AI governance for managers: frameworks, roles and controls.

In short: AI governance is how an organization decides which AI to use, under what rules, with what evidence, and who answers for the result. In practice it is an operating model (owners, a policy, a process and a review), built on two frameworks most organizations now align to: the NIST AI Risk Management Framework, which is free and voluntary,¹ and ISO/IEC 42001, which is certifiable.² The rest of this guide shows what each part looks like and where to start.

The operating model: four layers

Organizations that formalize this often create an AI Management Office to run it. Start with the policy: the AI policy template covers approved tools, the data table, human review by tier and incidents.

Who decides what

Decision rights are where governance most often fails quietly: everyone is consulted and nobody decides. A simple allocation for four recurring decisions:

The table is a starting point to adapt; what matters is that each row has one name in the "decides" column and that the people in it know.

The frameworks: NIST AI RMF and ISO/IEC 42001

The NIST AI Risk Management Framework (AI RMF 1.0, January 2023) organizes AI risk work into four functions: Govern, Map, Measure and Manage.¹ It is voluntary and free, it names seven characteristics of trustworthy AI,¹ and its Generative AI Profile (NIST AI 600-1, July 2024) lists twelve risks specific to generative systems.³ In Texas, substantial compliance with a recognized framework such as the AI RMF is a defense under the state's AI law.⁴ You do not implement all of it: you write a profile for each system.¹ See the NIST AI RMF in two pages.

ISO/IEC 42001:2023 is the certifiable standard for an AI management system, with 38 reference controls under nine objectives in Annex A.² Certification is by an accredited body in a two-stage audit, with surveillance audits over a three-year cycle; ISO/IEC 42006:2025 sets the requirements for the bodies that certify.^5,6 See ISO/IEC 42001: build first, certify second and choosing a certification body. ISO/IEC 42001 certifies organizations, not people; for the individual credentials that cover this ground, see how AIMP compares with AIGP, AAISM, AAIA, CPMAI and the ISO/IEC 42001 Lead Implementer.

They fit together: a NIST profile per system supplies much of the evidence an ISO 42001 audit samples.

Risk tiers: proportionate control

Not every AI use needs the same scrutiny. A common design uses three tiers: high-impact uses that affect people's rights, money or safety (full assessment, legal review, fairness testing, human review, quarterly review); medium-risk uses inside the organization (light assessment, a named owner, evaluation on a golden set, sampled review); and low-risk uses pre-approved by policy (registration in the inventory). The tier decides the work, so the effort goes where the risk is.

Governing AI agents

Agents that act on systems, not only answer questions, need controls that chatbots did not: bounded tools and permissions, approval before irreversible actions, logs written by the platform rather than the agent, and autonomy widened only on measured evidence. Singapore's Model AI Governance Framework for Agentic AI (January 2026)⁷ and OWASP's Top 10 for Agentic Applications (December 2025)⁸ are the clearest public guidance. Further reading: four commitments worth copying, the agentic risk list, the Replit incident and what makes an AI agent an agent.

The evidence: testing, fairness, security and data

• Evaluate on your own cases. A golden set of real cases with agreed answers, run before buying and after every change. See the golden set.
• Test decisions about people. Selection-rate comparisons such as the four-fifths rule, with significance testing and a look at proxies.⁹ See fairness testing.
• Secure what the system reads. Retrieved content can carry instructions; permissions should follow the user.¹⁰ See EchoLeak and the OWASP lists and permission-aware retrieval.
• Know the data. Inventory, permitted use, quality, lineage and retention. See AI-ready data.
• Govern suppliers. Evidence-based vendor scoring and contract terms. See vendor due diligence and AI contract clauses.

When things go wrong

Documented AI incidents rose from 233 in 2024 to 362 in 2025 by Stanford's count,¹¹ and some are now legally reportable on clocks from 24 hours to 15 days.^12,13,14 A one-page procedure per system (severity levels, who can pause it, who is told, what evidence is kept) turns an incident into a managed event. See AI incident response and three documented AI failures.

People: literacy and trust

Governance fails when the people using AI do not know the rules or do not trust them enough to report problems. The EU AI Act requires organizations to support AI literacy,^14,15 and the evidence on adoption points to managers: where they use AI openly and set quality standards, teams get more value from it.¹⁶ See AI literacy after the omnibus and trust as a management skill.

How to tell whether governance is working

Measure the system, not the paperwork. Useful indicators: the share of AI systems in the inventory with a named owner and a current assessment; time from intake to decision; the number of exceptions open past their end date; incidents by severity and time to contain; how many systems were retired or narrowed after review; and, for the business, whether high-impact systems meet the quality bar set before launch. A governance program that only ever approves is not measuring anything. Our briefing on why AI productivity numbers mislead covers the measurement traps that apply here too.

Where to start this quarter

1. Build the inventory: every AI system, its owner, its tier and the people it affects.
2. Adopt a use policy with the data table, and publish it.
3. Write a NIST-style profile for your highest-risk system.
4. Set up intake so new uses are assessed before they are bought.
5. Schedule the first management review, with the questions it must answer.

Governance frameworks and the operating model are taught in Module 5 of the AIMP AI management certification, with risk and responsible AI in Module 7 and data, privacy and security in Module 8. For the legal side, see AI regulation for managers.