AI agent governance is part of the controller's duty to implement appropriate technical and organisational measures under GDPR Articles 24, 25 and 32. On 28 October 2026 the Norwegian Data Protection Authority advised organisations to map their agents, restrict what the agents can access and do, give each agent its own identity and appoint a person responsible who can stop it.
On 28 October 2026 the Norwegian Data Protection Authority (Datatilsynet) published advice on the governance of AI agents under the heading «Gode rutiner hindrer at KI blir en usynlig trussel» (good routines prevent AI from becoming an invisible threat). The background is that language models have moved from tools that answer questions to agents that themselves retrieve data, update business systems and carry out tasks. The advice is not a regulation or a decision, but it shows what the authority expects of organisations that give agents access to email, file storage and case management systems. The question is which obligations the GDPR and the AI Act impose, and how the advice should be translated into procedures and contracts.
What is an AI agent, and why does it require separate governance?
An AI agent is an AI system that, on the basis of a goal, plans and performs actions in other systems, for example reading and answering email, retrieving documents, registering cases or updating customer records. The difference from a chatbot is that the agent acts with the organisation’s access rights and does not merely suggest text for a person to review. The agent’s actions are therefore the organisation’s actions, and the processing of personal data that occurs is the organisation’s processing.
The Norwegian Data Protection Authority identifies three risks. The first is shadow AI, where employees set up automated workflows that continue to run and process personal data after the employee has left. The second is AI embedded in applications the organisation has already approved, which erodes oversight without anyone having taken a decision. The third is that agents accumulate large amounts of context, such as customer data and meeting minutes, and that several interacting agents create new data flows that are difficult to trace. Liability for harm caused by agents to others is discussed in the article on AI agent liability.
What does the Norwegian DPA recommend on AI agent governance?
The Norwegian Data Protection Authority recommends that organisations have control of data quality, information security and access management before agents are deployed, and refers to guidance from the Norwegian Digitalisation Agency on getting one’s own house in order. The authority emphasises that the aim is safe use and not prohibition, and that staff awareness must be followed by technical controls and procedures. The core of the advice is a three-part inventory.
| Question | What the inventory should show |
|---|---|
| What exists? | All AI tools, assistants and automations in use, including AI embedded in approved applications |
| What can the agents do? | Which systems the agent can access and which actions it is permitted to perform |
| Who is responsible? | Who monitors the agent, owns the process and can deactivate it |
The authority concludes that an AI agent needs a defined identity, clear limits on access and a clear allocation of responsibility. In practice this means that the agent should have its own technical identity in the organisation’s access management, rather than borrowing the access rights of the employee who set it up.
Which GDPR requirements apply to AI agents?
The advice gives concrete form to obligations controllers already have under the GDPR, which applies in Norway through the Personal Data Act. Under GDPR Art. 24, the controller must implement appropriate technical and organisational measures and be able to demonstrate compliance. Under GDPR Art. 25, data protection must be built in, so that only the personal data necessary for the purpose are processed. An agent with broad read access to all mailboxes in order to perform a limited task will hardly meet the data minimisation requirement in GDPR Art. 5(1)(c).
The security requirement in GDPR Art. 32 covers access management, logging and the ability to detect and handle incidents. An agent performing actions that cannot be traced afterwards makes it difficult to determine whether a personal data breach has occurred, and therefore to notify it within the deadline. Processing carried out by agents must also be included in the record of processing activities under GDPR Art. 30. Under GDPR Art. 35, the organisation must assess the impact on data protection before processing likely to result in a high risk, and agents with independent access to business systems will often reach that threshold. The requirements for such an assessment are discussed in the article on DPIAs for AI tools.
The accountability principle in GDPR Art. 5(2) is in our assessment the most important legal basis for the authority’s advice. The controller must be able to demonstrate compliance, which presupposes that the organisation knows which agents process personal data, for which purpose and with which access rights. An automation that continues to run without an owner is in itself a breach of that obligation.
What does the AI Act require of deployers of AI agents?
The AI Act imposes specific requirements where an agent is a high-risk system, for example when it is used for recruitment, for evaluating employees or for deciding access to public services. Under AI Act Art. 12(1), the provider must ensure that the system automatically records events, and under AI Act Art. 14(4)(e) that a person can intervene in or stop the system so that it comes to a halt in a safe state. The deployer must under AI Act Art. 26(6) keep the logs for at least six months, to the extent they are under its control.
The AI Act has not been incorporated into the EEA Agreement, and the obligations for high-risk systems have been postponed by the omnibus Regulation (EU) 2026/1744. The requirements nevertheless apply to Norwegian organisations offering or using such systems in the EU, and they provide a sound template for responsible governance also outside the high-risk area. The authority’s recommendation of a named person who can deactivate the agent corresponds in practice to the requirement of human oversight.
What should the contract with the AI agent provider cover?
The contract with the provider must enable the organisation to carry out the governance the authority recommends. The provider will normally be a processor, and the data processing agreement under GDPR Art. 28 must cover the processing the agent performs, including the use of sub-processors for the language model. The requirements for such agreements are discussed in the article on data processing agreements for SaaS.
The contract should also ensure that the organisation can restrict the agent’s access and permitted actions, that actions are logged with time and identity, and that the logs are available to the organisation for a period at least equal to the requirement in AI Act Art. 26(6). It should further give a right to stop the agent immediately and oblige the provider to give notice of changes to the model or functionality that affect what the agent can do. Whether the organisation is bound by contracts the agent concludes with third parties is discussed in the article on AI agents and contract formation.
How should organisations implement AI agent governance?
Governance should be anchored in existing procedures for access management and internal control, so that agents are treated in the same way as other users with access to personal data. Further articles on AI in organisations are available on the artificial intelligence topic page.
- Map all AI agents and automations, including those set up by employees themselves and those embedded in existing applications.
- Give each agent its own technical identity with access limited to its task, and do not let it use an employee’s user account.
- Set out in writing which actions the agent may perform and which require approval by a person.
- Appoint a named owner for each agent with authority to stop it, and include the transfer of ownership in the procedures for departing employees.
- Enter the agents in the record of processing activities, and assess the data protection impact before agents with broad access go into operation.
- Log the agents’ actions so that a personal data breach can be detected, investigated and notified within 72 hours under
GDPR Art. 33.
Questions and answers
Is an AI agent a processor under the GDPR?
No. An AI agent is a tool and not a legal person. The organisation that deploys the agent for its own purposes is the controller of the processing the agent carries out. The provider operating the agent platform will normally be a processor, in which case a data processing agreement under GDPR Article 28 is required.
Must an organisation carry out a DPIA before deploying AI agents?
A data protection impact assessment is required under GDPR Article 35 where processing is likely to result in a high risk. Agents that independently process personal data in business systems, in particular data on employees, customers or vulnerable groups, will often meet that threshold because they combine new technology with extensive access.
Who is responsible when an employee who set up an agent leaves?
Responsibility lies with the organisation as controller and does not follow the employee. The Norwegian Data Protection Authority therefore recommends that every agent has a named owner and that ownership is transferred on departure, so that automations do not continue to process personal data without anyone in control.
- Norwegian Data Protection Authority, Gode rutiner hindrer at KI blir en usynlig trussel (28 October 2026)
- General Data Protection Regulation (EU) 2016/679 Arts. 5, 24, 25, 28, 30, 32, 33 and 35
- Regulation (EU) 2024/1689 on artificial intelligence (AI Act) Arts. 12, 14 and 26
- Norwegian Personal Data Act (personopplysningsloven)
Next legal review: 31 December 2026