Drafting a DPA for AI Tools That Process Your Data (2026)

What a DPA is and why AI tools need a special one A data processing addendum, or DPA, is the contract that governs how a service provider handles personal data on your behalf. Under the EU and UK GDPR, when you (the controller) let a vendor (a processor) process personal data for you, Article 28 requires a written contract setting out the subject matter, duration, nature and purpose of the processing, the type of data and categories of people, and a specific list of the processor's duties. Similar duties exist under other laws, such as the CCPA and CPRA in California. A DPA is how you meet that legal requirement. An AI tool needs more than a standard DPA because AI introduces uses of data that older templates never imagined: using your data to train or improve a model, generating new data from it, long-lived embeddings and logs, and foundation-model providers sitting as sub-processors behind the tool. A DPA for an AI tool keeps the Article 28 core and adds terms that nail down these AI-specific uses. This page is general information, not legal advice. Data-protection law and the drafting that satisfies it differ by jurisdiction and change over time, so have a qualified professional adapt the document for a live deal. ## The Article 28 core your DPA must contain Every GDPR-aligned DPA, AI or not, must require the processor to: - process personal data only on your documented instructions;

• ensure people who process the data are bound by confidentiality;
• apply appropriate security measures under Article 32;
• not engage sub-processors without authorisation, and flow equivalent terms down to them;
• assist you with data-subject requests and with your own security, breach and impact-assessment duties;
• notify you of a personal-data breach without undue delay;
• delete or return personal data at the end of the service; and
• make available the information needed to show compliance and allow audits. Set out the subject matter, duration, nature and purpose, the types of personal data and the categories of data subjects, usually in a schedule. ## The AI-specific terms to add 1. No training on your personal data without consent. State plainly that the processor will not use your personal data to train, fine-tune or improve its own or any third party's models unless you have given documented, specific authorisation. This is the clause that most often separates an AI DPA from a generic one. 2. Outputs and derived data. Say how embeddings, vectors, logs and other data derived from your personal data are treated, how long they are kept, and that they are deleted or returned with the source data. Derived data is easy to forget and can persist long after you think the data is gone. 3. Sub-processors, including model providers. Require a current list of sub-processors that names the foundation-model providers behind the tool, with the right to object to new ones and to be told where data is processed. 4. International transfers. Where personal data leaves its home jurisdiction, require a valid transfer mechanism, such as the EU Standard Contractual Clauses or the UK addendum, and a transfer-risk assessment. AI tools often route data through model providers in other countries, so this clause does real work. 5. Automated decision-making. If the tool makes or supports decisions about people, address the GDPR Article 22 rules on automated decisions, the information you must give people, and the support the processor will provide so you can meet those duties. 6. Security specifics and breach timelines. Beyond the Article 32 reference, set concrete breach-notification timelines and the security measures that matter for an AI pipeline, including access controls on training and inference data. 7. Deletion and exit. Require deletion or return of personal data and all derived artefacts within a set period at the end of the service, with written confirmation, and say what, if anything, the processor may retain and why. ## A clause checklist you can adapt ``` AI DATA PROCESSING ADDENDUM - CHECKLIST [ ] Roles stated (controller / processor) and processing described in a schedule (subject matter, duration, nature, purpose, data types, data subjects) [ ] Processing only on documented instructions [ ] No training / fine-tuning on personal data without documented consent [ ] Confidentiality of personnel [ ] Article 32 security measures, plus concrete controls for the AI pipeline [ ] Sub-processor list including model providers; authorisation and flow-down [ ] International transfer mechanism (SCCs / UK addendum) and TRA [ ] Breach notification without undue delay, with a stated timeline [ ] Assistance with data-subject rights and Article 22 automated decisions [ ] Treatment, retention and deletion of embeddings, logs and derived data [ ] Audit and information rights [ ] Deletion or return of all personal data and derived artefacts on exit, with written confirmation [ ] Allocation of liability and indemnity consistent with the main contract

- Signing a DPA does not make the processing lawful on its own. You still need a lawful basis for the processing, and the tool still has to be appropriate for the data.
- A no-training clause records a promise; verifying it is honoured is a matter of the processor's assurances, audits and track record, not the clause alone.
- A DPA cannot cure a tool that is fundamentally unsuitable for sensitive data. Sometimes the right answer is not to put the data in the tool at all. ## Keeping DPAs consistent across every AI vendor The real discipline is not drafting one good AI DPA; it is making sure every vendor that touches personal data has signed an equivalent one, that the no-training, sub-processor and transfer terms do not drift between suppliers, and that when a regulator or a data-subject request lands you can find which version each vendor signed and what it covers. Drafting from a clean DPA, spotting where a vendor has weakened the training restriction or the deletion duty, editing a clause in plain language across several addenda, and holding every version with its date can live in one AI workspace tuned to how your company contracts. [Adira](https://adiralaw.com) is built around that single-workspace idea for business contracts. For a single tool the checklist above is enough to work from; the point worth keeping is that a data-protection obligation only protects people if it is applied to every vendor and kept in one place.