Contracting a legal AI provider: the ten clauses no firm should sign without reviewing

Contracting a legal AI tool is not buying just another program. The provider will process client information subject to professional secrecy, and the firm remains the data controller under the GDPR. The difference between good and bad contracting lies in the clauses. These are the ten no firm should sign without reading twice.

1. Processing agreement under article 28 GDPR

The contract must incorporate a complete processing agreement: subject matter, duration, nature and purpose of the processing, type of data and categories of data subjects, and the processor's obligations. Without this annex, the firm breaches the GDPR from day one, regardless of the tool's quality.

2. Prohibition on training models with the firm's data

The clause that generates the most discussion: the firm's documents and queries cannot be used to train or improve models, neither the provider's own nor third parties'. It must appear expressly; contractual silence usually plays against the client. We connect this with the protection of secrecy in AI and professional secrecy: client data.

3. Data residency and transfers

Where the data is hosted, where queries are processed and whether there are international transfers. EU residency simplifies compliance; if there are transfers to third countries, the contract must identify the mechanism that covers them.

4. Subprocessors and third-party infrastructure

Almost no AI provider processes everything on its own servers: demand the list of subprocessors (cloud, foundation models, support), the right to object to changes and the guarantee that the data flow does not expand without notice.

5. Concrete security measures

Certifications (ISO 27001, ENS where applicable), encryption in transit and at rest, access management, activity logs and a commitment to notify breaches within short deadlines. "State-of-the-art security" without specifics is not a clause, it is advertising.

6. System classification under the AI Act

Regulation (EU) 2024/1689 gradates obligations by risk. The contract should reflect what class of system is being contracted, which provider obligations are already operative and how compliance responsibilities are shared. Our summary of the calendar is in The AI Act, the digital omnibus and the calendar for firms.

7. Ownership and licensing of outputs

Who owns the generated texts, under what licence the firm uses them and whether the provider can reuse them. It seems obvious, but it is not: some standard terms reserve exploitation rights over the results.

8. Liability for errors and infringing content

Liability limits in standard terms are usually heavily reduced. It is worth negotiating, at least, coverage of third-party claims for intellectual property infringement and a reasonable cap for damages from system errors.

9. Audit and transparency

Right to audit or to receive independent reports, access to the system's technical documentation and a commitment to report substantial model changes. The firm's internal AI policy, the AEPD policy and its application in firms, requires being able to justify these verifications.

10. Orderly exit

Reversibility: full export of data and histories in standard formats, certified deletion at termination and retention periods. Changing providers must not mean losing the documentary trail of matters.

Before signing

The CCBE guide on generative AI, which we summarised in The CCBE guide on generative AI for lawyers, recommends exactly this: vendor due diligence, contractual review and internal policy. Three steps, none optional. All our coverage, in the Legal AI section.