Privacy-Aware Infrastructure in the AI-Native Era: An Asset Classification Case Study

AI-native products introduce new data modalities, embeddings, multimodal inputs, and more, rapidly increasing the complexity of data management. In particular, even the same field name can be personal information (e.g., a person's age) in one context or simple system metadata (e.g., cache TTL) in another, meaning the name alone cannot determine privacy requirements.

Meta applies a hybrid Asset Classification pattern to address this problem. The core strategy is as follows:

• Build rich Context before the model performs inference
• Leverage LLMs to handle ambiguity, cold starts, and new data
• Separate human-reviewed labels from model-generated recommendations
• Convert stable behaviors into Deterministic Rules for execution

The ultimate goal is not to use LLMs everywhere, but to extract ambiguous signals once the LLM has learned them into low-latency, easily auditable deterministic logic for deployment in production. Through this, privacy-aware infrastructure (PAI) carries out four core operational tasks: understanding data, detection, enforcement, and proving compliance.