Seven principles in GDPR Article 5 govern every AI agent that touches personal data: lawfulness, fairness and transparency; purpose limitation; data minimisation; accuracy; storage limitation; integrity and confidentiality; and accountability. Each becomes a specific control on an agent: what it may read, what it may keep, what it may do, and what record proves it.
What does Article 5 require?
Article 5 sets the principles that all processing of personal data must follow. In substance, personal data must be:
- Lawful, fair and transparent. Processed lawfully, fairly and in a transparent manner in relation to the person.
- Purpose limited. Collected for specified, explicit and legitimate purposes and not further processed in a manner incompatible with them.
- Minimised. Adequate, relevant and limited to what is necessary for the purposes.
- Accurate. Accurate and, where necessary, kept up to date, with every reasonable step taken to correct errors.
- Storage limited. Kept in a form that identifies people for no longer than necessary.
- Secure. Processed with appropriate security, including protection against unauthorised or unlawful processing and accidental loss, destruction or damage.
And the controller is responsible for, and must be able to demonstrate, compliance with all of them. That is accountability, the seventh principle, and the one that turns the others into evidence.
Who must apply them?
Every controller processing personal data under GDPR, and in their own sphere, every processor. A firm deploying agents is the controller for what those agents do with personal data, regardless of who built the agent or the model behind it.
How does each principle apply when the processor of data is an agent that acts?
Principles written for databases and forms take on new shapes when the system reads widely, writes freely and acts on its own.
Lawfulness, fairness, transparency. Each purpose an agent serves needs a lawful basis. Fairness covers the agent's behaviour: an agent that pressures, misleads or treats groups unevenly processes data unfairly even if every field is lawful. Transparency means people know an agent is involved and what it does with their data.
Purpose limitation. An agent that reads a customer's complaint history to answer a balance question, or carries data from one case into another, has stretched purpose. Scope retrieval to the task.
Data minimisation. The most common agent failure. Agents tend to retrieve whole records and long histories because it helps them answer. Minimisation means field-level access per capability and pruning context between steps.
Accuracy. Agents can write unsupported inferences into records as fact. Every fact an agent writes should carry where it came from.
Storage limitation. Logs, conversation history, caches and memory all hold personal data. Each needs a retention period. Memory is the easiest to forget; our answer to whether GDPR applies to what an AI agent remembers is yes.
Integrity and confidentiality. An agent that reads outside content can be instructed by it. Security for agents includes making sure an instruction hidden in an email cannot cause disclosure or action.
Accountability. The controller must show compliance. For agents, that means a record of each action, its sources and its verdict.
How does each principle become a control?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Lawfulness | Each capability mapped to a purpose and lawful basis | Record of processing entries |
| Fairness | Outcome monitoring by group; brief bars pressure tactics | Monitoring reports; brief restrictions |
| Transparency | Notice that an AI agent is used and how; disclosure at interaction | Privacy notice; channel captures |
| Purpose limitation | Retrieval scoped to the case and capability; no cross-case context | Retrieval rules; session design |
| Data minimisation | Field-level access per capability; context pruned between steps | Access maps; agent configuration |
| Accuracy | Agent writes only to designated fields, with origin labels | Field permissions; provenance on records |
| Storage limitation | Separate retention for logs, history, caches and memory | Retention schedule; deletion job records |
| Integrity and confidentiality | Outside content marked and unable to authorise actions; capability bounds | Provenance in records; refusal records |
| Accountability | Append-only record of every action, source and decision | Retained records; DPIA |
The DPIA is where these controls are assessed together; see GDPR Article 35 DPIA for AI agents.
How do other principles-based regimes compare?
GDPR's principles are deliberately general. They tell you what must be true, not how. That leaves agent design to the controller, a point we make in principles-based regulation leaves the agent design to you. Singapore's MAS FEAT principles take a similar approach for financial firms' use of AI, with fairness, ethics, accountability and transparency at the centre; MAS FEAT principles applied to AI agents maps them in the same way.
What will a supervisor or auditor ask?
- For this agent, show us each principle and the control that meets it.
- What can the agent read that it does not need?
- How long do you keep what the agent produces and remembers?
- How do you know the agent treats people fairly?
- Could an outsider cause the agent to disclose data? How do you know?
What do the principles not settle?
They do not set technical standards, retention periods or fairness metrics; the controller decides and justifies them. They do not address the EU AI Act's separate duties, which overlap in places but add others. And they do not decide what actions an agent may take with money or records, beyond requiring that processing be lawful, fair and secure. For the full map, 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.
Where Pryme fits
Several principles meet in one design choice in Pryme: every fact reaching an agent arrives through the Pryme Context Protocol stamped with its origin, and every action meets five bounds, including role scope and sources. That serves accuracy, purpose limitation and security together, and the filed record of verdicts serves accountability.
Questions people ask
Which principle do agents breach most often?
Data minimisation. Agents perform better with more context, so they are often given access to whole records. Field-level access per capability is the fix.
Does fairness under GDPR mean the same as fairness under the AI Act?
They overlap. GDPR fairness covers how data is processed in relation to the person, including expectations and effects. The AI Act focuses on bias and discrimination in high-risk systems. Test for both.
Can a vendor's agent meet the principles for us?
A vendor can provide controls, but the controller decides purposes, access and retention and must demonstrate compliance. The principles stay with you.