CORE GUIDE
Privacy and data boundaries
AI systems need explicit rules for what data may enter models, logs, retrieval indexes, memory and third-party tools, because each boundary creates a different exposure and retention risk.
Mental model
Data moves through trust domains. Privacy engineering maps each class of information to allowed purposes, destinations, retention periods and access controls before it is serialized into prompts or persisted as AI context.
Why it matters
A useful AI feature can accidentally multiply sensitive data. User messages may be copied into traces, embeddings, long-term memory, support tickets and external model calls. Once data spreads across those surfaces, deletion and access control become harder. Clear data boundaries reduce unnecessary collection and keep private information from becoming ambient context simply because it might help the model.
01
Classify data before choosing where it flows
Identify sensitive fields and the purpose that justifies each use. Define whether the data may be sent to a model provider, indexed for retrieval, written to durable memory, included in observability or exposed to a tool. Minimize payloads, separate tenant or user scopes, apply retention and deletion policies, and record provenance so derived data such as embeddings can be traced back to the source lifecycle.
02
Example: building a support knowledge base
Product documentation can be broadly indexed, while customer tickets contain account identifiers and private conversation. The system stores tickets in tenant-scoped operational storage and retrieves only the small relevant excerpt for an authorized support task. It does not mix those tickets into a global knowledge index or persist the full conversation as reusable memory.
Common failure modes
- Sending every available customer field to the model because more context might help.
- Treating embeddings or summaries as non-sensitive just because they are derived data.
- Keeping AI traces indefinitely without a retention or deletion path.
Engineering heuristics
- Minimize data at every model, retrieval, tool and logging boundary.
- Preserve tenant and user scope through derived artifacts such as indexes and memory.
- Design deletion and retention behavior before data is copied into multiple subsystems.
Takeaways
- 01Privacy is a data-flow architecture problem.
- 02Derived AI artifacts can retain the sensitivity of their sources.
- 03Purpose, scope and retention should constrain context before generation begins.
Used in
This Concept is reused across these canonical learning paths.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.