AI Executive Office — CXO Assistant Platform (On-Premises)  ·  View 29 of 30  ·  7 · Assurance

AI Governance and Guardrails

Every control on the path of a single turn, and the three ways the platform is allowed to not answer.

Editable source SVG draw.io All views
Input Prompt shield jailbreak detection PII detection Presidio Scope check role · tenant Retrieval ACL trimming before scoring Sensitivity labels labels honoured Content is data never instruction Generation Pinned model version recorded Pinned prompt from the registry Tool allow-list no write tools Output Claim binding evidence id per claim Groundedness NLI verifier Protected material Egress DLP label-aware Fail safe Abstain "not enough evidence" Answer partially and say which part Escalate to a human analyst queue Record Full turn trace prompt · tools · model Feedback accepted · rejected Into the golden set next release unsupported AI Governance — Every Control on the Path of a Turn Security / platform Application we own Risk / gap Journey / task Person or role Data store failure / alternate Abstention is a first-class outcome with its own metric. Partial visibility answers partially and names the gap; a platform that never abstains is not more accurate, only less honest. v 1.0 · owner Data & AI Global Practice · date 2026-09

Decisions

  • Abstention is a first-class outcome with its own metric and its own review. A platform that never abstains is not more accurate, only less honest — and executives calibrate on the answers that turn out wrong
  • Groundedness is checked mechanically against retrieved evidence, not asserted by the generating model. Claims that fail are removed, and the removal is recorded
  • Model version and prompt version are pinned per tenant and written into the record. An answer is only reproducible if both are known
  • Tool descriptions are treated as untrusted input. They are pinned and reviewed with the server version rather than read live, because a description is text the model obeys

The three refusals

  • Abstain entirely: no sufficient evidence, and say so in those words
  • Answer partially: name which domain or which period is missing rather than quietly narrowing the question
  • Escalate to a human analyst: for questions the platform can frame but not resolve

Risks

  • Guardrails add latency to every turn and will be the first thing questioned when the 5-second target is missed. The groundedness check runs on the claim set rather than the full text to keep it affordable, and the budget is stated rather than discovered
◀  Sovereignty and Residency All views Audit and Lineage  ▶
Solution Architecture v1.1 · fully on-premise, open source, multi-tenant, sovereign-capable · Data & AI Global Practice · 2026-0929-ai-governance-and-guardrails