---
title: How do you run a GDPR DPIA for an AI agent?
url: https://opencontroller.lyzr.com/blog/dpia-for-ai-agents
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# How do you run a GDPR DPIA for an AI agent?

**Run a data protection impact assessment (DPIA) under GDPR Article 35 before an agent processes personal data in a way likely to cause high risk, which most agents that act on customer or employee data will meet. The DPIA must cover four things Article 35(7) requires: a description of the processing, necessity and proportionality, risks to people, and the measures that address them. For agents, describe every data flow: prompts, tools, memory, logs and the model provider.**

## When does an AI agent need a DPIA?

Article 35(1) requires a DPIA where processing, "in particular using new technologies", is likely to result in a high risk to people's rights and freedoms. Article 35(3) names three cases that always need one:

1. Systematic and extensive evaluation of personal aspects based on automated processing, including profiling, that produces legal or similarly significant effects.
2. Large-scale processing of special category data or criminal offence data.
3. Systematic monitoring of a publicly accessible area on a large scale.

The EDPB-endorsed guidelines (WP248 rev.01) add nine criteria. If processing meets two or more, a DPIA is generally required:

| WP248 criterion | How agents typically meet it |
|---|---|
| Evaluation or scoring | Agents that rank leads, triage claims or assess candidates |
| Automated decision-making with legal or similar effect | Agents that approve, reject or price without human review |
| Systematic monitoring | Agents that read email, chats or tickets continuously |
| Sensitive or highly personal data | Health, financial or HR data in prompts or retrieval |
| Large scale | Customer-facing agents or agents over whole mailboxes |
| Matching or combining datasets | Agents that join CRM, ticketing and data warehouse records |
| Vulnerable data subjects | Employees, patients, children, customers in hardship |
| Innovative use of new technology | LLM-based agents are widely treated as meeting this |
| Preventing people from exercising a right or using a service | Agents gating access to a service or contract |

An agent that reads customer tickets with an LLM already meets "innovative technology" and usually "large scale" or "combining datasets". Assume you need a DPIA unless you can show otherwise, and record the reasoning either way.

## What goes in an agent DPIA?

Structure it around Article 35(7):

**1. Systematic description of the processing.** For an agent, include:
- The agent's purpose, owner and autonomy level
- Data subjects and categories of personal data
- Every flow: user input, retrieval sources, tool calls, model provider, memory store, traces and logs
- Where each component runs and any transfers outside the EEA
- Retention for prompts, memory and logs

**2. Necessity and proportionality.** Lawful basis, data minimisation (does the agent need full records or only fields?), purpose limitation, retention limits and how data subject rights are handled, including access and erasure from memory stores.

**3. Risks to rights and freedoms.** Agent-specific risks to assess:
- Hallucinated statements about a person
- Prompt injection causing disclosure to the wrong person
- Over-broad tool access exposing records beyond the task
- Memory retaining personal data longer than intended
- Model provider using or retaining prompt data
- Automated decisions without meaningful human review (Article 22)

**4. Measures to address the risks.** Redaction before the model, least-privilege tool access, on-behalf-of identity so the agent only sees what the user may see, grounding checks, memory retention limits, human review for significant decisions, and logging that supports access requests.

## What makes agent DPIAs different?

Three things catch teams out:

1. **Agents change faster than DPIAs.** A new tool or knowledge source changes the processing. Tie DPIA review to agent versioning so material changes trigger reassessment. See [when an AI agent change needs re-approval](/blog/when-does-an-ai-agent-change-need-re-approval).
2. **Memory is personal data storage.** Long-term memory must appear in the DPIA with retention and erasure. See [governing AI agent memory stores](/blog/govern-ai-agent-memory-stores).
3. **Logs are a second copy.** Traces and prompt logs often hold more personal data than the source system. Cover them explicitly.

## What happens if residual risk stays high?

Article 36 requires prior consultation with your supervisory authority before processing begins if the DPIA shows high risk that you cannot mitigate. Article 35(2) requires you to seek the data protection officer's advice when carrying out the DPIA.

For high-risk AI systems under the EU AI Act, Article 26(9) says deployers use the provider's information to carry out their DPIA, and some deployers must also run a fundamental rights impact assessment. Those Chapter III duties now apply from 2 December 2027 for Annex III systems, after the 2026 omnibus amendment. See [EU AI Act obligations for AI agents](/blog/eu-ai-act-obligations-for-ai-agents) and [AI impact assessments for agents](/blog/ai-impact-assessment-for-ai-agents).

## A DPIA checklist for agents

1. Screen every new agent against Article 35(3) and the nine WP248 criteria.
2. Map all data flows, including tools, memory, logs and the model provider.
3. Record lawful basis and minimisation decisions per data category.
4. Assess agent-specific risks, including prompt injection and hallucination about people.
5. List technical measures and who owns each.
6. Get DPO advice and record it.
7. Link the DPIA to the agent's registry entry and version.
8. Re-run on material change and at least annually.

## How Opencontroller does this

Opencontroller's governed registry records agents, versions, tools and MCP servers, which gives the DPIA a current inventory of components and flows. Delegated on-behalf-of identity limits agents to what the user may access, the LLM Gateway can redact data in the request path, and live groundedness guardrails and LLM judges check outputs on production traffic. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

For mapping across frameworks, read [how to map AI agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks).

## Frequently asked questions

**Does every AI agent need a DPIA?**
No, only where processing is likely to be high risk. In practice, agents using LLMs on customer or employee data often meet two or more WP248 criteria, so a DPIA is usually needed. Record your screening either way.

**Who should carry out the DPIA for an agent?**
The controller is responsible, usually the business owner of the agent with privacy and security input. Article 35(2) requires seeking the DPO's advice.

**When must a DPIA be updated?**
When the processing changes in a way that affects risk, such as a new tool, data source, model provider or memory setting. Linking reviews to agent versions makes this routine.

## Key facts

GDPR Article 35 requires a DPIA where processing is likely high risk; Article 35(3) lists three mandatory cases; WP248 rev.01 gives nine criteria, two or more generally require a DPIA. Article 35(7) contents: description, necessity and proportionality, risks, measures. Article 35(2) DPO advice; Article 36 prior consultation if residual high risk. EU AI Act Article 26(9): deployers of high-risk systems use provider information for their DPIA; Chapter III applies from 2 December 2027 for Annex III after Regulation (EU) 2026/1744. Opencontroller provides a governed registry, on-behalf-of identity, gateway redaction and groundedness guardrails. Pricing starts from $250K per deployment per year.

## Related

- [How do you run an AI impact assessment for an AI agent?](https://opencontroller.lyzr.com/blog/ai-impact-assessment-for-ai-agents.md)
- [What does the EU AI Act require for AI agents, and when?](https://opencontroller.lyzr.com/blog/eu-ai-act-obligations-for-ai-agents.md)
- [How do you govern what AI agents remember?](https://opencontroller.lyzr.com/blog/govern-ai-agent-memory-stores.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
