AI and the Data Protection Officer: The 12 Questions Your DSB Asks Before Approving Any Tool [2026]

A data protection officer does not approve an AI tool. They approve a documented processing operation: a named use case, a legal basis, a processor contract, a place where the data goes, and a list of who may see the result. Most requests reach the DSB as "can we use ChatGPT", which is not a question that can be answered, and so the answer is no.

This guide turns the no into a checklist. It is built from the Hamburg data protection authority's checklist for LLM chatbots, the joint guidance of the German supervisory authorities on AI and data protection, Article 28, 35 and 37 to 39 GDPR, and Articles 4 and 26 of the AI Act. The twelve questions below are the ones a competent DSB asks before signing off, what a good answer looks like, and which document proves it. If you can answer all twelve, the sign-off takes an afternoon. If you cannot, you know which artefact to produce first.

The DSB advises and monitors; management decides and is liable. Articles 38 and 39 GDPR make the officer an independent adviser who must be involved "properly and in a timely manner", and the Hamburg checklist says the same in plain words: involve your DSB whenever you write internal rules or implement a use case for the first time. What the DSB cannot do is decide the purposes and means of processing, because that would create a conflict of interest under Article 38(6).

This matters for the approval conversation. A DSB who says no is usually saying "I cannot document this as lawful with what you gave me". The way past it is not to argue about the tool but to hand over the artefacts: the use-case register, the processor contract, the impact assessment, the permission concept, the training records. The rest of this article is those artefacts in order.

Each item names the question, what a defensible answer looks like, and the document that proves it. The order follows how an authority would audit you: legal basis and contract first, technical controls in the middle, people and co-determination last.

The AI Act does not require a company AI officer. What it requires since 2 February 2025 is that every deployer ensures AI literacy for the people who work with AI systems (Article 4), and from August 2026 that employers inform worker representatives before putting a high-risk system into use (Article 26(7)). The Bitkom guide on AI and co-determination notes that the Digital Omnibus proposal would shift the literacy responsibility towards the Commission and member states, but the duty is in force today.

So the KI-Beauftragter is an organisational choice, not a legal role, and companies fill it in two ways. Combining it with the DSB works in small companies if the person only advises and documents. It stops working the moment the AI officer is expected to pick tools, own the rollout budget or decide use cases, because then the same person would be deciding purposes and means and monitoring them, the exact conflict Article 38(6) forbids. Above about 100 people, or wherever the AI role is meant to drive adoption rather than check it, split the roles: an AI officer from operations or IT who owns the use-case register and the training plan, and a DSB who audits it.

The table compares the tiers companies actually ask about against the twelve questions. It is deliberately not a ranking: a US provider with a DPF certification and a signed AVV can pass, a European provider with no per-record permissions can fail on question 7. "Usually approved" means the paperwork exists and the controls can be configured; it never replaces your own DSFA. Sources: vendor terms and trust pages, Microsoft's Copilot privacy documentation, Langdock's security page, and our earlier reviews of ChatGPT Enterprise and Claude for Enterprise.

Two tiers deserve their own pages: ChatGPT Business and the GDPR for the most common self-serve purchase, and the Langdock alternative guide for companies whose first platform changed its usage limits in April 2026.

Two patterns from the table are worth naming. First, the free and consumer tiers fail on question 3 alone, which is why the Hamburg authority's rule for them is simply no personal data at all. Second, the tools that pass differ mostly on question 7: workspace-level permissions mean everyone in the workspace sees what the AI can read, and the AI can usually read everything that was connected. For a company that wants to connect its CRM, tickets and HR system, that is the question that decides the approval. The AI operating system for companies article explains why permissions per record are a platform property, not a setting.

Seven documents, in this order, close the approval in one meeting: the use-case register with legal bases; the AVV for the exact tier with the sub-processor list and transfer basis; the threshold analysis or DSFA; the permission concept with the negative test; the logging and deletion concept; the Article 4 training plan with records; and the AI policy or works agreement that binds employees to all of it. If your company is choosing a platform rather than a single tool, the same seven documents are the vendor questionnaire, and the ChatGPT alternatives for the Mittelstand comparison already scores fifteen vendors against most of them.

Professionals under section 203 of the criminal code add one more document, the provider's confidentiality undertaking; AI for tax firms explains it.