Governing AI use means knowing what is in use, who is responsible, how risky each use is, and what controls keep it safe. It does not mean slowing everything down. A small, clear structure and a register protect the organization and let low-risk uses move quickly.
This guide covers the main references (NIST AI RMF, ISO/IEC 42001, ISO/IEC 23894, the EU AI Act, and the OECD principles), a practical governance structure, risk tiers, FMEA applied to AI failure modes, and how to write an acceptable-use policy. A worked example scores eight AI risks at an illustrative plant before and after controls.
Laws and standards change. The EU AI Act dates here are as understood in late 2026; check the current text and ask qualified people what applies to you.
Before You Start
Why Govern AI Use
Governing AI use does not mean slowing it down. It means knowing what is in use, who is responsible, how risky each use is, and what controls keep it safe. Without that, people use AI anyway, in ways nobody has reviewed, with data nobody has classified.
Know What Is in Use
Staff use AI tools whether or not the company has approved them. A register brings that into view.
Match Controls to Risk
A tool that drafts internal notes needs less control than one that screens job applicants.
Protect Data and People
Confidential data, personal data, and decisions that affect people need extra care.
Meet Expectations
Customers, auditors, insurers, and in some places the law expect evidence of control.
The Reference Frameworks
You do not need to invent a governance approach. Several widely used references are available, and they fit together.
| Reference | What it is | How to use it |
|---|---|---|
| NIST AI Risk Management Framework (AI RMF 1.0, January 2023) | A voluntary US framework with four functions: Govern, Map, Measure, and Manage. A companion Generative AI Profile (NIST AI 600-1, July 2024) covers risks specific to generative AI | A practical way to structure risk work, for any organization |
| ISO/IEC 42001:2023 | A certifiable management system standard for AI, using the same plan-do-check-act structure as ISO 9001 and ISO 14001 | A structure for policy, roles, risk, controls, and audit; useful if you already run ISO systems |
| ISO/IEC 23894:2023 | Guidance on managing risk related to AI | Detail on risk assessment for AI |
| EU AI Act (Regulation (EU) 2024/1689) | A law that sorts AI systems by risk: prohibited practices, high-risk systems, systems with transparency duties, and minimal-risk systems | If you operate in or sell into the EU; see the note on dates below |
| OECD AI Principles | High-level principles for trustworthy AI, adopted by many governments | A statement of values for your AI policy |
In the US there is no single federal law for AI in operations at the time of writing, but existing rules on employment, credit, health, privacy, and consumer protection apply to AI-supported decisions, and some states regulate automated decision tools. Check the requirements that apply to you.
A Practical Governance Structure
| Element | What it is | A good test |
|---|---|---|
| AI policy | A short statement of what AI may and may not be used for, and the principles that apply | Can an employee tell in a minute whether a use is allowed? |
| Acceptable-use rules and data classes | Which kinds of data may go into which kinds of tool | Does everyone know what they must never paste into a public tool? |
| Roles | An executive sponsor, a governance group, an owner for every use, and a reviewer for every output | Is there a named owner for each AI use in the register? |
| AI register | A list of AI uses and tools, with purpose, owner, data, risk tier, and review date | Is there a use nobody can name? |
| Risk tiers | A simple sorting by impact, such as low, medium, high, with controls for each | Do high-impact uses get more review? |
| Testing and monitoring | Checks before use and while in use, with measures and limits | Would we notice if it got worse? |
| Vendor and data controls | Checks on suppliers and how data is stored and used | Do we know what the vendor does with our data? |
| Training | AI literacy for everyone, deeper training for owners and reviewers | Do users know the limits of the tool? |
| Incident response and change control | What to do when AI misbehaves, and how changes to models and prompts are approved | Is there a way to switch it off, and a record of changes? |
A governance group can be small and can build on an existing safety or quality committee. See Safety Team and Committee for charter and meeting practice, and the RACI Matrix Builder for roles.
Sorting Uses by Risk
Ask how much harm a wrong output could do, and how easily it would be caught. The tiers below are an example; align them to your own risk matrix and to any law that applies.
| Tier | Typical uses | Controls |
|---|---|---|
| Low | Drafting internal notes, summarizing public information, brainstorming | Acceptable-use rules, user training, no confidential data in public tools |
| Medium | Customer-facing drafts, internal classification and routing, reports used in decisions | Named owner, human review, testing on real examples, monitoring, vendor check |
| High | Decisions about people (hiring, pay, discipline), safety-related decisions, regulated or legally significant outputs, autonomous actions with real consequences | Formal risk assessment, legal and compliance review, strong human oversight, validation before use, logging, audit, and approval by the governance group |
| Not allowed | Uses the policy or law prohibits | Do not proceed |
Risk Assessment With FMEA
Failure mode and effects analysis (FMEA) suits AI well, because it asks the same questions: what can go wrong, how bad would it be, how likely is it, and would we catch it? For each AI use, list the failure modes, rate severity (S), occurrence (O), and detection (D) from 1 to 10, and multiply to get the risk priority number (RPN). Then add controls and re-rate.
| Common AI failure modes | Examples |
|---|---|
| Wrong output stated confidently | An invented part number, a wrong summary of a standard |
| Misclassification | A safety complaint filed as routine |
| Data leakage | Confidential data entered into a public tool |
| Manipulated input | Instructions hidden in a document or email that change the AI’s behavior |
| Drift or silent change | A vendor updates a model and the behavior changes |
| Over-reliance | Reviewers approve without checking |
| Bias and unfairness | Different error rates for different groups of people or items |
| Runaway action or cost | An agent loops, or takes actions beyond its role |
RPN has known limits, and the AIAG and VDA FMEA Handbook (2019) prefers Action Priority. Either way, treat severity first: a high-severity failure needs action whatever its RPN. See the FMEA guide.
Worked Example: An AI Risk Assessment at Riverside Plant
Riverside Plant, an illustrative 140-employee metal fabrication and assembly plant, assessed the risks of its complaint classification and work-order agent pilots. The team scored eight failure modes before and after planned controls. All scores are illustrative.
| # | Failure mode | Before (S × O × D = RPN) | Main control | After |
|---|---|---|---|---|
| 1 | Safety complaint classified as routine | 9 × 4 × 6 = 216 (High) | Keyword safety rule, plus human review of low-confidence items | 9 × 2 × 3 = 54 (Low) |
| 2 | Drafted reply cites a part number that does not exist | 6 × 5 × 5 = 150 (Medium) | Look-up check of every part number against the catalog before sending | 6 × 3 × 2 = 36 (Low) |
| 3 | Confidential customer data pasted into a public AI tool | 8 × 5 × 7 = 280 (High) | Approved tools only, data classification rules, and training | 8 × 2 × 3 = 48 (Low) |
| 4 | Agent closes a work order that is not finished | 7 × 3 × 7 = 147 (Medium) | Closure needs a technician confirmation step and a photo | 7 × 2 × 3 = 42 (Low) |
| 5 | Instructions hidden in a customer email change agent behavior | 8 × 3 × 8 = 192 (Medium) | Treat email text as data, restrict agent tools, and filter outputs | 8 × 2 × 4 = 64 (Low) |
| 6 | Vendor model update changes behavior without notice | 6 × 4 × 8 = 192 (Medium) | Pin the version; re-run the test set before changing | 6 × 3 × 3 = 54 (Low) |
| 7 | Reviewers approve outputs without checking | 7 × 6 × 7 = 294 (High) | Seed known errors into the queue; rotate reviewers; track the catch rate | 7 × 4 × 4 = 112 (Medium) |
| 8 | Agent loops and runs up usage cost | 4 × 3 × 4 = 48 (Low) | Step and spend limits per task, with an alert | 4 × 2 × 2 = 16 (Low) |
Results. Before the controls 3 risks were High and 4 Medium. After the controls 1 is Medium and 7 are Low. The one that stays Medium is the human one: reviewers who approve without checking (RPN 112). The team did not try to fix it with a reminder. It seeded a few known errors into the review queue, rotated reviewers, and made the catch rate a standing measure.
Record your own risks in the AI Risk Register, which has the scoring scales, a failure mode library, and a dashboard.
Writing an Acceptable-Use Policy
- Say what is allowed and what is not, in plain language, with examples.
- Classify data (for example public, internal, confidential, restricted) and say which classes may go into which kinds of tool: public tools, approved enterprise tools, or private deployments.
- Require review of AI output before it is sent outside the organization or used in a decision.
- Require disclosure where people would reasonably expect to know that AI was used.
- Name who to ask and how to report a problem, without blame.
- Review it every six to twelve months, because tools and laws change.
The AI Acceptable-Use Policy and Tool Register includes example policy text, a data classification matrix, a register of approved tools, and an acknowledgement log.
Common Mistakes
Policy Without Practice
A document nobody reads, and no register of what is in use.
Banning Instead of Governing
A ban drives use underground, where nobody can see it.
Treating All Uses the Same
Heavy process for harmless uses, and light process for risky ones.
No Owner for Each Use
Everyone can use it and nobody is responsible.
One-Time Review
Approving a tool once and never checking again.
Ignoring the Vendor
Not asking what happens to your data or how changes are announced.
Self-Assessment Questions
- Do we have a register of AI uses, with an owner, purpose, data, and risk tier for each?
- Do people know what data they must not put into which tools?
- Are higher-risk uses reviewed by more than the person who wants them?
- Can we switch a use off, and do we keep a record of changes?
- Do we know which laws and standards apply to our uses, and when we last checked?
AI Governance and Risk for Operations: Frequently Asked Questions
What is the NIST AI Risk Management Framework?
A voluntary framework from the US National Institute of Standards and Technology, released in January 2023 as AI RMF 1.0. It organizes AI risk work into four functions, Govern, Map, Measure, and Manage, which repeat over the life of an AI use. A Generative AI Profile (NIST AI 600-1, July 2024) adds guidance for generative AI.
What is ISO/IEC 42001?
A management system standard for AI, published in 2023, that can be audited and certified. It follows the same plan-do-check-act structure as ISO 9001 and ISO 14001, covering policy, roles, risk assessment, controls, and improvement for AI uses. It is most useful if you already run ISO management systems.
Does the EU AI Act apply to us?
It can if you place AI systems on the EU market, deploy them in the EU, or use their outputs in the EU, even if you are based elsewhere. Obligations depend on the risk category of the use. Dates for some obligations were moved in 2026, so check the current text and have qualified people assess your situation.
How do we assess the risk of an AI use?
List what can go wrong, rate how severe it would be, how likely it is, and how easily it would be detected, then choose controls and re-rate. FMEA does this well. Give priority to severity, and match the depth of review to the risk tier of the use.
Do we need an AI policy if we only use off-the-shelf tools?
Yes. The main risks with off-the-shelf tools are what data goes in, what comes out and is trusted, and who is responsible. A short acceptable-use policy with data classes, review rules, and a register of approved tools deals with all three.
Sources and Further Reading
- National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (26 January 2023), and AI RMF: Generative Artificial Intelligence Profile, NIST AI 600-1 (26 July 2024).
- ISO/IEC 42001:2023, Artificial intelligence: management system; ISO/IEC 23894:2023, Artificial intelligence: guidance on risk management; ISO/IEC 22989:2022, concepts and terminology (check current editions).
- Regulation (EU) 2024/1689 (the Artificial Intelligence Act), and the 2026 amending regulation (“digital omnibus”); check the Official Journal for the current text.
- OECD, Recommendation of the Council on Artificial Intelligence (2019, updated 2024).
- Automotive Industry Action Group and VDA, FMEA Handbook (2019), on severity, occurrence, detection, and Action Priority.
- ISO 31000:2018, Risk management: guidelines.