AI Data Processing Agreement (DPA)

How to use this document

• What it is for. This is a base contract addendum between a customer organization and an AI service vendor that processes personal data on the customer's behalf. It contains the usual terms of a data processing agreement and adds what is specific to AI: a ban on training models with customer data, a vendor report on the system's functioning, limitations and risks, documentation for responsible use, sub-processors (including model providers), international transfers, incidents and deletion of data from models and vector stores.
• Who uses it. It serves both the customer buying AI and the vendor selling it. If your organization is the vendor, adapt the clauses to your actual service and do not offer commitments your architecture cannot meet.
• Who should own it. The legal function, with the DPO or privacy function, information security and procurement. Enter the function in [RESPONSIBLE FUNCTION].
• Decisions to take first. Decide which option for using data to improve the service you accept (clause 5), the incident notification deadline (clause 13), the sub-processor authorization model (clause 8), what documentation you require from the vendor (clause 11 and Annex IV) and which optional clauses you include (clause 19).
• Fields to complete. Replace the bracketed fields and complete Annexes I to IV. [OPTIONAL] clauses are proposals for your counsel to assess; remove the tag when keeping them.
• Country variants. Keep the variants for the countries whose data protection law applies to the processing (because of where the customer, the data subjects or the vendor are) and add their content to the text or as an annex. Delete the rest.
• Review. A lawyer should review this addendum before signature, especially governing law, liability, transfer mechanisms and how it fits with the main contract and the vendor's standard terms.
• Signature and retention. Sign it with the main contract or as an amendment, keep the signed version and up-to-date annexes, and record the vendor in the AI system inventory and the record of processing activities.

AI Data Processing Agreement (DPA)

Field	Content
Customer (controller)	[COMPLETE: legal name, address and tax ID of the customer]
Vendor (processor)	[VENDOR], [COMPLETE: address and tax ID]
Main contract	[COMPLETE: name and date of the services or license agreement]
AI service	[AI SYSTEM]
Effective date	[DATE]
Customer privacy contact	[PRIVACY EMAIL]
Vendor privacy and security contact	[COMPLETE]

1. Purpose, scope and precedence

1.1 This agreement governs the processing of personal data that the Vendor carries out on behalf of the Customer when providing the AI service described in the main contract and in Annex I.

1.2 It forms part of the main contract. In case of conflict on data processing, use of data for training or security, this agreement prevails, unless the main contract is more protective of data subjects.

1.3 The Vendor's general terms, usage policies or online documents do not amend this agreement unless the Customer expressly accepts them in writing.

2. Definitions

• Customer Data: all information that the Customer or its users enter, upload or make available to the service, including personal data.
• Inputs: prompts, queries, files, documents, audio, images and other content sent to the AI system.
• Outputs: responses, content, scores, classifications, recommendations and other results the AI system generates from Inputs.
• Derived Data: embeddings, indexes, metadata, logs, user ratings and any other data derived from Inputs or Outputs.
• Train: train, retrain, fine-tune, distill, evaluate or otherwise improve an AI model.
• General Model: any model of the Vendor or of third parties that is or may be used to provide services to other customers.
• Sub-processor: any third party that processes the Customer's personal data on the Vendor's instructions, including AI model providers, cloud infrastructure providers and monitoring tool providers.
• Security Breach: any breach of security leading to the destruction, loss, alteration, unauthorized disclosure of or access to the Customer's personal data.
• AI Incident: a malfunction, unexpected behavior, manipulation or misuse of the AI system that causes or may cause significant harm to the Customer or to people, even if it is not a Security Breach, for example systematically wrong or biased Outputs, disclosure of data through Outputs or compromise of the model through prompt injection.
• Applicable Law: the data protection and AI laws applicable to the processing, including those indicated in the country variants the parties keep.