A Copilot agent helps HR staff draft performance review templates by retrieving historical review examples from a database. The database contains real employee reviews with names, performance ratings, and compensation data. The agent retrieves these as context and occasionally surfaces employee names and salaries in its responses to other users. Which data handling practice should be implemented to address this GDPR and privacy concern?
Select an answer to reveal the explanation.
Short Explanation and Infographic
Loading real PII into an AI agent's context is like putting personnel files on a conference table during a meeting—anyone in the room (or who enters the room) can see them. Anonymization before context loading means the agent gets the structural patterns it needs (how a good review is written) without the private data it doesn't.
Full explanation below image
![A 2048x1536 landscape educational infographic titled 'PII Anonymization Before Agent Context Loading'. Left: a database record showing 'Name: John Smith, Rating: 3/5, Salary: $95,000' entering an agent context window—red privacy violation arrows showing 'John Smith' appearing in another user's response. Right: the same record after an 'Anonymization Layer': 'Name: Employee A, Rating: [Performance Category], Salary: [Tier]'—enters context window → agent generates template using only structure and language, no PII surfaces (green). Center: GDPR data minimization principle icon. Footer: 'Agents need patterns, not personal data. Anonymize before context loading.' Clean data flow diagram, dark navy background.](/media/github/gh-600/GH-600-e5-q023-infographic.webp)
Full Explanation
The core privacy principle violated here is data minimization: agents should only process the personal data necessary for their task. To generate a performance review template, the agent needs to understand the structure and language patterns of reviews—it does not need the actual employee names, their specific ratings, or compensation figures.
Anonymization/pseudonymization approach: - Replace real names with synthetic identifiers: 'Employee A', 'Team Member B' - Replace specific compensation figures with ranges or generic placeholders: '$85,000' → '[Compensation Tier 3]' - Replace identifiable project names with generic references: 'Project Phoenix' → '[Major Product Initiative]' - This transformation should happen at the data retrieval layer, before records enter the agent's context window
Why GDPR applies: Even when HR staff are authorized to access the underlying data, feeding it to a third-party AI model (GitHub Copilot) may constitute a new processing activity that requires its own legal basis, data processing agreement, and data minimization obligations.
Option A (restrict to HR staff) addresses access control but not the data minimization principle—HR staff using the agent may still inadvertently expose PII to other users if the agent surfaces it in shared contexts. Option C (role-based response filtering) doesn't prevent the PII from being in the context; it only limits who sees the responses. If the model routes context across sessions, restriction on responses alone is insufficient. Option D (disclaimer) is a legal notice, not a technical control, and does nothing to prevent the privacy violation.