Inference Residency

Explanation

Inference Residency checks whether AI inference (the live model calls that process customer data) actually runs inside the approved geographic/residency boundary and that no data leaks out through unapproved network egress — closing the gap where a model endpoint or third-party API quietly sends sensitive banking data abroad. It matters for RBI data-privacy and data-localisation obligations: even a correctly governed model breaches the law if its inference endpoint sits in, or sends data to, an unapproved location. The metric and method are Inference Endpoint Residency & Egress — verifying where inference executes and what leaves the boundary. To implement, pin every inference endpoint (including embedding, retrieval and document-extraction services in RAG/chat pipelines) to approved regions, enforce egress controls and allowlists at the network layer, continuously verify endpoint residency, and log endpoint location, destination of any outbound calls and blocked egress attempts as evidence aligned to NIST MANAGE 3.2 data localisation and ISO Annex A transfer controls. The threshold is 100% of inference within approved residency and zero unapproved egress; any inference outside the boundary or any unapproved data egress is an immediate-priority breach that must block the flow and trigger investigation.