Skip to content
Blentera

Resources, Data protection

GDPR and AI agents: a practical checklist for companies

AI agents read, write and remember personal data, so the GDPR applies to them as it does to any other processing. This checklist covers what a company has to decide, document and contract before agents touch personal data.

Updated 20 November 2026. 10 minute read.

In short

  • Your company is normally the controller for what its agents do with personal data. The agent platform and the AI provider are usually separate processors, and each needs an Art. 28 agreement.
  • Every agent use needs a legal basis under Art. 6, a clear purpose and only the data that purpose requires.
  • Run a DPIA before go-live when processing is likely to be high risk. Agents that evaluate or decide about individuals, or that handle sensitive data at scale, usually meet that bar.
  • Decisions with legal or similarly significant effects should not be left solely to an agent unless an Art. 22 exception applies, with human review available.
  • The EU-US Data Privacy Framework is still in force as of November 2026, but it is under legal challenge. Keep a fallback transfer tool ready.

Controller, processor and the AI provider

The GDPR assigns duties by role. The controller decides why and how personal data is processed. A processor processes it on the controller's behalf (Art. 4(7) and 4(8)). The EDPB guidelines on these concepts stress that the label in a contract does not decide the role. The facts do.

When your company deploys an agent to handle customer emails, screen applications or update records, your company decides the purpose. That makes it the controller. An agent platform that runs the agents for you acts as a processor.

The AI provider is a separate party. Every prompt an agent sends, and every document it attaches, goes to the model provider's servers. If you connect your own provider account, you usually hold that contract directly. The provider then acts as your processor, not as a sub-processor of the platform. OpenAI's data processing addendum describes OpenAI as a processor acting on the customer's instructions. Anthropic's Privacy Center describes Anthropic as a processor for its commercial customers. Both state that API data is not used to train their models by default.

Processor agreements (Art. 28)

Art. 28(1) allows only processors that give sufficient guarantees. Art. 28(3) requires a binding contract that sets out the subject matter, duration, nature and purpose of processing, the types of personal data and the categories of data subjects. The contract must oblige the processor to:

  • act only on documented instructions
  • ensure staff are bound by confidentiality
  • apply the security measures required by Art. 32
  • use sub-processors only with your prior written authorisation and on the same terms
  • help you answer data subject requests and meet Art. 32 to 36 duties
  • delete or return the data when the service ends
  • give you the information needed to show compliance and allow audits

Map every party an agent's data passes through: the agent platform, each AI provider account, and each connected tool. Each one needs its own processor agreement or a clear controller role.

When a DPIA is required

Art. 35(1) requires a data protection impact assessment before processing that is likely to result in a high risk to individuals' rights and freedoms, "in particular using new technologies". Art. 35(3) makes it mandatory in three cases:

  1. systematic and extensive evaluation of personal aspects based on automated processing, including profiling, that leads to decisions with legal or similarly significant effects
  2. large-scale processing of special categories of data or criminal records data
  3. systematic monitoring of a publicly accessible area on a large scale

The EDPB-endorsed guidelines (WP248 rev.01) give nine criteria. They include evaluation or scoring, automated decisions with significant effect, systematic monitoring, sensitive data, large scale, combining datasets, vulnerable data subjects, innovative technology and processing that blocks access to a service or right. As a rule of thumb, processing that meets two criteria will usually need a DPIA.

Under Art. 35(4), each supervisory authority publishes a list of processing that always needs a DPIA. The EDPB keeps a register of these national lists. The Danish Datatilsynet's list builds on the EDPB criteria and states that it is not exhaustive. In its guidance on AI for public authorities, Datatilsynet says that, in its view, a DPIA will normally be needed when developing and using AI solutions that process personal data.

A DPIA must describe the processing, assess necessity and proportionality, assess the risks and set out the measures to address them (Art. 35(7)). If high risk remains after those measures, consult the supervisory authority before starting (Art. 36).

Automated decisions and transparency

Art. 22(1) gives individuals the right not to be subject to a decision based solely on automated processing, including profiling, that produces legal effects or similarly significantly affects them. Rejecting a job application, refusing credit or ending a contract are typical examples.

Art. 22(2) allows such decisions only when they are necessary for a contract, authorised by Union or Member State law, or based on explicit consent. In the first and third cases, Art. 22(3) requires safeguards, at least the right to obtain human intervention, to express a point of view and to contest the decision. In the SCHUFA case (C-634/21), the Court of Justice held that an automated score can itself be such a decision when a third party draws strongly on it.

The practical rule: let agents prepare, sort and recommend. When the outcome has legal or similarly significant effects, an employee with real authority reviews it and decides. A rubber stamp does not count as human involvement.

Telling individuals

Arts. 13 and 14 require you to tell individuals what you process, why, on which basis, who receives it and how long you keep it. Where Art. 22 decision-making takes place, you must also give meaningful information about the logic involved and the likely consequences (Art. 13(2)(f), 14(2)(g) and 15(1)(h)). Art. 12 requires concise, plain language. Update your privacy notice to name AI providers as recipients and to explain where agents are used.

International transfers and the Data Privacy Framework

Chapter V restricts transfers of personal data outside the EEA. Many AI providers are US companies, so check where each provider processes data and which transfer tool applies. The main tools are an adequacy decision (Art. 45) and appropriate safeguards such as the Commission's standard contractual clauses (Art. 46).

The Commission adopted its adequacy decision for the EU-US Data Privacy Framework on 10 July 2023 (Implementing Decision (EU) 2023/1795). It covers transfers only to US organisations certified under the framework. The Commission's first periodic review, published in October 2024, found that the framework was working.

Status as of November 2026

  • On 3 September 2025, the General Court dismissed Latombe's action to annul the decision (Case T-553/23).
  • Latombe appealed on 31 October 2025 (Case C-703/25 P). As of 20 November 2026, we found no judgment from the Court of Justice. Check the Court's case register for the latest position.
  • On 31 July 2026, the EDPB wrote to the Commission about the US Supreme Court judgment in Trump v. Slaughter. The letter notes that the adequacy decision relies on the independence of US authorities, including the Federal Trade Commission.
  • The adequacy decision stays in force until the Commission amends, suspends or repeals it, or the Court of Justice annuls it.

The framework can be used today, but its future is uncertain. Keep standard contractual clauses and a transfer impact assessment ready, or prefer providers that process data in the EEA.

Retention, agent memory, rights and security

Storage limitation means personal data is kept no longer than the purpose requires (Art. 5(1)(e)). Agents create new places where data lives: conversation history, company memory, logs, files and the provider's own retention. OpenAI and Anthropic both describe keeping standard API inputs and outputs for up to 30 days, with exceptions for some features, and both offer zero data retention to approved customers. Set and document a period for each store.

Data subject rights in agent memory

Rights of access, rectification, erasure, restriction, portability and objection (Arts. 15 to 21) apply to data in agent memory as they do to any database. Answer within one month, extendable by two further months for complex requests (Art. 12(3)). This only works if memory is searchable and editable. A model's weights are a different matter: information from a request is not stored there when the provider does not train on your data. That is one reason to keep company knowledge in a store you control rather than in a fine-tuned model.

  • keep agent memory in a store you can search by name, email or customer ID
  • make sure corrections and deletions reach memory, logs and connected systems
  • record which agent wrote which memory entry, and when

Security (Art. 32)

Art. 32 requires measures appropriate to the risk. It names pseudonymisation and encryption, ongoing confidentiality, integrity, availability and resilience, the ability to restore data after an incident, and regular testing. For agents, add least-privilege access, approval steps for actions that send data out or change records, spending limits that stop runaway loops, and an activity log. A personal data breach must normally be reported to the supervisory authority within 72 hours (Art. 33).

The checklist

  1. List every agent, its purpose and the personal data it can reach. Add each one to your records of processing (Art. 30).
  2. Confirm roles. Sign Art. 28 agreements with the agent platform, each AI provider account and each connected tool.
  3. Record a legal basis per purpose. Document the balancing test if you rely on legitimate interests.
  4. Limit access per agent role. Remove data the task does not need.
  5. Screen each use against Art. 35(3), the EDPB criteria and your national list. Complete a DPIA where needed.
  6. Find any decisions with legal or similarly significant effects. Put an employee with authority in charge of them.
  7. Update privacy notices to cover agents, AI providers and any automated decision-making.
  8. Map transfers outside the EEA. Check each provider's transfer tool and keep a fallback to the Data Privacy Framework.
  9. Set retention for conversations, memory, logs and provider data.
  10. Test that you can find, export, correct and delete one individual's data across agent memory.
  11. Apply Art. 32 measures: encryption, least privilege, approvals, budgets, logging and breach procedures.
  12. Review the setup when you add an agent, change a provider or change what an agent may do.

How Blentera helps

The platform is hosted in the EU and gives each agent a role in an org chart, so access follows the role. Approval rules let an employee sign off before an agent acts. Budgets warn at 80% and pause agents at the limit. One company memory holds what agents know, and an activity log records what they did. You connect your own AI provider accounts, so you choose the provider and its terms. A data processing agreement is offered on every plan, including Free. These tools support your GDPR work. They do not make your company compliant on their own.

Questions

Sources

  1. Regulation (EU) 2016/679 (GDPR), EUR-Lex
  2. EDPB Guidelines 07/2020 on the concepts of controller and processor in the GDPR
  3. EDPB Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models
  4. Article 29 Working Party, Guidelines on data protection impact assessment (WP248 rev.01), endorsed by the EDPB
  5. EDPB register of national DPIA lists (Art. 35(4))
  6. Datatilsynet: list of processing that always requires a DPIA
  7. Datatilsynet: Offentlige myndigheders brug af kunstig intelligens – Inden I går i gang (2023)
  8. Article 29 Working Party, Guidelines on automated individual decision-making and profiling (WP251 rev.01)
  9. Court of Justice, Case C-634/21 SCHUFA Holding (Scoring), judgment of 7 December 2023
  10. Commission Implementing Decision (EU) 2023/1795 on the EU-US Data Privacy Framework
  11. European Commission: report on the first periodic review of the EU-US Data Privacy Framework, COM(2024) 451
  12. General Court press release on Case T-553/23 Latombe v Commission (3 September 2025)
  13. Appeal in Case C-703/25 P Latombe v Commission, Official Journal notice
  14. EDPB letter of 31 July 2026 on the US Supreme Court judgment in Trump v. Slaughter
  15. European Commission: data protection adequacy for non-EU countries
  16. OpenAI Data Processing Addendum
  17. Anthropic: API and data retention
  18. OpenAI: data controls in the OpenAI platform
  19. Anthropic Privacy Center: does Anthropic act as a data processor or controller?
  20. Anthropic Privacy Center: is my data used for model training?

This guide is general information, not legal advice. Check how the rules apply to your company with your own advisers.

Get started

Put your first agent to work in minutes.

Three steps, and your first agent is working for your company.

  1. Step 1

    Create an account

    Sign up with your work email and invite your team. No sales call needed.

  2. Step 2

    Add your first agent

    Create an agent, give it a role, and connect your own AI key and the apps it may use.

  3. Step 3

    Give it a task

    Ask in Blentera or your chat app, and follow the work until it is done.