Back to Blog
Compliance21. november 202612 min

What Is a GDPR DPIA for AI Systems — and How to Pre-Fill One From Your Model Register

A DPIA for AI systems under GDPR Article 35 is mandatory in most enterprise use cases. Learn when it applies and how a model register pre-fills it.

What a DPIA Is and Why AI Makes It Unavoidable

A Data Protection Impact Assessment is the process GDPR Article 35 requires a controller to run before starting processing that is likely to result in a high risk to the rights and freedoms of natural persons. It is not a form to be filed after launch. It is a pre-deployment risk analysis that must describe the processing, assess its necessity and proportionality, evaluate the risks to individuals, and document the measures chosen to address them. Where residual risk remains high, Article 36 requires prior consultation with the supervisory authority.

AI systems trigger this duty more often than almost any other category of technology. Article 35(3) names three scenarios, and two of them describe routine AI behaviour: systematic and extensive evaluation of personal aspects based on automated processing, including profiling, where decisions produce legal or similarly significant effects; and large-scale processing of special category data. The European Data Protection Board's guidelines add nine criteria, including innovative use of new technology, evaluation or scoring, and processing of data about vulnerable subjects. Meeting two of them generally means a DPIA is required. A generative AI assistant connected to HR records, or a scoring model in credit or recruitment, will meet several at once.

The practical consequence is that any enterprise running more than a handful of AI systems that touch personal data is carrying a DPIA obligation per system, not one per company. Most organisations discover this late, when a data protection officer asks for the assessment on a tool that has already been live for six months. The question is no longer whether you need DPIAs for AI, but how to produce them at the pace at which AI is being adopted.

When a DPIA Is Required for an AI System

The trigger is likelihood of high risk, assessed on the nature, scope, context and purposes of the processing. National supervisory authorities publish lists of processing operations that always require a DPIA, and nearly all of them include some form of AI or machine learning use. The CNIL, the Dutch AP, the German DSK and the ICO all reference innovative technology combined with personal data as a trigger. Treat those lists as a floor.

In practice, a DPIA is almost certainly required when an AI system makes or materially influences decisions about employment, credit, insurance, education, healthcare or access to services; when it profiles individuals or predicts behaviour; when it processes biometric or health data; when it monitors employees or customers systematically; or when it ingests personal data at scale to train or fine-tune a model. It is also likely required when a general-purpose assistant is given access to mailboxes, document stores or CRM data, because the scope of personal data exposure is broad and hard to bound.

There are cases where a DPIA is not required, such as an internal drafting tool used on non-personal content with enforced data controls. But the burden of proof sits with you. The defensible approach is to document a screening decision for every AI system, recording why a full DPIA was or was not triggered. That screening record is itself evidence of accountability under Article 5(2), and it is the first thing a regulator will ask to see.

What a DPIA for AI Must Contain

Article 35(7) sets out four minimum elements. First, a systematic description of the envisaged processing operations and the purposes, including the legitimate interest pursued where relevant. Second, an assessment of the necessity and proportionality of the processing in relation to those purposes. Third, an assessment of the risks to the rights and freedoms of data subjects. Fourth, the measures envisaged to address those risks, including safeguards, security measures and mechanisms to demonstrate compliance.

For AI, each element needs more depth than a conventional system assessment. The description must cover the model, its provider, the hosting location, the training and inference data flows, any fine-tuning on your data, retention of prompts and outputs, and sub-processors in the chain. Necessity and proportionality must address whether a less intrusive method could achieve the purpose. Risk assessment must cover AI-specific harms: inaccurate outputs about real people, discriminatory outcomes, memorisation and leakage of personal data, inability to explain a decision, and function creep when a tool is repurposed. Article 22 matters whenever a decision is solely automated with significant effects, and the assessment should record whether human review is meaningful or merely nominal.

Measures should be specific and owned. Saying that data is encrypted is not a measure against model bias. Good DPIAs tie each identified risk to a named control, an owner, and a review date. They also record the view of the DPO, as Article 35(2) requires, and where appropriate the views of data subjects or their representatives under Article 35(9).

Why Most AI DPIAs Fail in Practice

The typical failure is not legal misunderstanding. It is information starvation. The person writing the DPIA, usually a privacy analyst or a business owner, does not know which model version powers the tool, where inference runs, whether the vendor trains on customer inputs, which sub-processors are involved, or what data categories the integration can reach. They email the vendor, wait weeks, paste fragments from a security questionnaire, and produce a document that is out of date the moment the vendor changes a model or a hosting region.

The second failure is duplication. The same vendor is assessed separately by three departments, each producing a different description of the same processing. The third is disconnection: the DPIA lives in a document library, while the actual AI estate lives in procurement records, IT asset lists and the shadow tools employees adopted on their own. No one can answer the simple question of which AI systems have a DPIA and which do not.

Underneath all three is a structural issue. Compliance tools that treat the DPIA as a standalone questionnaire start from a blank page. But roughly two thirds of the content in an AI DPIA is facts about the system, the vendor and the data flows, and those facts already exist somewhere in the organisation if you have bothered to record them. The assessment should be an act of judgement applied to known facts, not an act of discovery repeated for every system.

The Model Register as the Source of Truth for DPIA Inputs

A model register, sometimes called an AI vendor and model register, is the structured inventory of every AI system, model and vendor in use, with owners, purposes, data categories, hosting, contractual status and lifecycle stage. Built properly, it contains almost everything Article 35(7)(a) asks you to describe: the processing operation, the purpose, the provider, the model, the data categories, the recipients, the transfer mechanism and the retention terms.

This is the model-register-first approach to DPIAs. Instead of starting with an empty template and hunting for facts, you start from the register entry and let the assessment inherit those facts. Fronterio implements this directly. When a system is registered, its vendor, model, data classification, hosting region, sub-processor list and business owner are captured once. When the system crosses the screening threshold, Fronterio generates a GDPR Article 35 DPIA pre-filled from that register entry, with the systematic description, data flow summary, transfer basis and processor details already populated and linked back to their source.

The advantage compounds over time. When a vendor changes a hosting region or adds a sub-processor, the register changes, and every DPIA that depends on that vendor is flagged for review. The DPIA stops being a snapshot and becomes a living document tied to the real estate. Pre-filling does not remove the need for judgement. It moves human effort to where it belongs: assessing necessity, weighing risk and deciding on mitigations.

Aligning the DPIA With the EU AI Act: FRIA and Article 26

Enterprises deploying AI in the EU now face two overlapping assessments. The DPIA under GDPR Article 35 addresses risks to personal data and data subject rights. The Fundamental Rights Impact Assessment under EU AI Act Article 27 applies to certain deployers of high-risk systems, including public bodies and private entities providing public services, and deployers of credit scoring and life and health insurance pricing systems. Article 27(4) explicitly recognises the overlap: where a DPIA already covers an obligation, the FRIA complements it rather than repeating it.

Article 26(9) goes further, requiring deployers of high-risk systems to use the information provided by the provider under Article 13 to carry out their DPIA where applicable. That is a direct statutory link between the provider's technical documentation and your data protection assessment. If your model register already holds the provider's instructions for use, intended purpose and known limitations, the DPIA can draw on them without a separate collection exercise.

The efficient design is one evidence base feeding both assessments. The register supplies system facts, the DPIA analyses data protection risk, and the FRIA extends the analysis to non-data rights such as non-discrimination, access to services and human dignity. Fronterio's FRIA wizard reuses the same register inputs and DPIA conclusions, so the two documents stay consistent and a change in one prompts review of the other. Duplicated, divergent assessments are an audit liability. Aligned ones are a strength.

A Practical Workflow: From Register Entry to Approved DPIA

Start by making registration mandatory. No AI system should be procured, enabled or connected to personal data without a register entry, and discovery of unsanctioned tools should feed the same register. Capture at minimum the purpose, owner, vendor, model, data categories, data subjects, hosting and transfer details, and whether outputs influence decisions about people.

Second, run a screening step against Article 35 triggers and your supervisory authority's list, recording the outcome for every system. Where a DPIA is required, generate the draft from the register so that the description, data flows and processor details arrive pre-populated. Third, route the draft to the business owner to confirm purpose and proportionality, to security to validate technical measures, and to the DPO for advice, which must be documented. Fourth, record residual risk. If it remains high after mitigation, initiate Article 36 prior consultation before go-live.

Finally, tie the DPIA to monitoring. Set review triggers for model changes, new data sources, new purposes, vendor incidents and regulatory updates, and schedule periodic reviews regardless. Where the AI Act applies, link to post-market monitoring and incident reporting so that a serious incident prompts reassessment of both the DPIA and the FRIA. The goal is a DPIA that is current on the day an auditor asks for it, not the day it was signed.

Frequently asked questions

Is a DPIA mandatory for AI systems under GDPR?

It is mandatory whenever the processing is likely to result in a high risk to individuals, under Article 35. Many AI uses meet that threshold, including profiling, automated decisions with significant effects, large-scale special category data and innovative technology applied to personal data. Supervisory authority lists often name AI explicitly. Not every AI tool needs a full DPIA, but you should document a screening decision for each one so you can show why you concluded a DPIA was or was not required.

What is the difference between a DPIA and a FRIA?

A DPIA, under GDPR Article 35, assesses risks to personal data and data subject rights from processing. A FRIA, under EU AI Act Article 27, assesses the impact of certain high-risk AI systems on fundamental rights more broadly, such as non-discrimination and access to services. Article 27(4) allows a FRIA to complement an existing DPIA where obligations overlap, so the two should share evidence rather than be written separately.

Who is responsible for carrying out a DPIA for an AI system?

The controller is responsible, which usually means the deploying organisation, not the AI vendor. The controller must seek the advice of the data protection officer where one is designated, and the DPO should monitor performance. In practice the business owner, security, legal and the DPO all contribute. Vendors must assist as processors under Article 28, and under the AI Act, Article 26(9) requires deployers to use provider information to support the DPIA.

When do you need to consult the supervisory authority about an AI DPIA?

Under Article 36, you must consult the supervisory authority before processing when the DPIA indicates high residual risk that you cannot mitigate sufficiently. The authority has up to eight weeks to respond, extendable by six, so this must be planned before launch dates are committed. If you can reduce the risk to an acceptable level through technical and organisational measures, prior consultation is not required, but the reasoning should be documented.

Can a DPIA be pre-filled automatically for AI tools?

Yes, for the factual sections. If you maintain a model register recording vendor, model, data categories, hosting, sub-processors and purpose, those details can populate the systematic description and data flow sections of the DPIA. Judgement-based sections, such as necessity, proportionality and risk evaluation, still require human review and DPO input. Pre-filling removes the discovery work, not the accountability.

How often should a DPIA for an AI system be reviewed?

GDPR requires review when the risk represented by processing changes, and the EDPB recommends continuous review. For AI, triggers include a model version change, a new data source, a new purpose, a change of vendor hosting or sub-processors, a security incident, or a material shift in output accuracy. Set a scheduled review at least annually, and connect register changes to automatic review flags so nothing relies on memory.

Ready to get started?

Fronterio helps you implement everything discussed in this article, with built-in tools, automation, and guidance.