Our Expert in France
AI vendor contracts France buyers sign in 2026 carry a distinct legal fingerprint: French copyright law (droit d’auteur), the EU sui generis database right, GDPR obligations enforced by the CNIL, and the phased obligations of the EU AI Act all converge in a single procurement decision. In-house counsel, procurement leads and product owners negotiating access to AI models, datasets and SaaS integrations must now allocate risk across intellectual property, data protection and product-liability domains simultaneously. This guide sets out practical drafting positions, sample clause language and negotiation redlines tailored to French law, so that legal and procurement teams can move from a vendor’s standard paper to a defensible, France-ready contract.
Every recommendation below is framed as a drafting position, not a legal conclusion, adapt each clause to your facts and take advice before signing.
This content is for general information and does not constitute legal advice.
Busy readers should focus on five immediate risks before signing any AI vendor contract in France:
Before drafting any AI vendor contracts France teams should understand the legal architecture that governs the deal. French AI contracting sits at the intersection of the Code de la propriété intellectuelle, the Code civil, the GDPR, CNIL guidance and the EU AI Act. Each layer imposes obligations that cannot be waived away by boilerplate.
The Code de la propriété intellectuelle (CPI) governs copyright, the exclusive rights of reproduction and representation, and, critically for France, the author’s moral rights (droit moral). Moral rights in France are generally regarded as perpetual, inalienable and imprescriptible, and cannot be fully waived in advance, which affects any clause purporting to secure unrestricted downstream use of protected works within training data or outputs. On the contractual side, the Code civil governs the formation of contracts, remedies for breach, and the enforceability of limitation-of-liability clauses. French courts scrutinise liability caps that would deprive an essential obligation of its substance, so caps must be drafted with care to remain enforceable.
The EU sui generis database right (droit sui generis des bases de données), derived from Directive 96/9/EC and implemented in French law within the CPI, protects the substantial investment made in obtaining, verifying or presenting the contents of a database, independently of any copyright in the underlying content. For AI training, this matters enormously: extracting and reusing a substantial part of a protected database to train a model can infringe the database maker’s rights, subject to applicable exceptions (including the text-and-data-mining exceptions introduced by Directive (EU) 2019/790 and transposed into French law).
Your vendor’s IP warranty should therefore expressly cover lawful extraction and reuse under any applicable database right, and require evidence of consent or a licence for third-party databases incorporated into the training corpus.
Where training data includes personal data, the GDPR (Regulation (EU) 2016/679), together with the French Loi Informatique et Libertés, requires a lawful basis for processing, and a Data Protection Impact Assessment (DPIA) where the processing is likely to result in high risk to individuals. The CNIL, France’s data protection regulator, has published guidance (“how-to sheets”/recommendations) on artificial intelligence and data protection setting out its expectations on lawful basis, data minimisation, documentation of provenance and data subject rights. Layered on top, the EU AI Act (Regulation (EU) 2024/1689) introduces obligations calibrated to risk classification, including transparency and technical documentation, and, for high-risk systems, conformity requirements for providers, phased in over a transition period following its 2024 entry into force.
A well-drafted contract should allocate these obligations expressly between vendor and buyer rather than leaving them to statutory default.
IP warranties are the backbone of any AI vendor contract. In France, generic “the vendor warrants it has all necessary rights” language is inadequate because it does not engage the specific French rights, copyright, neighbouring rights, database rights and moral rights, that can be asserted against a deployed AI system. The goal is to obtain specific, actionable representations with meaningful remedies.
At a minimum, the buyer should extract representations that the vendor:
Model clause, sample language (for negotiation only, adapt to facts): “The Supplier represents and warrants that (i) it holds all rights, licences and consents necessary to provide the Model and Training Data and to grant the rights set out in this Agreement; (ii) the Model, Training Data and Outputs do not and will not infringe any third-party intellectual property right, including any copyright, neighbouring right (droits voisins) or sui generis database right under French or EU law; and (iii) all open-source components are used in compliance with their applicable licences.”
Many AI vendors aggregate third-party data and pre-trained components. The buyer needs assurance that the vendor’s own upstream licences permit sublicensing to the buyer for the buyer’s intended uses, including internal training, fine-tuning and commercial inference. Require the vendor to warrant the scope of upstream licences and to flag any field-of-use or territorial restrictions. Where the vendor cannot warrant a clean chain, negotiate a corresponding narrowing of your permitted uses or an enhanced indemnity.
Open-source software (OSS) is ubiquitous in AI stacks, and licence non-compliance, for example, failing to meet copyleft obligations under the GPL, or attribution requirements under Apache 2.0, creates real legal exposure. Ask the vendor to deliver a software bill of materials (SBOM) listing all OSS components and their licences, and to warrant that its use complies with each licence. Where copyleft licences appear in components that touch your proprietary code, insist on architectural separation or a written explanation of why no copyleft obligation is triggered.
Warranties are only as good as their remedies. Tie a breach of the IP warranties to a layered remedy: first, the vendor must, at its cost, procure the right to continue use, modify the deliverable to make it non-infringing, or replace it with equivalent non-infringing functionality; and second, the vendor must indemnify the buyer for third-party claims (see Section 6). For business-critical models, consider a source-and-documentation escrow so that the buyer can maintain the system if the vendor defaults or becomes insolvent.
Training-data licensing is where IP, data protection and provenance risk concentrate. In AI vendor contracts France buyers should treat the data licence as a distinct, carefully scoped grant, not an afterthought bundled into the software licence.
Define the licence with precision along several axes:
Model clause, sample language (for negotiation only, adapt to facts): “The Supplier grants the Customer a non-exclusive, worldwide, royalty-free licence to use the Training Data solely for the purpose of training, fine-tuning, validating and operating the Model, and to use resulting Outputs for the Customer’s internal and commercial purposes, subject to the Supplier’s warranties in Clause [X].”
Provenance is the chain-of-title for data. The CNIL’s guidance on artificial intelligence and data protection stresses documenting the sources of training data, the legal basis for processing personal data, and records of consent where consent is relied upon. Require the vendor to warrant the provenance of every dataset and to deliver supporting records: data source registers, licence agreements with upstream providers, consent records where applicable, and any records of web-scraping methodology and the rights basis on which it relied. Inadequate provenance should be a defined breach with specific remedial obligations.
Warranties without verification are weak. Negotiate a right to inspect provenance records, to sample dataset records, and to review processing logs, and the ability to appoint an independent auditor bound by confidentiality, on reasonable notice. Specify what happens if an audit reveals inadequate provenance: rectification, replacement of the offending data, re-training at the vendor’s cost, or termination with a refund.
Where personal data is involved, the contract must operationalise data subject rights under the GDPR, including access, rectification and erasure. Address the practical difficulty of “unlearning” personal data from a trained model: require the vendor to describe its process for honouring erasure requests, to delete personal data from active training corpora, and to document mitigations where full removal from model weights is technically infeasible. Allocate controller and processor roles clearly, and include instructions consistent with Article 28 of the GDPR where the vendor acts as a processor.
One of the most contested questions in AI contracting is deceptively simple: who owns the model? The answer in France is contractual, there is no statutory default that automatically transfers ownership of a model trained on a customer’s data to the customer. Silence generally favours the vendor.
There are three principal structures. First, the vendor retains ownership of the model but grants the customer a broad licence, ideally covering training, inference, export and continued use after termination. Second, the vendor assigns the model weights to the customer, transferring ownership outright. Third, the parties agree joint ownership, which sounds equitable but creates operational friction over exploitation and improvement rights. For most buyers, an explicit, broad and irrevocable licence delivers the commercial benefits of ownership without the negotiation cost of an assignment.
Distinguish carefully between the model architecture, the trained parameters or weights, and the outputs the model generates. A buyer may negotiate rights to the weights (the trained artefact reflecting the buyer’s data) even where the vendor retains its underlying architecture and platform. Outputs raise separate questions: address ownership of, and rights to use, generated content, and confirm that the buyer is free to use outputs commercially without further royalty. Note the French copyright constraint that droit d’auteur requires a human author and an original creation reflecting the author’s own intellectual creation; purely machine-generated outputs without sufficient human creative input may not attract copyright protection, a factor for buyers relying on IP protection for AI-generated deliverables.
If ownership is not transferred, “extractability” becomes vital: the practical right to export the trained model, weights and associated documentation on termination so the buyer is not locked in. Draft an express extractability clause specifying formats, timelines and the technical assistance the vendor must provide. A negotiation checklist for model rights should confirm: licence scope, weight export rights, permitted derivatives, post-termination continuation, and whether the vendor may reuse insights derived from the buyer’s data to improve models sold to competitors.
Liability allocation is where AI vendor contracts France negotiations most often stall, because the risks, third-party IP claims arising from training data, and downstream harm caused by model outputs, are novel and potentially large. The Code civil governs the enforceability of limitation-of-liability clauses, and French courts will generally not enforce a cap that empties an essential contractual obligation of meaning.
Identify the risk events that must be addressed: IP infringement arising from the training data or model; personal-data breaches; and harm caused by erroneous or unsafe model outputs. Each requires a defined trigger. IP infringement and data-protection breaches should typically be carved out of the general liability cap or subject to a higher, separate cap, because they can generate open-ended third-party exposure.
A warranty gives the buyer a claim for breach; an indemnity provides a direct promise to cover defined losses, often including third-party claims, without the buyer needing to prove breach in the same way. For third-party IP infringement claims, insist on an indemnity that covers defence costs, settlements and damages, with the vendor conducting or funding the defence subject to the buyer’s reasonable involvement.
Negotiate the overall liability cap by reference to contract value (for example, a multiple of annual fees), with a higher or unlimited cap for IP indemnities, data-protection breaches, confidentiality breaches, and losses arising from gross negligence (faute lourde) or wilful misconduct (faute dolosive), categories that French law generally will not permit a party to exclude by contract. Use a basket or threshold to filter trivial claims, but ensure it does not apply to indemnified third-party claims.
Contractual promises are only as good as the vendor’s balance sheet. Require the vendor to maintain professional indemnity / errors-and-omissions (E&O) cover and cyber insurance with defined minimum limits, to provide certificates on request, and, where appropriate and available under the relevant policy, to name the buyer as an additional insured. Insurance backstops the indemnity if the vendor cannot pay.
| Mechanism | What it covers | Pros for buyer | Limitations | Recommended drafting points |
|---|---|---|---|---|
| Indemnity for IP infringement | Third-party claims that the model or training data infringes IP, including defence costs, settlements and damages | Direct, loss-shifting promise; can cover claims without proving breach; vendor conducts/funds defence | Only as good as vendor’s solvency; may be capped or carved down; conduct-of-claims disputes | Carve out of general cap or set higher sub-cap; cover defence costs; define notice and conduct-of-claim procedure; include replacement/procure-rights remedy |
| Contractual liability cap | Caps the vendor’s aggregate financial exposure for breach and general liability | Predictable exposure; standard commercial mechanism | May be unenforceable if it guts an essential obligation; excludes categories French law won’t allow to be capped | Link to contract value; exclude gross negligence, wilful misconduct, IP and data breaches; ensure the cap does not defeat the essential obligation |
| Professional indemnity / E&O insurance | Insurer-backed cover for the vendor’s errors, omissions and, via cyber cover, data incidents | Financial backstop independent of vendor’s balance sheet; supports indemnity recovery | Policy exclusions and limits; claims-made timing; insurer defences | Specify minimum limits; require certificates; consider named-insured status; align policy scope with indemnified risks |
Beyond the headline IP and liability terms, operational clauses materially reduce risk over the life of the deployment and preserve evidence for any future dispute.
Require the vendor to maintain audit logs, provenance records, model version histories and rollback capability. Model versioning lets the buyer identify which model produced a given output, and rollback allows a return to a known-good version if a new release introduces infringing or unsafe behaviour. Where regulatory compliance or safety demands it, include explainability and transparency obligations, documented model descriptions, known limitations, and access to model cards or technical documentation consistent with EU AI Act transparency expectations.
Model clause, sample language (for negotiation only, adapt to facts): “The Supplier shall maintain, for a period of not less than [X] years, complete audit logs of Model versions, training-data provenance and material configuration changes, and shall provide the Customer with the ability to roll back to the immediately preceding Model version on request.”
Impose security standards, timely patching obligations, and controls over subprocessors, including a right to approve or object to new subprocessors and to receive notice of changes, consistent with GDPR requirements where personal data is processed.
Define service levels, remedies for failure (service credits), and mitigation obligations, so that operational failures have contractual consequences short of termination and litigation.
Use a priority matrix to focus negotiation capital where it matters most:
Typical buyer-side redlines include: converting a bare “sufficient rights” warranty into an enumerated non-infringement warranty; adding sui generis database right and neighbouring rights to the infringement carve-out; moving IP and data-protection claims outside the general cap; and inserting an express extractability clause. A useful fallback where a vendor resists an uncapped IP indemnity is a materially higher sub-cap combined with mandatory E&O insurance and a procure/modify/replace obligation.
Drafting robust AI vendor contracts France buyers can rely on in 2026 requires treating IP warranties, training-data licensing, model ownership and liability allocation as interlocking components rather than isolated clauses. The recurring theme is specificity: enumerate the French rights at stake, demand documented provenance, define model and output rights expressly, and back warranties with meaningful indemnities and insurance. Legal and procurement teams should build a standard clause library, a provenance-evidence checklist, and a priority matrix so that each new AI procurement starts from a defensible baseline rather than the vendor’s paper. As an immediate next step, preserve logs and provenance records from day one to support any future dispute, and confirm your GDPR lawful basis and DPIA position before deployment.
Consult qualified Intellectual Property lawyers in France before finalising any agreement.
This article was produced by Global Law Experts. For specialist advice on this topic, contact Nathalie Marchand at d’Alverny Avocats, a member of the Global Law Experts network.
posted 2 minutes ago
posted 4 minutes ago
posted 23 minutes ago
posted 46 minutes ago
posted 1 hour ago
posted 1 hour ago
posted 2 hours ago
posted 2 hours ago
posted 2 hours ago
posted 3 hours ago
posted 3 hours ago
Find the right Legal Expert for your business
Send welcome message
