Seven pillars from governance intent to technical evidence

Does your organization maintain a complete inventory of AI, GenAI, RAG, Copilot, and agentic systems currently in use?

Includes shadow AI, embedded vendor features, and internally built agents.

Evidence: Export from AI system registry, CMDB tags, procurement records.

Are AI use cases classified by risk tier, data sensitivity, and regulatory scope?

Aligned to internal risk taxonomy and applicable regulation (EU AI Act, NIST AI RMF).

Evidence: Risk classification policy and per-use-case classification records.

Is the inventory continuously updated as new AI capabilities are deployed?

Automated discovery vs. periodic manual attestation.

Evidence: Discovery pipeline logs, attestation cadence records.

Can your audit logs identify the original human user behind AI-triggered actions, even when service principals, APIs, or application identities are used?

End-to-end identity propagation through agent chains and downstream systems.

Evidence: Sample audit trail joining human identity to agent action.

Are non-human identities (agents, service principals, API keys) inventoried and owned?

Every workload identity has a documented owner and lifecycle.

Evidence: Workload identity inventory with owner and rotation policy.

Is delegated authority for AI agents scoped, time-bound, and revocable?

Prevents agents from acting with broader privileges than the requesting user.

Evidence: Token scoping policy, delegation logs.

Can AI-generated outputs be traced back to source data, retrieval context, model, prompt, tool, and authorization path?

Full provenance for a given inference or agent action.

Evidence: Trace record for a sample inference including retrieved chunks.

Is lineage maintained across ingestion, transformation, vector indexing, and retrieval?

Column-level or chunk-level lineage where applicable.

Evidence: Lineage graph export from catalog or observability tool.

Are model, prompt, and tool versions recorded with each inference?

Enables reproduction and forensic review of past outputs.

Evidence: Inference log schema and retention policy.

Are access controls technically enforced across data, models, prompts, tools, and downstream AI applications?

Not merely documented — enforced at the data plane and gateway.

Evidence: Policy engine configuration, deny-by-default proof.

Do RAG systems enforce the requesting user's data entitlements at retrieval time?

Prevents privilege escalation via shared vector indexes.

Evidence: Retrieval filter logic, entitlement check traces.

Are prompt injection, tool misuse, and data exfiltration risks technically mitigated?

Guardrails, output filtering, egress controls.

Evidence: Guardrail policy, red-team results.

Are data quality checks, schema validation, drift monitoring, and pipeline telemetry implemented before AI systems consume governed data?

Prevents silent degradation of AI outputs from upstream data issues.

Evidence: DQ dashboard, drift alert history.

Is sensitive data classified and masked before entering training, fine-tuning, or retrieval corpora?

Automated classification and redaction with audit.

Evidence: Classification coverage report, redaction logs.

Are model and agent behaviors monitored in production for accuracy, safety, and drift?

Observability beyond infrastructure metrics.

Evidence: Model observability dashboard, incident review records.

Can your organization produce evidence showing who approved an AI use case, what data it uses, what risks were reviewed, and what controls were implemented?

On-demand evidence package for auditors and regulators.

Evidence: Sample evidence pack from a live use case.

Are AI incidents, near-misses, and control failures logged and reviewed?

Feeds continuous improvement of controls.

Evidence: Incident register with root cause and remediation.

Is evidence retention aligned with regulatory and contractual obligations?

Retention schedule mapped to regulatory scope.

Evidence: Retention policy and enforcement configuration.

Are AI governance policies technically enforced, or are they primarily documentation and training?

The gap between written policy and runtime behavior.

Evidence: Policy-to-control mapping with enforcement evidence.

Is there a defined AI governance operating model with clear accountability?

Named owners across risk, data, security, legal, and platform.

Evidence: RACI and governance charter.

Are governance decisions and exceptions tracked with a defined escalation path?

Enables oversight of risk acceptance and deviations.

Evidence: Exception register and review cadence.