decides
Accountable executive
Owns AI risk appetite, signs off the policy, and answers to the board for AI use.
Guide · AI governance
An AI governance framework you can run: five parts, one loop.
Who is accountable, what AI you have, how risky each system is, which controls that risk demands, and how you know they still hold.
Published frameworks describe outcomes. This guide turns them into an operating design: the roles to name, the fields an AI inventory needs, a tiering rule, the controls each tier owes before and after release, and the monitoring loop that keeps an approval valid — then maps every part to the NIST AI Risk Management Framework and ISO/IEC 42001.
Updated · 7 sources
The five parts of an AI governance framework
01
Roles and accountabilityan executive owner, a review body, system owners, independent challenge
NIST Govern 2ISO 5.302
AI inventoryevery model, agent and AI feature, with purpose and owner
NIST Govern 1.603
Risk tiersimpact × autonomy decides how much scrutiny a system gets
NIST Map 5ISO 6.1.204
Controls per tiertesting, oversight, change control, records
NIST MeasureISO Annex A05
Monitoring and reviewdrift, incidents, re-approval, retirement
NIST Manage 4ISO 9What is an AI governance framework?
An AI governance framework is the set of roles, policies, processes and controls an organization uses to decide which AI systems it builds or buys, how much risk each carries, what must be shown before each goes live, and how each is monitored, changed and retired. A usable one has five parts — roles, an inventory, risk tiers, controls per tier and monitoring — and can be mapped to the NIST AI RMF and certified against ISO/IEC 42001.
Part 1 · Accountability
Roles in an AI governance framework: who owns, who checks, who audits.
NIST asks that roles and responsibilities for AI risk be documented and that executive leadership take responsibility for decisions about AI risk. In practice that means separating the people who build a system from the people who challenge it.
decides
AI governance committee
Legal, risk, security, data and business leads. Approves high-tier systems and exceptions.
first line
System owner
The business lead for one AI system: purpose, inventory entry, controls, incidents.
first line
Builders
Engineers and analysts who configure models and agents, write the tests and fix findings.
second line
Independent reviewer
Model risk or a review team that did not build the system — NIST names internal experts who did not serve as front-line developers.
third line
Internal audit
Checks that the framework is followed and works, without redoing the validation.
Also named in a complete framework
- Data ownerapproves the sources a system learns from or reads at run time
- Privacy and legaldecide which laws apply and review impact assessments
- Securityreviews tool access, secrets and attack testing
- Human overseersthe people who review, override or stop the system in use — trained for it
Parts 2 and 3 · Inventory and risk tiers
The AI inventory and the risk tiers that decide scrutiny.
You cannot govern what you have not listed. Each entry carries enough to tier the system; the tier then decides which controls it owes. Tier on two things: how much harm a wrong output can do, and how freely the system acts without a person.
Fields every inventory entry needs
- Name and ownerone accountable person
- Intended purposewhat it decides or produces, and for whom
- Model and versionincluding the provider for bought models
- Tools and actionswhat an agent can change, not only read
- Datasources it learns from, is tested on and retrieves
- People affectedstaff, customers, applicants, the public
- Risk tierwith the reasoning written down
- Status and datesin build, approved, live, under review, retired
High impact
Tier 2Drafts or recommends where a wrong answer is costly — a person decides on each case.review each output
Tier 1Acts or decides where errors hurt people, money or compliance. Full controls.full controls
Low impactTier 3Summaries, search, drafting for internal use. Light controls.register and monitor
Tier 2Automates routine, reversible steps. Tested, monitored, sampled.test and sample
Assists a personActs on its ownBefore tiering, check the legal class (EU AI Act, where it applies).
A prohibited practiceArticle 5 — such as social scoring or emotion recognition at work. Do not build or deploy it.
An Annex III useCredit, hiring, education, essential services and more — high-risk duties apply from 2 December 2027. Tier 1 at least.
Talks to people or makes contentArticle 50 transparency duties apply from 2 August 2026: tell people they deal with AI, mark synthetic content.
None of theseTier by your own impact and autonomy rule.
Part 4 · Controls
Controls for each risk tier, before release and after.
The tier decides the depth of each control. The table is an example to adapt — set your own thresholds — but every row should exist in some form for every tier, because a light control is still a control.
| Control | Tier 3 | Tier 2 | Tier 1 |
|---|---|---|---|
| Risk and impact assessment | Short form by the owner | Full form, reviewed | Full form, committee approval; legal impact assessments where law requires |
| Testing before release | Sample check | Golden dataset with pass thresholds | Golden dataset, independent validation, attack testing |
| Human oversight | Users can flag errors | Review of low-confidence cases | Approval on actions; people can override and stop |
| Change control | Owner approves | Re-test on every change | Re-validate on a new model, prompt or tool |
| Records | Run logs | Run logs and decisions, kept to policy | Tamper-evident audit trail, retention set by law |
| Monitoring | Periodic review | Monthly metrics and drift | Continuous drift alarms, incident process |
| Transparency | Labelled as AI | Sources shown with outputs | Explanations for affected people, notices |
Generative AI adds its own risks. The NIST Generative AI Profile lists twelve, from confabulation and data privacy to information security and value-chain integration. Use it as the test catalogue for Tier 1 and Tier 2 assistants and agents.
Part 5 · Monitoring and review
Monitoring: the loop that keeps an approval valid.
An approval is a statement about a system on the day it was tested. Inputs drift, models are updated, and agents get new tools. NIST’s Manage function asks for post-deployment monitoring plans that include override, incident response, recovery and change management.
The AI system lifecycle under governance
- Registerthe system enters the inventory with an owner
- Assesspurpose, people affected, legal class, tier
- Testagainst golden data and known risks
- Approvego or no-go, by the right body for the tier
- Operatewith human oversight and full records
- Monitorquality, drift, complaints, incidents
A material change, a drift alarm or an incident sends the system back to assessment — or to retirement.
Mapping
Mapping the AI governance framework to NIST AI RMF and ISO/IEC 42001.
The two references play different roles. NIST describes outcomes to aim for and is voluntary; ISO/IEC 42001 sets requirements for a management system an auditor can certify. A framework built from the five parts above satisfies the structure of both.
NIST AI RMF 1.0
Published by the US National Institute of Standards and Technology in January 2023, with a Generative AI Profile added in July 2024. NIST states it is being revised.
- Four functions, categories and subcategories
- Outcomes, not mandatory requirements
- A Playbook of suggested actions
- No certification
ISO/IEC 42001:2023
The international standard for an AI management system, published in December 2023, built on the same clause structure as other ISO management systems.
- Requirements in clauses 4 to 10
- Annex A reference controls
- Statement of which controls apply
- Certifiable by an accredited auditor
Govern 2: roles, training, executive responsibilityRoles and accountabilityClause 5: leadership, AI policy, roles; Annex A.3
Govern 1.6: mechanisms to inventory AI systemsAI inventoryClause 4: context and the scope of AI systems covered
Govern 1.3, Map 1.1 and Map 5.1: purpose, likelihood, magnitudeRisk tiersClause 6.1.2 risk assessment, 6.1.4 impact assessment; Annex A.5
Measure 2: tests and evaluations; Manage 1–2: treat and respondControls per tierClause 6.1.3 risk treatment; Annex A.6 to A.10
Measure 2.4, Manage 2.4 and 4.1: monitor, disengage, respondMonitoring and reviewClause 9 evaluation, internal audit, review; clause 10 improvement
Govern 1.7: decommissioning safelyRetirementClause 8 operation; Annex A.6, the AI system life cycle
- Name the owner
An accountable executive and a committee with a written charter.
- Write the policy
Scope, risk appetite, tiers and who approves what.
- Build the inventory
Survey teams and vendors; include AI features inside bought software.
- Tier everything
Legal class first, then impact and autonomy.
- Close Tier 1 gaps
Tests, oversight, records and change control for the riskiest systems first.
- Start the loop
Monitoring reports and a review calendar; certify later if you need to.
On Botminds
How Botminds turns an AI governance framework into running controls.
A framework lives in documents; its controls have to run where the AI runs. On Botminds, the controls for parts 1, 4 and 5 are part of the platform the agents run on, and each leaves evidence for the inventory entry and the review.
Botminds platformthe controls a framework names, enforced where agents run
Roles at three layersplatform, workspace and project roles; Access Explorer shows who reaches what
Tools granted by nameeach agent reaches only the tools it is given
Approvals on toolschosen tools wait for a person to approve, reject or edit
Evaluation gatesevery agent publish scored on golden data, verdict kept
Drift and calibrationscore drift watched; confidence fitted from reviewers’ decisions
Change controlsnapshots, diff, rollback and freeze with a reason
Separation of dutiessignatures and stage moves need a different person
Verifiable audit trailhash-chained, append-only, verified on demand
Read how each control works in AI governance on Botminds. The Academy covers access and roles, human in the loop and policy and guard rails.
Questions people ask
AI governance framework: common questions.
The technology behind itAI governanceAudit, signatures, access and proof→What is an AI governance framework?
An AI governance framework is the roles, policies, processes and controls an organization uses to decide which AI systems to build or buy, how risky each is, what must be shown before release, and how each is monitored, changed and retired.
What is an example of an AI governance framework?
The NIST AI Risk Management Framework (Govern, Map, Measure, Manage) and ISO/IEC 42001, the AI management system standard, are the most used references. Singapore’s Model AI Governance Framework is another published example. Most organizations build their own on top of one of them.
What is the difference between NIST AI RMF and ISO/IEC 42001?
NIST AI RMF is a voluntary framework describing outcomes for managing AI risk. ISO/IEC 42001 sets requirements for an AI management system and can be certified by an accredited auditor. Many organizations use NIST for the risk method and ISO/IEC 42001 for the management system.
Can you be certified against an AI governance framework?
Against ISO/IEC 42001, yes — through an accredited certification body. The NIST AI RMF has no certification. EU AI Act conformity assessment is a separate legal process for high-risk systems.
How do you tier AI systems by risk?
Check the legal class first — under the EU AI Act some uses are prohibited and Annex III uses are high-risk. Then tier on two axes: the harm a wrong output can do, and how freely the system acts without a person. Write down the reasoning.
What should an AI inventory include?
Name, accountable owner, intended purpose, model and version, tools and actions, data sources, people affected, risk tier with its reasoning, and status. NIST’s Govern 1.6 asks for mechanisms to inventory AI systems.
Who is responsible for AI governance?
An accountable executive owns AI risk; a committee approves high-tier systems; system owners run the controls; an independent reviewer challenges; internal audit checks the framework works.
How does Botminds support an AI governance framework?
By running the controls a framework names: roles and tool grants, approvals on chosen tools, evaluation gates on every agent publish, drift monitoring, versioned change control, separation of duties and a verifiable audit trail.
Sources
- NIST AI 100-1: Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 2023nvlpubs.nist.gov
- NIST: AI Risk Management Framework — status, Playbook and profilesnist.gov
- NIST AI 600-1: Generative Artificial Intelligence Profile, July 2024nvlpubs.nist.gov
- ISO/IEC 42001:2023 Information technology — Artificial intelligence — Management systemiso.org
- PDPC Singapore: Model AI Governance Framework and Singapore’s approach to AI governancepdpc.gov.sg
- Regulation (EU) 2024/1689 (EU AI Act), EUR-Lexeur-lex.europa.eu
- European Commission: AI Act — regulatory framework and application datesdigital-strategy.ec.europa.eu
Read next
AI governance vs data governanceWhat each one controls, who owns it, and where the two meet→EU AI Act for AI agentsRisk classes, provider and deployer duties, and the dates that apply→SR 11-7 and generative AIModel risk guidance, its 2026 replacement, and how banks govern LLMs and agents→All guidesEvery explainer, by topic→
See it on
your own work.
Bring one process and its documents. We’ll show you the agents, the controls and the record, running on your data.