A standard Data Processing Agreement assumes a processor does one thing with personal data: stores it, moves it, maybe runs a calculation on it, and hands the same data back or deletes it when the relationship ends. That assumption breaks down the moment the processor is an AI or machine learning vendor, because the whole point of a model is to change based on what it sees. A payment processor that touched a customer’s data yesterday doesn’t behave differently today because of it. A model that trained on that data does, and a DPA written for the first kind of vendor has nothing to say about the second. If you haven’t confirmed you need a DPA with a given vendor at all, start with Do You Need a Data Processing Agreement?; this guide picks up once you already know the answer is yes and the vendor in question is an AI or ML tool.
Why a Standard DPA Doesn’t Cover What an AI Vendor Actually Does
Article 28(3)(a) of the GDPR requires that a processor handle personal data only on the controller’s documented instructions. For a conventional SaaS processor, that clause does most of the work on its own: the vendor stores what you send it, serves it back when asked, and doesn’t touch it for anything you didn’t authorize. Our broader clause-by-clause guide, What to Include in a Data Processing Agreement, covers that baseline: subject matter and duration, the sub-processor list, security measures, breach notice, audit rights, and international transfer terms.
An AI vendor’s commercial terms often reserve a use that a standard DPA never anticipates: using what you send it, the prompts, documents, or records containing personal data, to train or improve the underlying model. That’s not a data storage decision or a data transfer decision, the two things a conventional DPA is built to constrain. It’s a decision to let personal data shape a system that will later serve other customers, and “documented instructions” doesn’t automatically cover it just because the controller signed a DPA that never mentioned training in either direction.
Standard SaaS DPA vs AI/ML vendor DPA
| Standard SaaS DPA | AI/ML vendor DPA | |
|---|---|---|
| Data use scope | Limited to service delivery | Must exclude model training |
| Sub-processors | Hosting, email, analytics | + model API, human review |
| Retention | Input data only | + outputs, embeddings, logs |
| Deletion at termination | Primary data deleted | Must include cached outputs |
None of this means a standard DPA is wrong. It means it’s incomplete for this category of vendor, and the gap sits in three specific places: whether the vendor can use your data to train anything, how long it keeps what the model produces, and who else touches the data once it leaves the vendor’s own infrastructure.
Figure 1: The three places where a standard SaaS DPA says nothing about an AI or ML vendor. Source: GDPR Article 28.
The Model Training Prohibition Clause
This is the clause that matters most and the one a generic DPA template is least likely to have. It needs to say, in terms specific enough that there’s no room to read it either way, that the vendor will not use the data it processes, meaning both what the controller sends in and what the model produces in response, to train, fine-tune, retrain, or otherwise improve any model or product, for any customer other than the one the data came from, without that customer’s prior written consent.
The reason this has to be explicit rather than inferred from the general “process only on instructions” language is that many AI vendors’ default commercial terms run the other direction: free and lower tiers of a given product commonly reserve the right to use customer inputs for service improvement unless the customer opts out, while enterprise or API tiers more often default to not training on customer data unless the customer opts in. A DPA that’s silent on training doesn’t create a prohibition; it just leaves whatever the vendor’s underlying commercial terms already say, and those terms are written by the vendor, for the vendor’s benefit, not with your DPA’s Article 28 obligations in mind. If your business is a controller supplying personal data to an AI vendor as processor, that vendor’s default terms are exactly the gap the training-prohibition clause needs to close, in the DPA itself, not by reference to whatever the product terms happen to say this quarter.
Output Retention: Prompts, Completions, and Embeddings
A standard DPA’s retention and deletion clause covers the data the controller sent in. An AI vendor generates something new from it: a completion, a summary, a classification, or a vector embedding stored for later retrieval, and each of those can carry personal data forward even after the original input is deleted. Embeddings in particular are easy to overlook because they don’t look like personal data on the surface, a list of numbers rather than a name or an email address, but research on embedding inversion has repeatedly shown that enough of the original content can be reconstructed from a stored vector to make it personal data under most regulators’ reading of the term.
A DPA covering an AI vendor needs retention terms for each category of derived data separately: how long raw prompts and completions are kept for debugging or abuse monitoring (a short, defined window, not indefinite), whether embeddings persist in a vector store past the life of the underlying record they were built from, and what “deletion” actually means when a vector index doesn’t delete records the way a relational table does. The termination clause needs to say explicitly that deletion covers cached outputs and embeddings, not just the primary dataset, because a vendor that reads “delete customer data” narrowly will delete the input table and consider the job done while a vector store built from that same data quietly survives.
Training-use clause: generic vs AI-specific
- Processor uses data only per customer instructions
- No mention of model training or improvement
- Silent on whether outputs are new personal data
- Sub-processor list covers hosting and email only
- Explicit prohibition on training without consent
- Defines outputs and embeddings as in-scope data
- Sets retention limits for prompts and completions
- Names the underlying model provider as sub-processor
Sub-processor Disclosure for the Model Layer
Article 28(2) and 28(4) require a processor to disclose its sub-processors and give the controller a mechanism to object before a new one goes live, covered in general terms in What to Include in a Data Processing Agreement. AI vendors add a sub-processor category that’s easy for a vendor to leave off the list because it doesn’t look like a typical vendor relationship: the foundation model provider itself. A product that wraps a third-party model API under its own branding is routing your customers’ personal data through that model provider’s infrastructure whether or not the vendor’s sub-processor list mentions it, and “we treat the model API as internal infrastructure, not a sub-processor” isn’t a characterization that changes what Article 28 requires. If personal data reaches that provider’s systems, it’s a sub-processor, and it needs to be disclosed and covered by the same notice-and-objection mechanism as any other.
Figure 2: A foundation model provider that receives personal data is a sub-processor under GDPR Article 28(2) and 28(4).
The same applies to a second category vendors sometimes leave out: any third party doing human review of model outputs, for quality evaluation, safety moderation, or RLHF-style fine-tuning feedback. A person reading a flagged output containing a customer’s personal data is a form of processing that a controller has a right to know about and, in a lot of cases, a right to exclude by contract. Because a controller’s own international-transfer obligations also depend on where each of these sub-processors actually operates, foundation model providers are frequently outside the EU/EEA, this is also where Standard Contractual Clauses vs a DPA for SaaS becomes relevant: naming the model provider as a sub-processor only solves half the problem if that provider’s infrastructure sits somewhere without an adequacy decision and the DPA doesn’t also cover the transfer mechanism.
Putting the AI-Specific Terms Together
A DPA for an AI or ML vendor should read like a standard Article 28 agreement with three additions layered on: a training-use clause that closes the gap the vendor’s own commercial terms leave open, retention and deletion terms that reach outputs and embeddings and not just the original input, and a sub-processor list that names the model provider and any human-review vendor by their actual function rather than a euphemism for infrastructure. None of the three shows up automatically in a template built for a conventional SaaS processor, which is exactly why they’re worth checking for by name before signing rather than assuming a DPA labeled correctly covers them. Our DPA generator builds the underlying Article 28 agreement your AI vendor relationship still needs either way, giving you a complete document to layer these AI-specific terms onto rather than starting the whole agreement from a blank page.








