An AI privacy guard is middleware: it sits between your application input and the model provider, transforming sensitive values before the model request is sent.
Send textual input to the privacy engine.
Replace detected spans with scoped placeholders.
Call whichever supported AI provider the customer chose.
Replace exact placeholders in the model response using short-lived application mapping.
The privacy layer does not need to own the model.
A text-wrapper should not be assumed to cover files, images, audio, embeddings or tool payloads until those paths are explicitly integrated and tested.