DPIA for AI: template and process

When is a DPIA for AI mandatory?

Art. 35(1) GDPR requires a DPIA where processing, “in particular using new technologies”, is likely to result in a high risk to the rights and freedoms of natural persons. According to the guidance of the German data protection conference (Datenschutzkonferenz, DSK), this “will often be the case” for AI. Three routes lead to the obligation:

• Art. 35(3): for example large-scale processing of health data (point b) or systematic evaluation of people with legal effects (point a).
• DSK mandatory list for the non-public sector: no. 11 names the “use of artificial intelligence to process personal data in order to steer interaction with the data subjects or to evaluate personal aspects”. No. 3 names large-scale processing of data subject to social, professional or special official secrecy. Public bodies follow the lists of their supervisory authority.
• Criteria from WP 248: if two of the nine criteria apply, a DPIA is usually required. With AI in a hospital or public authority, these often come together: confidential or highly personal data, vulnerable data subjects (patients, children, employees), innovative technology.

An internal writing aid without personal data does not need a DPIA. An assistant that drafts doctor’s letters, administrative decisions or social reports as a rule does.

What must the DPIA contain?

Art. 35(7) GDPR names four minimum contents:

Part	What you describe for AI
Systematic description	Purposes, types of data, data subjects, data flow from the input field to the AI model, storage locations, deletion periods, service providers involved
Necessity and proportionality	Why AI is needed for this purpose, data minimisation, legal basis, storage period
Risks to data subjects	for example wrong answers, disclosure to unauthorised persons, training by the provider, third-country access, monitoring of employees’ performance, decisions without genuine human review
Measures to address the risks	Permissions and groups, logs, server on your premises or data centre in the EU, exclusion of training, source references, training of staff, internal instructions

How does a DPIA work?

1. Threshold assessment: go through the mandatory list and the WP 248 criteria, document the result.
2. Description: record the areas of use, the data and the data flows. The provider supplies the details of the system; the DSK advises paying attention to this when selecting a provider.
3. Assessment: estimate the likelihood and severity of each risk. Many supervisory authorities recommend the Standard Data Protection Model (SDM) for this.
4. Measures: assign technical and organisational measures and assess the residual risk.
5. Opinion: seek the advice of the data protection officer (Art. 35(2)) and, where appropriate, the views of the data subjects or their representatives (Art. 35(9)).
6. Decision and review: if a high residual risk remains, consult the supervisory authority before going live (Art. 36). When something changes, such as a new AI model or new data sources, review the DPIA again (Art. 35(11)).

Which information do you need from the provider?

• Where the platform and the AI model run and who has technical access
• Whether inputs are stored, logged or used for training
• List of sub-processors, third-country links
• Permissions and roles concept, logging, deletion
• TOMs under Art. 32 and proof (certificates, attestations)
• For high-risk AI: the instructions for use under Art. 13 of the AI Act; deployers use them for their DPIA (Art. 26(9))

How does this work with basebox?

For AI with health or social data, a DPIA is as a rule required. You prepare it and are responsible for it as the controller. For the cloud and the server there is one template each in Word format, structured according to Art. 35 GDPR and the Standard Data Protection Model; the details on basebox are already filled in. In addition there are the TOMs, the list of sub-processors and a template for the record of processing activities, all on /sicherheit. basebox sends the description of the architecture and data flows (in English) on request. Many questions from the assessment are answered on the page Security questions.

Not legal advice: Whether a DPIA is required and how you assess risks is decided by your organisation together with its data protection officer. The mandatory lists of the supervisory authorities may change. Research as of 11 October 2026.