An AI governance framework is the set of owners, rules and controls that decides which AI tools a business uses, what data may go into them, who checks the output and what happens when something goes wrong. For a small business it fits on a few pages and covers eight components: ownership, an approved-tool register, data classes, vendor review, identities and access, human review, logging and incident response. It is reviewed on a schedule, not written once and filed.
A framework is the program around your AI policy. The rules staff read day to day belong in the policy itself, which our AI usage policy guide covers, and the software that automates parts of the program is compared in our guide to AI governance tools. This article gives you the structure, a starter template, the mapping to NIST and ISO, and the checks that show it works. Our AI governance consulting builds and runs it with you.
What is an AI governance framework, and how does it differ from a policy or a model?
Three terms get used interchangeably, and they are not the same thing:
| Term | What it is | Who reads it |
|---|---|---|
| AI governance policy | The rules for staff: approved tools, what data may go where, when a person must check AI output | Everyone who uses AI at work |
| AI governance framework | The structure behind the policy: owners, the tool register, vendor review, controls, logs, incident handling and the review cycle | Leadership, IT and whoever approves tools |
| AI governance model | How decisions get made: one accountable owner or a small committee, and which decisions need whose approval | Leadership |
For a small firm, the simplest workable model is one accountable executive, with IT or a managed provider as administrator and a short monthly check-in instead of a committee. Your AI governance strategy is then the order in which you put the eight components in place, starting with whatever touches client data first.
What are the components of an AI governance framework?
Each of the eight components needs a decision, an owner and evidence that it is happening.
| Component | What you decide | Starter rule | Evidence |
|---|---|---|---|
| 1. Ownership and scope | Who is accountable, and which teams, tools and data are in scope | One named executive owns AI governance; IT administers it | A one-page charter with names and the next review date |
| 2. Approved-tool register | Which AI tools and built-in AI features staff may use, on which accounts, for what | Only registered tools, on business accounts with company sign-in | The register, with an owner and review date per tool |
| 3. Data classes | Which kinds of data may go into which tools | Public data in any approved tool; internal data in approved business tools; client, health, financial and credential data only in tools approved for that class | A data rules table staff can check in seconds |
| 4. Vendor review | What a new AI vendor must show before approval | No approval until retention, training on your data, subprocessors and data location are answered in writing | A completed review form per vendor |
| 5. Identities and access | How people and AI agents sign in, and what they can reach | Company sign-in with MFA; admin approval before any AI app connects to mail or files; agents get their own identities, never a staff login | Sign-in and app consent logs |
| 6. Human review | Which AI output a person must check, and who | A person reviews anything sent to a client, published, or used in a decision about a person | Review notes or approvals in the workflow |
| 7. Logging and monitoring | What is recorded, and who looks at it | Sign-ins, app consents and agent actions logged and reviewed monthly | The monthly review record |
| 8. Incident response | What counts as an AI incident and what happens next | Report client data pasted into an unapproved tool, a wrong answer sent to a client, or an agent action nobody approved, within one business day | An incident log with outcomes |
What does a starter AI governance framework template look like?
Copy this outline into a document. Expect three to five pages for a small firm, with the register and data rules kept in a spreadsheet the document points to.
- Purpose and scope. Why the framework exists, the teams it covers and the date of the next review.
- Roles. The executive owner, the administrator, the people who approve tools, and what every employee is expected to do.
- Approved-tool register. Columns for the tool, account type, business owner, approved uses, highest data class allowed, date approved and next review.
- Data rules. Each data class with examples from your own business and the tools allowed for it.
- New-tool requests. How staff ask, the vendor review questions, and a promised turnaround so people do not work around the process.
- Access rules. Company sign-in, MFA, who may connect AI apps to company data, and how AI agents get identities and permissions.
- Human review. The outputs that need a second pair of eyes, and whose.
- Logging and monitoring. What is logged, where, for how long and who reviews it.
- Incident reporting. What to report, to whom, and the first steps: stop, contain, notify, fix and record.
- Training. A short session per role at onboarding and once a year.
- Review and change log. A quarterly review of the register, logs and incidents, with changes recorded.
The outline doubles as an AI governance checklist: any section that is empty or has no named owner is your next task.
How does the framework map to the NIST AI RMF and ISO/IEC 42001?
You do not need a formal standard to govern AI well, but mapping to one helps when a client or insurer asks which framework you follow.
NIST AI Risk Management Framework (AI RMF 1.0). NIST released it on January 26, 2023, for voluntary use. Its core has four functions. Govern is a cross-cutting function that runs through the other three. Map establishes the context to frame the risks of an AI system, Measure analyzes, assesses, benchmarks and monitors those risks, and Manage puts resources against them on a regular basis. Once governance is in place, NIST says the other functions can be performed in any order and the process should be iterative. In July 2024 it added a Generative AI Profile, NIST AI 600-1, for risks specific to generative AI. As of October 2026, NIST also says AI RMF 1.0 is being revised under the White House AI Action Plan, so check its page for a newer version before you map to it.
| NIST AI RMF function | Components above that deliver it |
|---|---|
| Govern | Ownership and scope, roles, training and the review cycle |
| Map | The approved-tool register, data classes and vendor review |
| Measure | Logging and monitoring, human review records and testing of AI agents |
| Manage | Access controls, incident response and retiring tools that fail review |
ISO/IEC 42001:2023. Published in December 2023, it specifies requirements for establishing, implementing, maintaining and continually improving an AI management system, and it applies to any organization regardless of size. Unlike the NIST framework, it can be certified, by an accredited certification body. A small firm can borrow its management-system structure (scope, roles, risk assessment and internal review) without pursuing certification.
Two other names come up in searches. Singapore's Model AI Governance Framework, first released in January 2019, gives private-sector organizations practical guidance on deploying AI responsibly. Microsoft publishes six responsible AI principles (fairness, reliability and safety, privacy and security, inclusiveness, transparency and accountability) and a Responsible AI Standard; they describe how Microsoft approaches AI, while your framework governs how your business uses it.
How do you govern AI agents?
Agents raise the stakes because they act: they send messages, update records and call other systems. Singapore's IMDA published a Model AI Governance Framework for Agentic AI in January 2026 and updated it in May 2026. It works across four dimensions: bound the risks up front by choosing suitable use cases and limiting what agents can do, keep humans accountable with checkpoints where a person must approve, apply technical controls throughout the agent's life, and make end users responsible through transparency and training.
In practice, an agentic AI governance framework adds these controls to the eight components:
- Each agent runs under its own identity, never a staff member's login, with permissions limited to the data and actions its job needs.
- Read access by default, and write access only for named actions.
- Anything touching money, client records, outbound messages or system settings waits for a person's approval.
- Every action is logged, spend and rate limits stop runaway loops, and a named person can switch the agent off.
- Before launch, the agent is tested against instructions hidden in emails, documents and web pages.
Who maintains the framework, and how do you validate it?
The executive owner is accountable and chairs a short quarterly review. The administrator, usually IT or your managed provider, keeps the register, configures the controls and pulls the logs. Business owners approve the tools their teams use and confirm human review is happening. Every employee reports incidents.
Validate the framework against evidence, not intentions:
- Register against reality. Compare the register with sign-in and app consent logs. Any AI tool in the logs that is missing from the register is shadow AI to resolve.
- Human review sampling. Pull a sample of client-facing work that used AI and confirm a reviewer signed it off.
- Vendor terms. Re-read the terms of your main AI vendors at each review, because retention and training terms change.
- Incident drill. Once a year, walk through a scenario such as a client file pasted into a personal chatbot account, and time each step.
- Training records. Confirm every current employee completed the session for their role.
How does NetSys help with AI governance?
We start with what is already happening. Sign-in logs, app consents, managed browsers and a short staff survey show which AI tools are in use and on which accounts. The policy is drafted with your managers and reviewed by your counsel, and each approved tool gets an owner, an account type and a data rule. Controls in Microsoft 365 or Google Workspace start in audit-only mode, so you see what they would block before they block it, and the register, logs and policy are then reviewed on a schedule.
We map the program to the NIST AI RMF and can give it an ISO/IEC 42001-style structure. We do not perform ISO/IEC 42001 certification audits or give legal sign-off; those stay with an accredited certification body and your counsel. Ongoing reviews run month to month.
Book a call with an engineer to scope your framework, and bring a list of the AI tools your staff use and the kinds of client data you hold.
Frequently asked questions
What is an AI governance framework?
It is the structure a business uses to control its use of AI: named owners, a register of approved tools, rules on what data may go into them, vendor review, access controls, human review of output, logging and incident response, all reviewed on a schedule. The staff-facing AI policy sits inside it.
What are AI governance best practices for a small business?
Name one accountable owner, approve a small set of business-grade tools on company sign-in, write data rules staff can apply in seconds, require a person to review anything that reaches a client, log AI app sign-ins and consents, and review the whole program every quarter. Make the approved path easier than the unapproved one.
Is there a free AI governance framework PDF or template?
NIST publishes the AI RMF 1.0, its Playbook and the Generative AI Profile free of charge on its website. ISO/IEC 42001 is a paid standard. The eleven-section outline in this article is a working template you can copy into your own document.
Does a small business need ISO/IEC 42001 certification?
Usually not. Certification makes sense when clients or contracts require it. Otherwise a framework mapped to the NIST AI RMF, using ISO/IEC 42001 as a structural guide, gives you the same discipline without an audit. When certification is needed, it comes only from an accredited certification body.
How often should an AI governance framework be reviewed?
Quarterly for the register, logs and incidents, and at least once a year for the full document and the training. Review sooner when you adopt an AI tool that touches client data, deploy an agent, or a vendor changes its terms.
Who should own AI governance in a small company?
An executive with authority over both operations and risk, often the owner, COO or managing partner, with IT or a managed provider as administrator. Ownership by IT alone tends to stall, because the hard decisions are about which business uses are worth their risk.
Sources and further reading
- NIST: AI Risk Management Framework: AI RMF 1.0, the Playbook, NIST AI 600-1 and the note that AI RMF 1.0 is being revised, checked October 2026.
- NIST AI Resource Center: the AI RMF Core: the Govern, Map, Measure and Manage functions.
- IEC webstore: ISO/IEC 42001:2023: the scope and publication date of the AI management system standard.
- IMDA Singapore: AI governance: the Model AI Governance Framework and the framework for agentic AI, checked October 2026.
- Microsoft: responsible AI principles and approach: the six principles and the Responsible AI Standard.
Discuss ai governance consulting for your business.
Tell us about your current systems, the result you need and your timeline. We will discuss the work, responsibilities and pricing before you decide on an engagement.