk classification and mapping: analysing your systems to determine whether they present an "unacceptable" risk (prohibited), a "high risk" (subject to strict requirements) or a "limited risk" (transparency obligations).
Public authorities increasingly use AI to detect fraud, calculate entitlements or guide users (chatbots, allocation systems such as Parcoursup). The use of AI by public authorities is very tightly regulated:
Artificial intelligence is shaking the foundations of intellectual property law and data law:
High-performing AI is often "data-hungry" AI. However, the use of personal data to train or feed an algorithm must comply with the GDPR.
We provide legal support to strip data of its identifying features before models are trained. (Discover our expertise in public data and GDPR)
As far as the public service user is concerned, liability remains with the authority: it cannot hide behind its contractor's malfunction to justify an unlawful decision, and the administrative court will assess the lawfulness of the act, not the quality of the software. Internally, the allocation of liability depends on the contract and on how the parties are classified under the AI Act: an authority that makes the system available under its own name, or substantially modifies it, may switch from deployer to provider status and take on the corresponding obligations. Hence the importance of negotiated clauses on compliance warranties, indemnification and liability caps, as well as a reserved right of recourse in the event of an administrative penalty. As to the timetable: the prohibitions and the AI literacy obligation have applied since February 2025 and the transparency obligations since August 2026, while the Digital Omnibus regulation adopted in June 2026 postponed the obligations relating to high-risk systems to 2 December 2027 (Annex III) and 2 August 2028 (Annex I). Note, however, that this postponement is a preparation period, not an exemption.
Yes, but... provided that the requirement is justified by the subject matter of the contract and proportionate. Article 31 of the SREN Law requires the use of qualified cloud solutions where the data processed are sensitive and their compromise would be likely to undermine public order or public security. The Prime Minister's circular No. 6519/SG of 5 February 2026 on digital public procurement goes further, requiring central government buyers to apply a grid of nine decision criteria, explicitly including digital sovereignty and security. In practice, these requirements come into play at the application stage (SecNumCloud 3.2 qualification or equivalent measures, absence of non-European control), at the tender evaluation stage (security and data location criteria) and in the contract (audit, reversibility, termination in the event of a change of control). The drafting must remain compatible with European competition law and the principle of non-discrimination: be careful, this is where the legal certainty of the contract is won or lost.
This is one of the most common blind spots in AI contracts. In the absence of an express provision, the software vendor may argue that it remains the owner of the model retrained with the authority's data, and freely reuse the performance gains obtained thanks to the public service for the benefit of its other clients, sometimes competitors of the local authority. The CCAP (special administrative clauses) must therefore deal separately with three items: the data supplied by the buyer, the data generated during performance of the contract (annotations, corrections, feedback from staff) and the model parameters resulting from training. The CCAG (general administrative clauses) also provides for clauses to this effect. Reuse of public data for other purposes should be prohibited or strictly controlled, the fate of training datasets at the end of the contract should be provided for, and these provisions should be reconciled with the obligations to publish essential contract data, trade secrets and the GDPR.
In practice, the CCAGs (general administrative clauses), including the CCAG-TIC (for information and communication technology contracts), were not designed for AI systems: they must be supplemented in the CCAP (special administrative clauses). Six families of clauses deserve particular attention. First, the allocation of roles under the AI Act: the contractor must be contractually identified as the provider and undertake to supply the technical documentation, the instructions for use and the information enabling the public deployer to fulfil its own obligations. Next, performance, with measurable indicators (error rate, false positive rate, robustness requirements) rather than mere best-efforts commitments. Then come human oversight and traceability (logging, explainability of outputs), the regime governing data and intellectual property, the management of changes to the model (a substantial update must not silently degrade the service or alter the legal classification of the system) and, finally, reversibility: return of data, configurations and annotation sets, and an enforceable exit plan.
Buying an AI system means anticipating, from the moment the need is defined, what no conventional software specification provides for: the classification of the system under the AI Act, the nature and origin of the training data, the expected level of performance and how it is measured, and the division of roles between provider and deployer. The choice of procedure is also a key structuring decision: prior market sourcing, the innovation partnership, the research and development public contract or the direct award of innovative purchases, whose threshold was raised to €140,000 excluding VAT by the law on simplifying economic life of 26 May 2026, do not meet the same needs or carry the same risks. The State Purchasing Directorate (Direction des achats de l'État) and DINUM have also published a dedicated guidance sheet on the responsible purchasing of AI solutions, which offers sourcing questions and sample clauses. Our firm provides legal assistance to contracting authorities (AMO) across this entire chain, from defining the need to notification of the contract.
Yes, it is legally possible, but strictly regulated. An individual administrative decision may be taken solely on the basis of an algorithm, unless that decision involves assessing a person's behaviour (prohibition of blind profiling). In addition, the administration must ensure that a human official has command of how the algorithmic system works and can explain it to the user.
Yes, this is an absolute legal obligation (French Code of Relations between the Public and the Administration). The administration must explicitly inform the user when an individual decision (such as the award of a grant or an allocation) is based on algorithmic processing. The user has the right to request an explanation of the criteria and weighting used by the AI. Failure to include this notice renders the decision unlawful.
The AI Act imposes several transparency obligations: informing users that they are interacting with an AI system (particularly in the case of chatbots); labelling AI-generated content (deepfakes, synthetic text); and documenting the training data and performance of high-risk systems. Providers of general-purpose AI models (such as LLMs) are required to publish summaries of their training data and to comply with copyright law.
The penalties under the AI Act are graduated according to the seriousness of the breach: up to €35 million or 7% of total worldwide annual turnover for the use of AI systems posing an unacceptable risk; up to €15 million or 3% of turnover for other infringements; up to €7.5 million or 1.5% of turnover for supplying incorrect information. These ceilings are halved for SMEs and start-ups.
The AI Act is the European regulation on artificial intelligence that came into force in 2024. It establishes a regulatory framework based on the risk level of AI systems: systems posing an unacceptable risk are prohibited (manipulation, social scoring); high-risk systems (health, justice, HR, education) are subject to strict obligations of transparency, documentation and human oversight; other systems are subject to lighter obligations. Any company developing or deploying AI systems within the European Union is affected.
Nothing prohibits it in principle, but three limits apply. Confidentiality, first: Article L. 2132-1 of the French Public Procurement Code (code de la commande publique) prohibits the buyer from disclosing confidential information submitted by candidates. Uploading technical proposals to a consumer generative AI tool whose terms of use allow inputs to be reused therefore exposes the buyer to a clear breach of trade secrets: this is the central risk of shadow AI in purchasing departments. Second, the principles of public procurement: freedom of access, equal treatment and transparency require the buyer to retain control over, and understanding of, the tool it uses, and to ensure that algorithmic pre-screening does not systematically disadvantage certain types of operator. Finally, the decision must remain a human one: AI-assisted analysis cannot on its own constitute the choice of the successful tenderer, and the rejection of a tender must remain reasoned and explainable before the court. Drafting an internal policy on the use of AI can be the first step towards securing the process.
These are AI systems whose use may have a significant impact on health, safety or fundamental rights. In the public sector, this includes in particular AI used for migration management, access to education, assessing eligibility for social benefits, or the use of AI by law enforcement (biometric recognition).

Whether you are a start-up developing an innovative algorithmic solution or a local authority seeking to secure the procurement and deployment of AI in its services, NOVLAW is your strategic partner.