AI Privacy Report
A person filling out a printed assessment form on a clipboard
policy

DPIA Template for LLM Deployment: A Working Structure

A practical Data Protection Impact Assessment structure for LLM-integrated workflows. Includes the risk factors GDPR Article 35 requires, the AI Act overlay, and the sections most often skipped.

By Hannah Linden · · 8 min read

A Data Protection Impact Assessment is required under Article 35 GDPR when processing is likely to result in a high risk to the rights and freedoms of natural persons. LLM deployments almost always trigger DPIA because they involve novel technologies, large-scale processing of personal data, and frequently automated decision-making. Under the EU AI Act, high-risk AI systems also require a Fundamental Rights Impact Assessment (FRIA, Article 27), which overlaps with — but does not replace — the DPIA. This is the working template I use with practitioner clients.

When a DPIA is required for LLM workflows

GDPR Article 35(3) lists three mandatory DPIA triggers. The EDPB’s nine-criteria test (WP 248) operationalizes these. For LLM workflows, the typical triggers are:

If two or more criteria apply, DPIA is required. For most production LLM workflows, three or four apply.

Template structure

The DPIA should address the following sections. Section names mirror Article 35(7) text; subsections reflect operational practice.

1. Systematic description of the processing operations

1.1 Purposes and lawful bases

State the processing purpose specifically. “Improve customer service” is not specific enough. “Generate suggested first-response messages for human review in our complaints queue” is. State the Article 6 lawful basis for each purpose. For special-category data, state the Article 9 condition.

1.2 Data flow

Document the data flow end-to-end:

Include the third parties involved (model provider, infrastructure provider, fine-tuning data processors). Each is either a processor or joint controller; document which.

1.3 Legitimate interest assessment (where applicable)

If you rely on Article 6(1)(f), document the three-part test: purpose test, necessity test, balancing test.

2. Necessity and proportionality

State why the processing is necessary and why a less-intrusive alternative is not sufficient. For LLM workflows, common alternative-considered framings:

Document why each alternative was rejected. This section is where regulators look for evidence of substantive consideration.

3. Risks to data subjects

Risk identification specific to LLM workflows. Categories to address:

3.1 Memorization and re-emission

Risk that the model emits training data containing personal information. Probability and severity depend on model provenance and fine-tuning data composition.

3.2 Hallucination of personal data

Risk that the model generates plausible-but-false statements about real persons. Particularly relevant for customer-facing assistants that name external persons in outputs.

3.3 Inference of sensitive attributes

Risk that the workflow infers protected-category attributes from non-protected inputs (e.g., inferring health status from search behavior).

3.4 Automated decision-making impact

If Article 22 applies, the risk of consequential decisions without meaningful human review. See the linked Article 22 framework.

3.5 Cross-context use

Risk that data collected for one purpose is reused inappropriately when fine-tuning or RAG pipelines blur context boundaries.

3.6 Security risks

Prompt injection enabling exfiltration of personal data. RAG-store poisoning. Tool-use abuses. See defensive cluster controls at sentryml.com for the technical-control inventory.

3.7 Cross-border transfer risks

If the model is operated outside the EEA, document the Article 46 mechanism (SCCs, BCRs, adequacy decision). The Schrems II case-law overlay applies.

4. Risk mitigation measures

For each identified risk, document the mitigations in place and the residual risk. Mitigations relevant to LLM workflows:

5. Consultation

Article 35(2) requires consultation with the DPO. Article 35(9) recommends consulting data subjects “where appropriate.” For high-stakes deployments, consultation with affected populations (or representative groups) is increasingly expected.

6. AI Act overlap

If the system is high-risk under the AI Act, document the FRIA elements. There is substantial overlap with the DPIA, but the FRIA additionally requires:

Many practitioners maintain a single integrated DPIA+FRIA document with clear cross-references.

Sections most often skipped

In reviews of practitioner DPIAs, the sections most consistently underdeveloped are:

These are also where regulators most often find deficiencies on review.

Maintaining the DPIA

A DPIA is not a one-time document. The EDPB recommends review on a regular basis. For LLM workflows, the triggers for review include:

A practical cadence is quarterly review for production systems, with re-issuance on material change.

Cross-references

For the Article 22 framework that the DPIA must address head-on, see GDPR Article 22 and LLM-driven decisions. For the EU AI Act transparency layer that the DPIA should cross-reference, see Article 50 transparency obligations. For the incident-reporting framework that informs the security-risk section, aiincidents.org’s incident logging methodology describes the public record DPIA authors should consult.

What a regulator will look for on inspection

In reviewing DPIAs during inspections, member-state DPAs have consistently flagged:

A DPIA written specifically for the deployment and maintained as the deployment evolves is the artifact regulators credit. A generic template filled in once and never revisited is the artifact that produces findings.

For more context, AI policy watch covers related topics in depth.

Sources

  1. GDPR Article 35
  2. EDPB DPIA Guidelines (WP 248 rev.01)
  3. EU AI Act — Article 27 (FRIA)
#dpia #gdpr-article-35 #compliance #llm #risk-assessment
Subscribe

AI Privacy Report — in your inbox

AI privacy regulation, compliance, and enforcement, sourced. — delivered when there's something worth your inbox.

No spam. Unsubscribe anytime.

Related

Comments