GDPR for AI Agent Projects: What to Check Before You Build
If your AI agent reads, stores or acts on information about identifiable people, such as customer emails, CRM records, support tickets or CVs, the GDPR applies to it exactly as it applies to any other software. The short GDPR AI agent checklist for a small company has six items: know which personal data the agent touches, choose a lawful basis for each purpose, minimise the data it sees, sign processor contracts with model vendors, secure a valid mechanism for international transfers, and screen for a data protection impact assessment (DPIA). Then design logging, retention and human oversight, and check which EU AI Act duties already apply.
This is practical guidance for founders and CTOs, not legal advice. For your specific case, involve your data protection officer or counsel.
When does an AI agent process personal data?
An AI agent is software that uses a language model to decide which steps to take and which tools to call to reach a goal, instead of following a fixed script. If you are still choosing between a fixed pipeline and an agent, our guide on choosing between an AI workflow and an AI agent explains the trade-offs; the data protection picture is usually simpler for a fixed workflow.
Under Article 4 of the GDPR, personal data is any information relating to an identified or identifiable person, including names, location data and online identifiers, and processing covers almost any operation on it, from consultation to storage. In a typical agent, personal data appears in five places:
- Inputs: prompts, uploaded files and forwarded emails.
- Retrieval: documents, tickets or CRM records used as context.
- Tool calls: data read from or written to other systems.
- Outputs: drafted replies, summaries and decisions about people.
- Logs and traces: stored prompts, responses and tool results.
Map these flows on one page before any code is written.
Which lawful basis fits your use case?
Article 6(1) GDPR lists six lawful bases: consent, contract, legal obligation, vital interests, public task and legitimate interests. For business agents, the realistic options are usually contract (handling a customer’s own request) or legitimate interests (triaging inbound support email), which requires a documented balancing test. Two further rules often catch agent projects: purpose limitation in Article 5(1)(b) means support data is not automatically available for a new purpose such as fine-tuning, and special categories of data under Article 9, such as health data, need an additional legal condition.
For DACH teams, the German data protection authorities published an orientation guide on AI and data protection on 6 May 2024, focused on selecting and using LLM-based systems.
How do you apply data minimisation to an agent?
Data minimisation (Article 5(1)(c) GDPR) requires personal data to be limited to what is necessary, and Article 25 requires data protection by design and by default. Agents tend to over-collect, because it is easy to hand the model a full CRM export. Practical measures:
- Give each tool the narrowest scope: read-only where possible, specific fields instead of whole records.
- Redact or pseudonymise identifiers the task does not need before text reaches the model.
- Retrieve only records relevant to the current task.
- Use synthetic or anonymised data in development and testing.
Our AI agent security guide for engineering leads covers tool permissions, prompt injection and approval gates.
What must your contracts with model vendors cover?
When you decide why and how personal data is processed, you are the controller; a model provider processing it on your behalf is a processor. Article 28(3) GDPR requires a written contract covering the subject matter, duration, nature and purpose of processing, the types of data and data subjects, and the processor’s duties: acting only on documented instructions, security, sub-processors, assistance with data subject requests, and deletion or return of data. The same applies to a development partner with access to production data.
Check each vendor’s current terms. OpenAI’s enterprise privacy page states that it does not train on business data by default, offers a Data Processing Addendum for API customers, and may retain API data for up to 30 days for abuse monitoring, with zero data retention for eligible cases. Anthropic’s privacy centre states that, by default, it does not use inputs or outputs from its commercial products for training. For every vendor, record the contracting entity, retention, sub-processors, hosting region and deletion process.
How do international transfers work for US-hosted models?
Sending personal data to a provider outside the EEA is a transfer under Chapter V of the GDPR. For the United States, the European Commission adopted an adequacy decision for the EU–US Data Privacy Framework on 10 July 2023. It covers only US companies that have self-certified, so check the vendor on the official Data Privacy Framework list.
The General Court dismissed an action to annul the framework on 3 September 2025, and an appeal to the Court of Justice was lodged in October 2025. As a fallback, use a vendor DPA that also includes standard contractual clauses under Article 46(2)(c) GDPR.
When do you need a DPIA?
Article 35 GDPR requires a DPIA when processing, in particular using new technologies, is likely to result in a high risk to people’s rights and freedoms. The French regulator CNIL explains that, under the European criteria, a DPIA is generally expected when processing meets at least two of them, such as sensitive data, large scale, vulnerable people, combined datasets or innovative technology. It also names AI-specific risks, including automated discrimination and fabricated content about real people.
| Agent scenario | Main GDPR focus | DPIA likely? |
|---|---|---|
| Drafts replies to B2B support tickets, a person sends them | Minimisation, vendor DPA, retention | Often not; document the screening |
| Summarises sales calls and updates the CRM | Transparency to participants, lawful basis | Possible, depending on scale |
| Screens or ranks job applicants | Article 22, fairness, special categories | Yes, in most cases |
| Monitors employee activity | Systematic monitoring, workplace rules | Yes, in most cases |
| Handles health, financial or children’s data | Article 9 conditions, security | Yes, in most cases |
What should you log, and for how long?
Stored prompts and tool results are personal data too. Storage limitation (Article 5(1)(e) GDPR) means keeping them only as long as needed; Article 30 requires records of processing, Article 32 appropriate security, and Article 33 notifying the supervisory authority of a breach within 72 hours where feasible, unless it is unlikely to result in a risk. A workable pattern is short retention for full traces, longer retention for redacted metrics, and restricted access to raw logs. Align this with your agent evaluation and release criteria, since test sets built from real conversations need the same treatment, and with your incident response plan for AI agents.
Where does human oversight fit?
Article 22 GDPR gives people the right not to be subject to a decision based solely on automated processing that has legal or similarly significant effects, with narrow exceptions. Keep a person in the loop for hiring, credit or account termination decisions, and add approval steps before irreversible actions such as external emails, deletions or payments.
How does the EU AI Act relate to GDPR for agent projects?
The AI Act does not replace the GDPR; both apply in parallel. According to the European Commission, it entered into force on 1 August 2024, prohibitions and AI literacy rules have applied since 2 February 2025, general-purpose AI model obligations since 2 August 2025, and the regulation became generally applicable on 2 August 2026.
The AI Omnibus, Regulation (EU) 2026/1744, published on 24 July 2026, moved high-risk obligations to 2 December 2027 for stand-alone systems in Annex III, such as recruitment tools, and to 2 August 2028 for AI in products under Annex I. It also reworded the AI literacy duty so that providers and deployers must take measures to support staff AI literacy. Transparency duties under Article 50 of the AI Act apply from 2 August 2026: an agent that talks directly to people must make clear they are interacting with AI, unless that is obvious. For marking AI-generated content, systems already on the market before that date have until 2 December 2026. Our article on the EU AI Act in 2026 for engineering teams covers classification and documentation.
Pre-build checklist
- Data flow map: inputs, retrieval, tools, outputs, logs.
- Lawful basis per purpose, with a legitimate interests assessment where used.
- Minimisation: tool scopes, redaction, test data.
- Signed DPAs with model vendors and any development partner.
- Transfer mechanism: DPF certification, SCCs or EU hosting.
- DPIA screening documented; full DPIA where two or more criteria apply.
- Retention periods for traces, test sets and outputs.
- Human approval for significant decisions and irreversible actions.
- AI Act check: AI disclosure, risk classification, staff AI literacy.
- Privacy notice updated to describe the agent.
Frequently asked questions
Does GDPR apply if we only call a model through an API?
Yes. If prompts or retrieved context contain personal data, you are processing it and the API provider is your processor. You need a DPA under GDPR Article 28 and, for providers outside the EEA, a valid transfer mechanism.
Do we need a DPIA for every AI agent?
No, but screen every agent and document the result. A DPIA is generally expected when two or more high-risk criteria apply, according to the CNIL’s DPIA guidance.
Can we use customer conversations to evaluate or fine-tune the agent?
Only if the new purpose is compatible with the original one or has its own lawful basis (GDPR Article 6(4)). Redacted or synthetic test sets reduce the risk.
Is the EU–US Data Privacy Framework still valid?
Yes, the adequacy decision remains in force. The General Court upheld it in September 2025 and an appeal was lodged with the Court of Justice, so keep standard contractual clauses as a fallback.
Next step
If you are planning an AI agent and want data flows, vendor choices and oversight points clarified before the build, Powercode Group can scope it with you through its AI and data science services, with transparent costs and timelines before you commit. Powercode Group is an EU-based engineering group. Contact Powercode Group to discuss your use case.
