AI Governance Framework: The 8 Parts and a Template
An AI governance framework sets who owns, approves and monitors each AI system. Get the 8 parts and a free template, mapped to the EU AI Act.

On this page
In short: an AI governance framework is the written set of roles, rules and records that decides which AI systems you use and who answers for each one. It also covers who approves a system before go-live and what evidence you keep while it runs. If your board asked tomorrow which AI systems you run and who signed each one off, a framework with a register and named owners lets you answer from records instead of from memory.

TL;DR
- An AI governance framework has 8 parts: inventory, classification, roles, policy, approval gates, data controls, monitoring and incidents, and AI literacy.
- Each part should also produce a record someone outside your AI team could check: a register, a signed approval, a log, a training record.
- The EU AI Act, ISO/IEC 42001 and the NIST AI RMF overlap on most of these parts, but they differ in force: the Act is law, ISO/IEC 42001 is a certifiable standard and the NIST AI RMF is voluntary guidance. So build one framework and map it to the ones that apply to you.
- Above all, start with the inventory. You can’t govern a system you don’t know you have.
What is an AI governance framework?
An AI governance framework answers four questions for every AI system in your organisation:
- What is it, and what does it decide?
- How risky is it, and under which rules?
- Who is accountable, and who approved it?
- How do we know it still works as intended?
However, a policy document alone doesn’t answer those questions. Instead, a framework connects the policy to named people, specific systems and records you can produce on request. It is also different from the frameworks you may have heard of. In particular, the NIST AI Risk Management Framework and ISO/IEC 42001 are reference frameworks: they describe what good looks like. By contrast, your AI governance framework is your organisation’s own version, written for your systems and your regulators. If you’re choosing between those two, see NIST AI RMF vs ISO 42001: which AI governance framework should you adopt first?
The 8 parts of an AI governance framework
1. An AI inventory
A register of every AI system you build, buy or use: what it does, who owns it, which provider and model it uses, and what data it touches. Also include AI features switched on inside software you already pay for, and tools staff use without approval. The second group has a name: shadow AI. Evidence: the register itself, with a named owner on every row.
2. Risk classification
First, sort each system by the rules that apply to it. Under the EU AI Act, systems fall broadly into prohibited practices (Article 5), high-risk systems (Article 6 and Annex III, which covers uses such as recruitment, credit scoring and access to essential public services), systems with transparency duties (Article 50, such as chatbots and generated content) and minimal risk. However, these categories overlap. For example, a high-risk system can also have transparency duties. In addition, general-purpose AI models have their own rules. Evidence: the classification and the reason for it, recorded against each system in the inventory.
3. Accountable roles
Each system in your AI governance framework has one accountable person, not a team. Then decide who approves new systems and material changes, who handles incidents, and how the board or trustees get oversight. Also, for high-risk systems, Article 26 of the EU AI Act requires deployers to assign human oversight to people with the competence, training and authority to carry it out. Evidence: named owners and approvers, recorded against each system.
4. Policy and acceptable use
A short policy that says which uses are allowed, which need approval and which are banned. Above all, keep it short enough to read in ten minutes, so staff can check it before they use a tool. Evidence: the approved policy, its version and date, and who signed it off.
5. Approval gates
In short, approval gates are where an AI governance framework stops being paperwork. Specifically, a gate is a point before go-live, and before any material change, where a named person reviews the system and signs it off. In other words, the gate is where the classification, the data checks and the oversight plan come together. Our rule, therefore: AI agents can’t sign a gate. Only a person can. Evidence: a signed approval for each release, with what was reviewed and when.
6. Data controls
What data each system may use, where that data goes, and whether personal data is involved. If it is, your GDPR duties apply as well, including a data protection impact assessment where processing is likely to be high-risk. This part also overlaps with your data governance programme. AI governance vs data governance explains where the line sits. Evidence: data sources, data protection impact assessments and the sensitive data each system touches.
7. Monitoring, logs and incidents
Next, your AI governance framework should show how you check that each system still does what it was approved to do, and what happens when it doesn’t. For high-risk systems, Article 26(6) of the EU AI Act requires deployers to keep the logs the system automatically generates, to the extent they are under their control, for at least six months. However, other laws may require longer. Article 26(5) also requires deployers of high-risk systems to inform the provider, and then the relevant market surveillance authority, of serious incidents. Evidence: logs, monitoring results and an incident register.
8. AI literacy
Finally, Article 4 of the EU AI Act already requires providers and deployers to take measures to ensure a sufficient level of AI literacy among the staff who operate and use AI systems. In practice that means people know what the system does, where it fails and when to escalate. Evidence: who was trained, on what and when.
AI governance framework template
First, copy this into a spreadsheet, a document or your governance tool. Then fill in one row per AI system.

| Part | What to write down | Who owns it | Evidence to keep |
|---|---|---|---|
| 1. Inventory | System name, purpose, provider and model, data used, users | AI owner | AI register |
| 2. Classification | EU AI Act category and reason; other laws that apply | Compliance or risk lead | Classification record |
| 3. Roles | Accountable owner, approver, oversight people, incident contact | Senior responsible owner | Named roles per system |
| 4. Policy | Allowed, approval-needed and banned uses | Policy owner | Signed, dated policy |
| 5. Approval gates | What is reviewed before go-live and before changes | Approver | Signed approvals |
| 6. Data | Data sources, personal data, where data is processed | Data protection officer | DPIAs, data register |
| 7. Monitoring | What is measured, how often, what triggers a review | AI owner | Logs (6 months minimum for high-risk), incident register |
| 8. Literacy | Who needs training, on what | HR or learning lead | Training records |
How an AI governance framework maps to the EU AI Act, ISO 42001 and NIST AI RMF
| Part | EU AI Act (Regulation (EU) 2024/1689) | ISO/IEC 42001:2023 | NIST AI RMF 1.0 |
|---|---|---|---|
| 1. Inventory | Supports classification under Article 6 | Clause 4 (context and scope) | Govern (GOVERN 1.6), Map |
| 2. Classification | Articles 5, 6, 50 and Annex III | Clause 6 (planning, risk and impact assessment) | Map |
| 3. Roles | Article 26 (human oversight by deployers) | Clause 5 (leadership, roles) | Govern |
| 4. Policy | Supports Articles 4, 5 and 26 | Clause 5 (AI policy) | Govern |
| 5. Approval gates | Article 26 (deployer oversight); Article 14 sets what providers must build in to allow oversight | Clause 8 (operation) | Manage |
| 6. Data | Article 10 (data governance, high-risk providers); GDPR | Clause 8 and Annex A controls | Map, Measure |
| 7. Monitoring | Article 26 (logs, monitoring, incidents) | Clause 9 (performance evaluation) | Measure, Manage |
| 8. Literacy | Article 4 (AI literacy) | Clause 7 (support, competence) | Govern |
Mappings are indicative. The three documents do not match clause for clause, and obligations under the Act differ for providers and deployers. Even so, one framework lets you collect each piece of evidence once and point to it from all three documents. In other words, the mapping is a guide to overlap, not a legal opinion that meeting one satisfies another.
A worked example: a council’s housing triage tool
This is a hypothetical example, not a customer story. For example, suppose a local council wants to use an AI system to prioritise housing applications. Here is how, step by step, its AI governance framework could apply the eight parts.
- Inventory: first, the system goes on the register with its supplier, model, the application data it reads and the housing team as users.
- Classification: Annex III lists AI used by public authorities to evaluate eligibility for essential public benefits and services, so a tool like this is likely to be high-risk. Next, the council confirms the classification against the final text, including the Article 6(3) exception. If it is high-risk, the council, as a public body, must also carry out a fundamental rights impact assessment under Article 27 before first use.
- Roles: above all, the head of housing is accountable. In addition, two named caseworkers provide human oversight.
- Policy: for instance, the policy says the system may rank applications but may not refuse one.
- Approval gate: the head of housing and the data protection officer sign off before go-live, and again before any change to the scoring.
- Data: meanwhile, a DPIA covers the personal data in applications, which is likely to include special-category data.
- Monitoring: after go-live, decisions are logged and kept for at least six months, and overturned rankings are reviewed every month.
- Literacy: finally, caseworkers are trained on what the ranking means and when to override it.
Together, those records are what the council would show an auditor, an ombudsman or a resident who asks why their application was ranked where it was.
When do you need an AI governance framework?
Now. The EU AI Act’s bans and its AI literacy duty already apply. Meanwhile, under the Digital Omnibus on AI, the high-risk rules apply from 2 December 2027, and from 2 August 2028 for AI built into regulated products. Before that, two new bans and the end of a labelling grace period arrive on 2 December 2026. What still applies after the EU AI Act delay sets out the dates. In any case, check the final text of the Act before you rely on any article number or date. Still, the deadline matters less than the size of your AI estate. In fact, every month without a framework is another month of systems going live without an owner, a classification or an approval on record.
How @Human supports an AI governance framework
@Human by Planet Crust is an open-source, self-hosted platform for building AI agents, automations and applications, governed by design. As a result, every system you build is classified against the EU AI Act and ISO/IEC 42001. You also get accountable roles, signed approval gates, runtime guardrails, monitoring and incident tracking, risk scenarios and a sensitive-data register. For example, AI agents can’t sign a gate: 0 of the 296 tools on a live @Human instance can. To be clear, @Human gives you the controls and the evidence. Whether you are compliant still depends on how you use them.
AI governance framework FAQ
What should an AI governance framework include?
At minimum: an AI inventory, risk classification, accountable roles, a usage policy, approval gates, data controls, monitoring and incident handling, and AI literacy training.
Is an AI governance framework a legal requirement?
Partly. The EU AI Act doesn’t use the term, but it requires many of the parts: AI literacy (Article 4), human oversight and log keeping for high-risk systems (Article 26), and risk classification in practice. ISO/IEC 42001 requires a documented AI management system if you seek certification.
Is ISO 42001 an AI governance framework?
It’s a standard for an AI management system, which is a formal version of one. Therefore, you can be certified against it. Similarly, your own framework can be built to meet it.
How long does it take to set up?
First, the inventory and policy are documents, so you can draft them quickly. However, classification, approval gates and monitoring take longer because they have to run on every system, not just exist on paper.
Who should own the AI governance framework?
One senior person, named, with a reporting line to the board or trustees. Each system then has its own accountable owner.
Do we need a separate framework for AI agents?
Partly. Agents need extra controls, because they act rather than answer. See our agentic AI governance framework.
What to do today
First, open a spreadsheet and list every AI system your organisation uses, including AI features inside tools you already pay for. Then put one named owner next to each. Any row without an owner is your first governance gap. To see accountable roles, signed approval gates and incident tracking working together, book a free walkthrough. A real person replies within one working day.
Sources
- Regulation (EU) 2024/1689 (the EU AI Act), Articles 4, 5, 6, 10, 14, 26, 27, 50 and Annex III, Official Journal of the EU, 12 July 2024, https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- ISO/IEC 42001:2023, Information technology — Artificial intelligence — Management system, December 2023, https://www.iso.org/standard/81230.html
- NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, January 2023, https://doi.org/10.6028/NIST.AI.100-1
- Planet Crust, The EU AI Act delay is real. Here’s what still applies on 2 December 2026, 10 November 2026, https://www.planetcrust.com/eu-ai-act-delay-what-still-applies/