AI governance used to belong to data and security teams. Now employees need a simple way to check whether a new AI use case is OK, who decides, and what to pause while they wait.

Governance is well known to the average data engineer or your cybersecurity specialist. AI brings it a bit closer to the average employee. Except not everyone knows what it is and why it’s important when using modern AI tools.
AI use can unintentionally break laws or internal policy. Governance is a way to keep AI use within accepted use cases. Employees can understand why a use case is acceptable, the data they can use, the checks in place, and the owner for any process or data.
Plus, it shows employees whom to ask when a new use case isn’t documented. Before approving it, the reviewer needs to check who could be affected if the AI gets something wrong and whether someone can catch that mistake before the output is used. Employees also need to know what they must pause while waiting for a decision.
This guide adapts the AI Risk Management Framework from NIST, the National Institute of Standards and Technology.
TL;DR
- AI governance best practices connect use-case approval, employee AI literacy, and clear ownership.
- Review the task and its consequences, not just the tool employees want to use.
- Article 4 addresses AI literacy; Articles 5 and 50 introduce separate requirements where applicable.
- Reopen approval when data, purpose, affected people, or system permissions change.
Four possible review outcomes
NIST’s Map function establishes the context and potential impacts of an AI use, while Manage is about deciding whether deployment should proceed and how to respond to risks.
Why are we focused on use cases, rather than tools? It’s based on our experience with the EU AI Act, and its overlap with standards like ISO/IEC 42001. General-purpose AI tools can be used broadly and it’s easy to break a law or violate policy without intending to do so. Starting from specific use cases is an easier way to govern AI use. And a tool approval does not answer whether a particular use is acceptable.
For employee requests, this guide adapts those functions into four practical review outcomes: proceed within defined boundaries, test under supervision, obtain specialist review, or stop.
Record the reason, the person responsible, and the conditions that trigger another review.
| Review outcome | Best for | Required next action | Limitation |
|---|---|---|---|
| Approve with boundaries | A defined task with understood data and checks | Record permitted use, restrictions, and owner | Approval does not cover a different task |
| Supervised pilot | A bounded task with unresolved performance questions | Set evaluation criteria and keep a named reviewer | A successful demonstration does not establish everyday reliability |
| Specialist review | Uses involving sensitive data, consequential decisions, or legal questions | Assign the relevant legal, privacy, security, or operational reviewer | Employees must wait for the necessary decision |
| Stop | Prohibited uses or uses your organization cannot adequately control | Explain the reason and identify an acceptable alternative where possible | Training cannot make an impermissible use acceptable |
Why use NIST as the foundation for this review?
The NIST AI Risk Management Framework provides the foundation for the review guidance below because it connects an AI use case’s purpose and impacts with ownership, testing, and ongoing oversight. Its four functions — Govern, Map, Measure, and Manage — cover the questions an employee use-case review needs to address without prescribing one approval process for every organization.
NIST is the National Institute of Standards and Technology, a U.S. government science agency within the Department of Commerce. Its mission includes advancing measurement science, standards, and technology. That makes its work relevant to a practical governance question: how can an organization assess AI risks and check whether its controls work? See NIST’s role and mission.
NIST is qualified to develop this guidance through its AI measurement and evaluation work and its ability to bring public- and private-sector expertise together. Congress directed NIST to develop a voluntary AI risk framework in collaboration with those sectors. The framework, released in January 2023, was developed through public requests for information, draft reviews, and workshops. See the AI Risk Management Framework background.
The connection to the employee review is specific:
- Govern: Assign accountable owners, define responsibilities, and prepare employees to carry them out.
- Map: Describe the task, data, intended use, affected people, and potential consequences.
- Measure: Test performance and relevant risks under conditions similar to actual use.
- Manage: Decide whether to proceed, document risk responses, and monitor changes or failures.
The five-step sequence below is an adaptation, not a NIST-prescribed or independently validated approval method. NIST explicitly states that its framework functions are not an ordered checklist. Adapt the review to the use case and involve the appropriate specialists; following this guidance does not establish EU AI Act compliance or replace legal assessment. NIST identifies AI RMF 1.0 as under revision as of when this article was published, so check the cited source when updating your process.
Five steps to review a new AI use case
Use the following sequence as a decision record. Keep the employee-facing request short, then route specialist questions to the people who can answer them. The employee proposing a use should describe the work; they should not have to interpret every applicable law.
Approve task
Ask the employee to describe the intended output and its recipient. Record whether AI drafts, recommends, classifies, decides, or takes an action, because those roles require different checks.
Example: A communications employee proposes drafting an internal event announcement from approved public information. That request is different from generating a statement about an employee investigation, even if both use the same writing tool.
Name the business owner who accepts responsibility for the workflow. Approval should identify the task, not simply say that a department can use AI.
Check data
List the information entering the system and the places the output will go. Include uploaded documents, connected repositories and information retrieved automatically, rather than looking only at what an employee types.
Ask the privacy and security owners to assess relevant handling conditions. Employees need a clear operational instruction: which information they can use, which they must remove, and which they must keep out entirely. A training reminder does not replace access restrictions or a supplier assessment.
Check boundaries
Identify who the use affects and what happens if the output is wrong. A draft that an informed employee checks before circulation differs from an automated action affecting a customer or job applicant.
Review the proposed use against company policy, applicable law, and the system’s permitted purpose. Assign unresolved questions to a named reviewer. Employees should know whether they can continue a bounded test or must pause until that reviewer responds.
Test results
Define what an acceptable result looks like before the pilot begins. Use representative tasks and include cases where the system should refuse, flag uncertainty, or require human intervention.
Name the person who checks factual accuracy, inappropriate disclosure, and suitability for the task. A human review requirement has little value if the reviewer lacks the knowledge, source material, or authority to challenge the answer. Record failures as well as successful outputs.
Record decision
Document the outcome, restrictions, owner, and reason for approval or rejection. Add the circumstances that require reassessment, such as a new data source, a different audience, or permission to take actions.
Give the employee instructions they can follow without interpreting a policy document. State what they can do, what they must check, and where they should take an exception. Keep that instruction connected to the approval record.
What changes the depth of an AI use-case review?
Use the same intake process across the company, then vary the specialist assessment according to the proposed work. A familiar tool can still support an unfamiliar or consequential use.
- Information sensitivity: Identify personal information, confidential business material, and other restricted content before approving inputs.
- Effect on people: Examine whether the output influences employment, access to services, or another consequential decision.
- System permissions: Distinguish generating a suggestion from sending messages, modifying records, or executing actions.
- Ability to check: Determine whether the assigned reviewer can detect errors before anyone relies on the result.
- Reach and reversibility: Consider who receives the output and whether your organization can correct an error after release.
These factors help reviewers choose the right route. They are not a substitute for the legal classification of a particular system or use.
Which EU AI Act requirements belong in the review?
For global companies, assess the EU AI Act’s territorial scope rather than assuming it applies only to organizations headquartered in the EU. Relevant connections include providers placing systems on the EU market, deployers established in the EU and certain cases involving outputs used in the EU.
Legal teams should determine which obligations apply to your organization’s role and use. Training should then translate those decisions into employee instructions.
Article 4: Teach people to use AI with appropriate understanding
Article 4, as amended in July 2026, requires providers and deployers to take measures to support the development of AI literacy among staff and others operating or using AI systems on their behalf. It takes account of technical knowledge, experience, education, training, context of use, and the people affected, and it does not require a guaranteed level of AI literacy for any individual.
For a use-case review, ask whether the employee understands the approved task, limitations, information rules, and escalation route. Article 4 does not prescribe a particular course, certificate, or universal passing score. Course completion alone does not establish that an employee can perform the approved task safely.
Article 5: Screen for prohibited practices
Article 5 prohibits specified AI practices, subject to the provision’s definitions and exceptions. A relevant workplace example is using AI to infer emotions in workplace settings, except where the system is intended for medical or safety reasons.
Route proposals involving employee emotion inference to legal review before a pilot. Do not treat employee agreement or a manager’s approval as an answer to whether a prohibition applies. Train employees to recognize the issue and raise it rather than asking them to resolve the legal exception themselves.
Article 50: Check applicable transparency duties
Article 50 has applied since Aug. 2, 2026. It assigns different transparency duties to providers and deployers, so the review must establish your organization’s role rather than treating disclosure as a single universal requirement.
For example, providers of systems intended to interact directly with people must ensure people are informed they are interacting with AI, unless that is obvious in the circumstances. Deployers generating or manipulating deep fake image, audio or video content face disclosure duties, with specific treatment for evidently artistic, creative, satirical, fictional, or analogous works. Have legal teams confirm the relevant requirement and exception before employees publish.
Employment uses need additional scrutiny
Certain systems used for recruitment, worker management, and employment decisions fall within the Act’s high-risk framework, subject to the applicable classification rules. Do not classify every HR writing task as high-risk, or assume that human involvement automatically removes that classification. The Digital Omnibus moved the application date for these high-risk obligations to Dec. 2, 2027.
A proposal to help draft an interview invitation differs from a proposal to rank applicants. Document the actual purpose and decision process so legal teams can assess the correct use. This guide supports planning; it is not legal advice.
How can you measure whether employees understand the rules?
Measure decisions employees can demonstrate, alongside participation. Completion tells you who finished training; a task-based assessment shows whether they can apply the approved boundaries.
Use hypothetical workplace scenarios that ask employees to:
- Identify information they must exclude from an AI prompt.
- Check a generated claim against an authoritative source.
- Recognize when an output affects a person and requires escalation.
- Explain who approves a tool and who owns the resulting work.
- Stop when a proposed action exceeds the approved use.
Record the assessment, its date, and the policy version it tests. Review mistakes by topic so HR can assign relevant learning and governance owners can clarify confusing rules. Treat recurring errors as a reason to improve the instruction, not just repeat the quiz.
Should you build training internally or buy it?
Choose the training route after defining the employee behaviors you need. Buying content and buying a system that delivers content solve different problems.
| Training route | Best for | Advantage | Limitation to address |
|---|---|---|---|
| Internal training | Organization-specific policies and workflows | Your policy owners can explain local rules directly | Someone must maintain the content and assessments |
| Large learning platforms | Evaluating available courses across learning needs | You can compare existing course content | Check whether the chosen course teaches your required workplace decisions |
| Specialist companywide AI training | Establishing shared foundational understanding | The buying brief can focus on employee AI literacy | Verify assessment evidence and add company-specific instructions |
| Learning management system | Delivering and administering learning | Provides a delivery route for selected content | Delivery records do not themselves establish employee capability |
Ask providers to demonstrate what an employee must do to pass, what evidence your team receives, and how company policy enters the learning experience. Keep technical security controls and legal decisions with their responsible owners.
FAQ
What are the most useful AI governance best practices for employee use cases?
Review the task, data, effects on people, human checks, and accountable owner before approval. Record restrictions and reopen the decision when the workflow changes.
Who should approve a new employee AI use case?
A named business owner should own the workflow, with legal, privacy, security, and other specialists deciding questions within their responsibilities. Give employees a clear route to that decision rather than asking them to collect informal permissions.
Does an approved AI tool cover every employee use?
An approved tool does not automatically authorize every use case. A different purpose, data source, audience, or action permission requires assessment against the approved boundaries.
Does Article 4 require a particular AI training certificate?
Article 4 does not prescribe a particular training certificate. It requires measures to support AI literacy, taking account of staff knowledge and experience, the context of use, and affected people.
Can AI training guarantee EU AI Act compliance?
AI training cannot guarantee EU AI Act compliance. It helps employees understand and carry out rules, while legal assessment, technical controls, and operational oversight remain separate responsibilities.
How do we help employees who are unsure about using AI?
Give unsure employees an approved task with clear input rules, a checking method, and a named person to ask for help. Assess the task they perform rather than expecting confidence to follow from course completion alone.
When should we review an approved AI use case again?
Review an approved AI use case when its purpose, data, affected people, system permissions, or required controls change. Also reassess when failures show that the current checks do not work.



