AI Governance Frameworks

01 Why AI governance matters

The shift from generative AI to agentic AI changes the stakes — autonomous systems take real actions on real systems, and "the model said so" is no longer an acceptable audit trail.

02 NIST AI RMF

NIST AI Risk Management Framework — issued by the U.S. National Institute of Standards and Technology. A voluntary, sector-agnostic framework for managing AI risks systematically across the AI lifecycle.

Purpose

Provide organizations with a structured way to identify, measure, and reduce risks from AI systems — covering bias, hallucinations, privacy violations, lack of explainability, robustness failures, security weaknesses, and inadequate human oversight.

Four core functions

Focus areas

When to use NIST AI RMF

• You operate in or sell to U.S. federal/regulated industries (finance, healthcare, defense).
• You want a vendor-neutral, lifecycle-oriented risk vocabulary your engineering, legal, and product teams can share.
• You need a framework that maps cleanly onto ISO 42001 and the EU AI Act for unified compliance.

03 EU AI Act

The European Union's landmark AI law — the world's first comprehensive horizontal regulation of AI systems. It classifies AI by risk level and imposes obligations proportional to the risk.

Risk-based classification

Requirements for high-risk AI

General-purpose AI (GPAI)

Foundation models like GPT-class LLMs face additional obligations: technical documentation, copyright compliance for training data, and — for models with systemic risk — model evaluations, adversarial testing, incident reporting, and cybersecurity protections.

04 Responsible AI (RAI)

Responsible AI is the broader philosophy and practice — adopted by Microsoft, Google, OpenAI, IBM, Anthropic, and most enterprise AI teams. While NIST is a framework and the EU AI Act is law, RAI is a set of principles operationalized into engineering practice.

Common RAI principles

How big tech operationalizes RAI

05 Framework comparison

The three are complementary, not competing. NIST gives you vocabulary, the EU AI Act gives you obligations, RAI gives you the principles your engineering culture lives by.

06 Governance for agentic AI

Agents change the threat model. A chatbot returns text; an agent calls APIs, modifies databases, sends emails, deploys code, moves money. Governance must shift from "what did the model say" to "what did the system do."

What makes agents different

07 Agent governance controls

Concrete controls you can put in production today. Each control answers a specific failure mode.

08 Reference architecture

A typical enterprise AI governance architecture for an agentic system. Each layer enforces a different control surface — bypass any one and you lose accountability.

Layer responsibilities

Sample policy rule

# Deny destructive DB operations in production
package agent.tools

default allow = false

allow {
  input.tool == "db_query"
  not contains(input.args.sql, "DELETE")
  not contains(input.args.sql, "DROP")
  input.env != "production"
}

allow {
  input.tool == "db_query"
  input.env == "production"
  input.args.sql == "SELECT"  # read-only in prod
  input.user.role in ["analyst", "engineer"]
}

09 Modern hot topics

Where the field is moving. Each of these is a live area of standards, tooling, and research investment.

10 Implementation checklist

A pragmatic starting point. Don't try to ship all of it on day one — sequence by risk.

Foundations (week 1–4)

• Inventory every AI system and assign a named owner.
• Classify each system by risk tier (use EU AI Act tiers as a starting point even if you're not in the EU).
• Stand up centralized audit logging with immutable storage.
• Add identity propagation so every model call is bound to a verified user.

Hardening (month 2–3)

• Introduce a policy engine in front of all tool calls.
• Define per-tool allowlists per agent role.
• Wire human approval gates for irreversible / high-cost actions.
• Add safety classifiers on inputs and outputs (PII, toxicity, jailbreaks).
• Stand up an eval pipeline with golden tests + adversarial probes.

Maturity (month 4+)

• Continuous red-teaming with rotating attack suites.
• Drift detection on inputs, outputs, and tool-call distributions.
• Memory governance: namespaces, TTLs, redaction, audits.
• Incident response playbook with kill-switch drill quarterly.
• Map controls to NIST RMF + EU AI Act articles for unified compliance reporting.

11 Glossary

Quick definitions for terms that recur across NIST, the EU AI Act, and RAI documentation.

12 One-line summary