Skip to main content
Praxikon

Article 30 GDPR and AI: what goes in your records of processing activities

··19 min read

Every organisation that processes personal data with an AI application includes that processing in its records of processing activities (Article 30 GDPR), often called the ROPA. For each processing activity the record states, among other things, the purpose, the data subjects and the data, the recipients, including the AI vendor, any transfer outside the EEA, the retention periods and the security measures. An AI vendor acting as processor keeps a record of its own as well. The exemption for organisations with fewer than 250 employees rarely helps with AI, because it falls away as soon as processing is not occasional or is likely to result in a risk. The record is also the starting point for deciding which AI processing needs a DPIA.

Picture this: you are the privacy officer of a municipality. The customer service team has put a chatbot on the website to answer residents' questions, and the social welfare department is testing a tool that summarises benefit application files. Your data protection officer asks whether both are in the records of processing activities. The honest answer is often: the chatbot is not, because nobody saw it as processing, and the tool only half, under the existing entry for welfare benefits, without the vendor as a recipient. This post walks through what belongs in each item.

In short

Article 30 GDPR requires the controller to keep a written record of its processing activities with seven fixed items (paragraph 1, points (a) to (g)), and the processor to keep its own record with four items (paragraph 2). The record may be electronic (paragraph 3) and must be made available to the supervisory authority on request (paragraph 4). Organisations with fewer than 250 employees are only exempt for processing that is occasional, unlikely to result in a risk and does not involve special category or criminal offence data (paragraph 5). An AI application is rarely occasional. Failing to keep records can lead to a fine of up to 10 million euros or, for an undertaking, 2 percent of worldwide annual turnover, whichever is higher (Article 83(4)(a) GDPR). According to the Court of Justice, a missing record does not in itself make processing unlawful (C-60/22), but it remains a separate infringement.

What Article 30(1) requires, and what you fill in for AI

The record describes processing activities, not systems. One AI application can serve several purposes, and each purpose then belongs to the processing activity it supports. A chatbot that answers questions and whose conversations are also used to analyse service quality touches two purposes. Conversely, one processing activity, such as handling benefit applications, can use several systems. The AI application then becomes part of that existing processing activity, with its own recipients and retention periods.

Item in Article 30(1)What the GDPR requiresWhat you fill in for an AI application
(a)Name and contact details of the controller, any joint controller, the representative and the data protection officerYour own organisation and the DPO. If you determine purposes and means together with another party, for example a group of municipalities sharing one AI application, that party as well (Article 26 GDPR)
(b)The purposes of the processingThe specific purpose for which you use the AI, such as "answering residents' questions about waste collection and passports", not "using AI". Each additional purpose, such as analysis or improvement of the application, separately
(c)Categories of data subjects and of personal dataWho appears in the input, the output and the logs: residents, customers, applicants, pupils, staff. Which data: what people type in themselves, what the application retrieves from other sources, and what it infers, such as a score or a summary. Note separately whether special category data (Article 9) or criminal offence data (Article 10) may end up in it
(d)Categories of recipients, including in third countriesThe AI vendor and its hosting provider. A processor is also a recipient: the GDPR defines a recipient as any party to which personal data are disclosed, whether a third party or not (Article 4(9)). Plus internal departments that use the output
(e)Where applicable, transfers to a third country or international organisation, naming that country and, for transfers under the second subparagraph of Article 49(1) GDPR, the documentation of suitable safeguardsWhere the vendor or a sub-processor processes the data or provides support outside the EEA, and the country. Recommended: also note the transfer tool, for example an adequacy decision or standard contractual clauses (Article 45 and Article 46)
(f)Where possible, the envisaged time limits for erasure of the different categories of dataSeparate periods for input (prompts), output, conversation logs and any copies held by the vendor. Check the vendor's default settings: they sometimes retain data longer than you would choose
(g)Where possible, a general description of the technical and organisational security measures referred to in Article 32(1)Access control, encryption, logging, agreements that the vendor does not use input for its own purposes, and a reference to the security annex of the data processing agreement (Article 32)

Two things often go wrong in practice. The first is the purpose. If the vendor also uses your input to train its own models, it determines the purpose of that processing itself and is no longer a processor for it but a controller (Article 4(7); if it does so in breach of the agreement, Article 28(10) applies). Your record must then show that you disclose data to a third party, and you need a legal basis for that. The second is the question of which data enter the application. With an open input field, the user decides what goes in. A resident who tells a chatbot that he receives benefits because of an illness is supplying health data, even if the chatbot was not designed for it.

What the AI vendor keeps as processor (paragraph 2)

If the vendor processes the data only on your behalf, it is a processor and keeps its own record of all categories of processing it carries out for its customers (Article 30(2) GDPR). That record contains its name and contact details and those of each controller on whose behalf it acts, the categories of processing per customer, transfers outside the EEA and a general description of security. Purposes, data subjects and retention periods are not in it: you determine those, so they belong in your record.

If the vendor engages other parties itself, such as a provider of the underlying language model or a cloud platform, those are sub-processors. The vendor needs your prior specific or general written authorisation, and under a general authorisation it must inform you in advance of changes so that you can object (Article 28(2)). The sub-processor is bound by the same obligations (Article 28(4) GDPR). Ask for the list of sub-processors and the countries in which they operate: that is how you fill in items (d) and (e) of your own record.

In writing and available on request (paragraphs 3 and 4)

The record must be in writing, including in electronic form (paragraph 3). A spreadsheet or a privacy management tool is therefore fine, as long as it is kept up to date. The controller and the processor make the record available to the supervisory authority on request (paragraph 4). Recital 82 GDPR explains why: the record serves to demonstrate compliance and is a tool for supervision.

The exemption for fewer than 250 employees (paragraph 5)

The obligation does not apply to an enterprise or organisation employing fewer than 250 persons, unless one of three situations arises: the processing is likely to result in a risk to the rights and freedoms of data subjects, the processing is not occasional, or the processing includes special categories of data or criminal offence data (Article 30(5) GDPR).

The Article 29 Working Party, the predecessor of the European Data Protection Board (EDPB), stated in a position paper of 19 April 2018 that each of those three situations on its own triggers the obligation, and that it concerns a risk, not just a high risk. A small organisation then only needs to record those processing activities; genuinely occasional processing without risk and without special category data may stay out of the record. According to the Working Party, processing is only occasional if it is not carried out regularly and falls outside the regular course of business. The EDPB endorsed this position paper on 25 May 2018 (Endorsement 1/2018).

For AI this means: an application you use on a structural basis, such as a chatbot, a tool for assessing applications or software that marks pupils' work, is not occasional. A small organisation must therefore record that processing.

Case law: does a deficient record make processing unlawful?

No, the Court of Justice held on 4 May 2023 in C-60/22, UZ v Bundesrepublik Deutschland (ECLI:EU:C:2023:373). A German federal office had not produced a complete record of processing activities despite a request from the court, and had no arrangement for joint controllership either (Article 26 GDPR). The Court found that keeping a record is not one of the grounds for lawful processing under the first subparagraph of Article 6(1) GDPR. Where the absence of a record or of such an arrangement does not, as such, infringe the accountability principle read with Article 5(1)(a) and Article 6(1) GDPR, the processing is therefore not unlawful and the data subject has no right to erasure or restriction. The infringement must be remedied through the other measures of the GDPR, such as the corrective powers of the supervisory authority, a complaint or compensation (paragraphs 59 to 61, 65 to 67 and 69 and the operative part).

That does not make the record optional. It remains a separate obligation backed by a fine; whether a public body can be fined is for national law to decide (Article 83(7) GDPR), and in the German case it was excluded (paragraph 68). Moreover, without a record it is hard to meet the accountability principle in Article 5(2).

Fine

An infringement of Article 30 falls under Article 83(4)(a) GDPR: an administrative fine of up to 10 million euros or, for an undertaking, up to 2 percent of total worldwide annual turnover in the preceding financial year, whichever is higher (Article 83).

How the record connects to the DPIA, the AI register and the AI Act

DPIA. A data protection impact assessment is required where a type of processing, in particular using new technologies, is likely to result in a high risk (Article 35(1)). The record is the source for asking that question per processing activity: without an overview of what you process, you cannot know which processing you need to assess. Note in the record, for each AI processing activity, whether a DPIA is required and where it is kept. When a DPIA is required for AI is covered in a DPIA for AI systems: when it is required and in the Dutch DPA list of mandatory DPIAs.

The AI Act. For a high-risk system, the deployer uses the information the provider supplies under Article 13, such as the instructions for use, to carry out its DPIA (Article 26(9) AI Act). The same information helps with the record: which data the system processes, which logs it generates and for how long. If you as deployer keep the logs of a high-risk system, the period must be appropriate to the intended purpose and at least six months, unless other law, in particular data protection law, provides otherwise (Article 26(6) AI Act). That period belongs in item (f). The obligations in Article 26 apply to high-risk systems in Annex III from 2 December 2027 (Article 113, as amended by Regulation (EU) 2026/1744), with a transitional rule for systems already in use before then: they fall under it only after a significant change in design, and systems intended for use by public authorities by 2 August 2030 at the latest (Article 111(2) AI Act). The records of processing activities are required now.

AI register. The AI Act does not require an internal AI register for every AI system, but without an overview it is hard to demonstrate compliance with its obligations, see is an AI register required?. Registration in the EU database under Article 49 AI Act is something else: it applies to providers of high-risk systems in Annex III (except point 2), to providers that conclude under Article 6(3) that an Annex III system is not high-risk, and, for their use, to deployers that are public authorities or act on their behalf (Article 49(1) to (3) AI Act). The practical advice: keep one inventory and link each entry in the AI register to the processing activity in the records of processing activities, instead of two lists that drift apart. How that works within a privacy function is covered in AI governance through the privacy function; for AI agents, see the AI agent register.

Transparency. What is in the record is also what you must tell data subjects: purpose, recipients, transfers and retention period also appear in Article 13 and Article 14 GDPR. A complete record entry is therefore the basis for the privacy notice that goes with the AI application.

Three examples

The examples are fictional and serve to make the items concrete.

Municipality. The website chatbot runs on a language model that the vendor buys from a US provider; conversations are stored in the EU, but the model provider may have administrative access from the United States. The record gets a new processing activity, "answering residents' questions through the chatbot". Data subjects: residents and other visitors. Data: whatever people type in, with the note that health or income data may be included despite the warning in the chatbot. Recipients: the vendor as processor and the model provider as sub-processor. Transfer: the United States, with the transfer tool the vendor relies on, for example an adequacy decision or standard contractual clauses (Articles 45 and 46 GDPR). Retention: for example thirty days for conversation logs, with confirmation that the model provider does not retain the input or use it for training. The summarising tool in the social welfare department becomes part of the existing processing activity "handling applications for social assistance", with the vendor as a new recipient and its own retention period for the summaries. Because it concerns access to public assistance, it may be a high-risk system (Annex III, point 5(a) AI Act), unless it only performs a preparatory task without profiling (Article 6(3) AI Act). If it is, and it is in use before 2 December 2027, the municipality must meet the high-risk obligations by 2 August 2030 at the latest, or earlier if the system undergoes a significant change (Article 111(2) AI Act). The DPIA question arises now: the municipality assesses whether a DPIA is required before the tool goes into use.

Insurer. A non-life insurer with 900 employees uses a vendor's AI model to triage incoming personal injury claims and flag possible fraud. The paragraph 5 exemption does not apply: the insurer has more than 250 employees. The record lists "handling personal injury claims" and, separately, "fraud detection" as purposes, because the second purpose may involve other data and other recipients. The data entry notes that health data are involved, which requires an exception under Article 9(2). The recipients entry lists the vendor, and the question whether it also uses the claims data to improve its model for other customers. If so, it is a controller for that processing and the insurer must assess this as a disclosure to a third party. An automated flag that decides which claims are investigated further, using health data on a large scale, is processing for which a DPIA is mandatory (Article 35(3)(b) GDPR).

Education. A secondary school with 120 staff has pupils submit writing assignments in a program that gives AI feedback. The school believes it is below the 250 threshold and therefore needs no record. That is wrong: the processing runs throughout the school year and is therefore not occasional. Because it concerns data of minors, it is moreover likely to result in a risk. The record lists pupils and teachers as data subjects, the assignments, the feedback and user data as data, the vendor of the program and its sub-processors as recipients, and, for example, the end of the school year as retention period for the assignments. Ask the vendor for its list of sub-processors and whether pupils' texts are used for training.

Five steps to a record that covers AI

  1. Map which AI applications process personal data. Include AI features built into existing software, such as a summarise button in the case management system or an assistant in office software, and applications that staff use on their own initiative.
  2. Determine the roles per application. Are you a controller, jointly with another party, or a processor? And does the vendor use the data only for you, or also for itself?
  3. Fill in items (a) to (g) per processing activity. Attach the application to the processing activity it is used for, or create a new one if the purpose is new. Take the details on sub-processors, countries, retention periods and security from the data processing agreement and the vendor's documentation.
  4. Make the links. For each processing activity, note whether a DPIA is required and where it is kept, which data processing agreement applies, which privacy notice informs data subjects and which AI register entry belongs to it.
  5. Keep the record up to date. Update the entry when the vendor switches to a new model, adds a sub-processor, changes the retention period, or when you start using the application for a new purpose. A notice of a new sub-processor under Article 28(2) GDPR is a fixed moment to review the record.

Want to know, for a specific use case, which GDPR and AI Act obligations come into view, and whether a DPIA is required? With the case check you describe the use case in a few questions.

Frequently asked questions about records of processing activities for AI

Newsletter

Every Tuesday, the AI Act week ahead in 5 minutes

A practical briefing on deadlines, new guidance and enforcement, so you know what matters this week. No spam and you can unsubscribe in one click.

Practical and short · No spam · One-click unsubscribe

Written by

Zahed Ashkara, freelance AI and privacy consultant and legal adviser

Zahed Ashkara

Freelance AI and privacy consultant and legal adviser · AI governance & safety

Zahed Ashkara is a freelance AI and privacy consultant and legal adviser for the GDPR and the EU AI Act. Organisations hire him as AI governance consultant, privacy counsel, external DPO or AI governance lead, have him carry out a DPIA, FRIA or AI Act classification, or book an in-company training. He focuses on AI governance and safety, including AI agents. He works exclusively as a freelancer, through Embed AI (trade name Praxikon). He is a Certified AI Compliance Officer (CAICO) and has been a member of the NEN standards committee Artificial Intelligence & Big Data since July 2026.

Follow on LinkedInView expert profile

Use these sources in your AI assistant too: Praxikon in Claude and other assistants

Real-World Stories

Discover how organizations apply this in practice

Related articles

Is your AI system on the Dutch DPA's mandatory DPIA list?

An AI system is never on the list of the Dutch Data Protection Authority as AI. The list looks at what you do with which data, and for most categories only when that happens on a large scale or systematically. What the 17 categories actually say, with examples from a municipality, a bank, an insurer, an employer and a school.

DPIAGDPRDutch Data Protection Authority

Regulators will look at automated candidate screening in 2027: what to put in place now

On 11 November 2026 the European data protection authorities chose the topic of their joint enforcement action for 2027: automated decisions in recruitment, including with AI. What it means for the privacy officer at a municipality, a bank or a school, which questions to expect under Articles 15 and 22 GDPR, and how the AI Act adds to that from 2 December 2027.

GDPRArticle 22 GDPRRecruitment