AI clause in a SaaS contract: the 6 key points to include

1. Describe the AI function precisely

An AI system should never be described generically. The annex must specify the purpose of the function, the intended uses, the excluded uses and the exact role of the AI in decision-making. This precision protects the provider (it limits its liability to the intended uses) and frames the client's expectations.

2. Clarify the allocation of responsibilities

The AI Act distinguishes between provider and deployer, each with their own obligations. The contract must clarify who does what: the provider guarantees compliance 'by design' (documentation, testing, transparency); the deployer ensures compliant use (human oversight, informing individuals, training). See provider or deployer.

3. Organise transparency and access to documentation

Article 13 requires the provider to supply comprehensible documentation and instructions for use. This documentation may be made available via a trust centre, on request, subject to confidentiality. This approach protects your intellectual property whilst satisfying the transparency obligation.

4. Specify human oversight

As soon as an AI system influences a decision, the contract must specify whether human validation is required, who is responsible for it and when it takes place. The more critical the system, the more this requirement becomes structural: it is a reminder that the tool assists the decision, without substituting for it.

5. Frame data and training

Even if GDPR-compliant, you must state in black and white that the prompts sent by the client are not used to train the models, that the generated results are not fed back into the training data, and that any secondary use requires express agreement. This is a point increasingly scrutinised by buyers.

6. Set limits and guarantees

Results generated by AI are inherently uncertain. The contract must set clear limits: the provider does not guarantee the accuracy or completeness of the content generated; the use of this content remains the client's responsibility. Also allow for the possibility of modifying or replacing the underlying model, without renegotiation at every change.

Methodological point: an AI clause cannot be copied and pasted. It must be custom-built, according to the role of AI in your product, its level of autonomy, the type of content generated and the permitted uses.