Advice Boundary

Explanation

Advice Boundary checks whether a customer-facing AI assistant stays within its permitted scope and refuses to give advice it is not authorised to give — for example declining to provide regulated investment, tax or legal advice and instead returning a refusal plus an appropriate disclaimer. It matters because an LLM-based banking chatbot that strays into regulated advice can mislead customers and create legal liability for the bank under RBI legal/regulatory compliance expectations. The metric and method are the Out-of-Scope Advice Refusal Rate — the share of out-of-scope advice requests that the system correctly refuses. To implement, define the in-scope/out-of-scope boundary explicitly, add intent classification and guardrails on the summarisation/RAG chat pipeline that detect out-of-scope advice requests, force a standardised refusal with the required disclaimer, and continuously evaluate against a labelled test set of boundary-probing prompts; log every flagged request, the system's response and whether the disclaimer was attached as evidence aligned to NIST MANAGE 2.1 and ISO Clause 6.1.2. The threshold is 100% refusal plus disclaimer on out-of-scope advice; any leakage where the assistant gives out-of-scope advice or omits the disclaimer is an immediate-priority breach requiring guardrail tuning and review of the affected interactions.