GDPR Article 35 DPIA for AI agents

Most AI agents that process customer data at scale will need a data protection impact assessment under GDPR Article 35, because they use new technology to evaluate people or influence decisions about them. A useful agent DPIA describes capabilities as well as data: what it can read, what it can do, and what happens if it acts wrongly.

What does Article 35 require?

A controller must carry out a DPIA before processing that is likely to result in a high risk to the rights and freedoms of people, taking into account the nature, scope, context and purposes of the processing, and in particular where new technologies are used.

Article 35 names three cases where a DPIA is required in particular:

1. systematic and extensive evaluation of personal aspects based on automated processing, including profiling, on which decisions are based that produce legal or similarly significant effects;
2. processing on a large scale of special category data or data relating to criminal convictions;
3. systematic monitoring of a publicly accessible area on a large scale.

Supervisory authorities also publish lists of processing that requires a DPIA. In substance, a DPIA must contain:

• a systematic description of the processing and its purposes, including the legitimate interest where relied on;
• an assessment of the necessity and proportionality of the processing;
• an assessment of the risks to people's rights and freedoms;
• the measures to address the risks, including safeguards and security, to demonstrate compliance.

The data protection officer's advice must be sought. If high risk remains after mitigation, the supervisory authority must be consulted before processing starts. The detailed test of when one is needed is in GDPR DPIA triggers for AI agents.

Who must carry one out?

The controller. Processors, including agent platform providers, must help, and the EU AI Act requires providers of high-risk systems to give deployers information they can use for it. In the UK, the equivalent duty sits under UK GDPR; see UK GDPR DPIA for AI agents.

What should a DPIA capture when the system is an agent?

Many DPIAs for AI describe a model: what data trains it, what data it scores, what output it gives. That misses most of what an agent does. An agent's risks to people come from its actions as much as its data.

A DPIA for an agent should describe, per capability:

• What it can read. Sources, fields and whether it can reach special category data.
• What it sends out. Prompts to the model provider, calls to tools, messages to customers.
• What it can do. Read, draft, act within limits, move money or take hard-to-undo actions.
• Who approves. Which actions wait for a person, and who that is.
• What it keeps. Logs, conversation history and memory, with retention.
• What it can be told by outsiders. Content from emails, documents and web pages that could carry instructions.

Then assess risks against that description. A DPIA written this way also reveals design choices no one made deliberately, such as an agent that can reach health notes because they sit in the same record as the address it needs. Agent-specific risks to include:

1. Disclosure of one person's data to another through a wrong record or a manipulated instruction.
2. Inaccurate data written into records and relied on later.
3. Decisions that become solely automated through gradual design change.
4. Excessive access, where the agent can read far more than tasks need.
5. Retention creep in logs and memory.
6. Unfair treatment of some groups in effort, tone or outcome.

How do you run a DPIA that reflects what the agent can do?

A ready structure is in our DPIA template for an AI agent.

What will a supervisor or auditor ask?

• Show us the DPIA. Does it describe the agent as it runs today?
• What can the agent do, and which actions did you assess?
• What data can it reach that it does not need?
• What residual risks did you accept, and who signed off?
• What did the DPO advise, and did you follow it?
• When was the DPIA last reviewed, and why?

What does a DPIA not cover?

It is not a fundamental rights impact assessment under the EU AI Act, though the two can run together. It is not a security assessment, though it includes security measures. It does not approve the processing; it shows you assessed and addressed the risk. And it does not stay valid when the agent changes. For the full map of duties, see AI agent regulation, mapped to controls.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.