An AI agent asked to schedule a medical follow-up may read a discharge note, infer a health condition, and add it to a persistent profile. The task has shifted from scheduling to categorizing sensitive health data—a form of purpose creep. section 5.6.1
Under the GDPR, purpose limitation requires personal data to be collected for specified, explicit, and legitimate purposes and not further processed in incompatible ways. Agentic systems put this principle under strain when they pursue evolving goals, reuse data in real time, or retain memories that support new inferences. section 5.6.1
From scheduling to profiling
An agent scheduling a follow-up accesses an inbox for context and sees a discharge note mentioning diabetes. It then creates an “endocrinology” profile tag for future conversations. Although the note helped provide context, the new categorization serves a different purpose from the requested appointment. section 5.6.1
Persistent memory can extend this risk across months of interactions: an agent may infer patterns in mood, stress, or decision-making and build a profile the user never agreed to. Purpose limitation also matters when customer data is repurposed beyond its initial scope or vendors use it for continued training. section 5.6.1 section 5.7
For monitoring logs that may contain personal data:
- Document a specific monitoring or compliance purpose; a general statement that logs are needed for the EU AI Act is too vague. section 7.5.1
- Log only what monitoring and compliance require, document each personal data attribute, and do not repurpose the logs. section 7.5.1
- Treat prompt and output logs as containing personal data unless evidence shows otherwise, and apply minimization and access controls. section 7.5.1
- Document the basis for retaining personal data after deletion requests; logs necessary for a legal compliance obligation may remain for the required retention period. section 7.5.1
