AI Acceptable Use Policies

Domain Context: AI Risk Governance and Framework Integration (ISACA Advanced in AI Risk – AAIR)

1. Why AI Acceptable Use Policies Matter

Generative AI tools such as ChatGPT, Copilot, Gemini and Claude have arrived in organizations faster than almost any technology in history. Employees are already pasting customer records into public chatbots, generating code with unknown licensing, and relying on hallucinated outputs for business decisions. Without clear rules, this creates shadow AI: unsanctioned, unmonitored use that bypasses every control the organization has built.

An AI Acceptable Use Policy (AI AUP) is the single most visible governance artifact that translates abstract AI risk appetite into concrete, day-to-day behavioral expectations. From an ISACA risk perspective it matters because it:

• Reduces data leakage risk – prevents confidential, personal, or regulated data from being submitted to external models that may retain or train on inputs.
• Protects intellectual property – clarifies ownership of AI-generated content and prohibits exposing trade secrets.
• Supports regulatory compliance – aligns employee behavior with GDPR, the EU AI Act, HIPAA, sector rules, and emerging AI legislation.
• Enables accountability – establishes that humans, not models, are responsible for outputs they use.
• Creates an auditable control – gives assurance professionals a documented standard against which to test compliance.
• Demonstrates due care – shows boards, regulators and courts that management acted reasonably.

In exam language: the AI AUP is a preventive, administrative (directive) control that operationalizes governance decisions at the user level. It is a key first-line control that supports the Govern function of the NIST AI RMF and the policy requirements of ISO/IEC 42001.

2. What an AI Acceptable Use Policy Is

An AI AUP is a formally approved document that defines who may use AI systems, which AI systems and use cases are permitted, prohibited, or require approval, what data may be used with them, and how outputs must be reviewed and disclosed. It also sets out consequences for violations and the process for exceptions.

It is distinct from, but closely related to:
• The AI Governance Policy / AI Strategy – sets the enterprise direction, principles and accountability structure (the 'why' and 'who owns it').
• The AI Development or Model Lifecycle Standard – governs how the organization builds AI (data quality, testing, validation, model risk management).
• The general IT Acceptable Use Policy – governs email, internet, and device use; an AI AUP may be a standalone document or an annex to this policy.
• Procurement / third-party risk standards – govern vetting of AI vendors.

Remember the hierarchy tested frequently by ISACA: Principles → Policy → Standards → Procedures → Guidelines. The AI AUP sits at the policy level (mandatory, approved by senior management), supported by standards (e.g., approved tool list, data classification mapping) and procedures (e.g., how to request a new tool).

3. Core Components of an Effective AI AUP

a) Purpose, Scope and Applicability
States objectives and who is covered: employees, contractors, third parties, and which systems – enterprise-licensed tools, public/consumer tools, embedded AI features in SaaS, internally developed models, and AI agents.

b) Definitions
Clear definitions of AI, generative AI, large language model, prompt, output, training data, hallucination, and 'approved tool'. Ambiguity is a common audit finding.

c) Roles and Responsibilities
• Board / executive leadership – approve and set risk appetite.
• AI governance committee or Chief AI Officer – own and maintain the policy.
• Risk, legal, privacy, security, compliance – review and advise.
• Business unit managers – enforce locally.
• All users – comply, complete training, report incidents.

d) Permitted, Restricted and Prohibited Uses
Typically a tiered model:
• Permitted – approved tools for drafting, summarizing non-sensitive material, brainstorming, code assistance with review.
• Restricted / requires approval – customer-facing deployments, decisions affecting individuals (hiring, credit, health), use of personal data, high-risk categories under the EU AI Act.
• Prohibited – entering confidential or regulated data into unapproved public tools, using AI to generate deceptive content, bypassing security controls, autonomous decisions without human review, uses that violate law or ethics.

e) Data Handling Rules
Maps to the data classification scheme: which classes (public, internal, confidential, restricted) may be processed by which tool tier. Addresses retention, opt-out of model training, and data residency.

f) Output Accountability and Human Oversight
Users remain responsible for verifying accuracy, bias, copyright and appropriateness. Requires human-in-the-loop for consequential decisions and prohibits presenting AI output as sole authority.

g) Transparency and Disclosure
When AI use must be disclosed to customers, regulators, or internally (e.g., labeling AI-generated content, informing individuals they are interacting with a chatbot).

h) Security Requirements
Use of enterprise accounts and SSO, prohibition of API keys in code, prompt injection awareness, no uploading of credentials.

i) Intellectual Property and Legal
Ownership of outputs, respect for third-party IP, licensing of AI-generated code, contractual obligations with clients.

j) Monitoring, Enforcement and Consequences
States that use may be monitored (consistent with privacy law), how violations are investigated, and disciplinary consequences.

k) Exception and Approval Process
Defined pathway to request new tools or use cases, with risk assessment, legal/privacy review and documented approval.

l) Training and Attestation
Mandatory awareness training and periodic acknowledgment of the policy.

m) Review Cycle
At least annual review, plus triggered reviews when new tools, regulations, or incidents arise.

4. How an AI AUP Works in Practice – The Lifecycle

Step 1: Establish governance foundation. Define AI principles, risk appetite and accountable executive. The AUP must flow from these; a policy with no owner will fail.

Step 2: Inventory current AI use. Survey users, analyze proxy/CASB logs, and review SaaS contracts to discover shadow AI. You cannot govern what you have not identified.

Step 3: Perform risk assessment. Classify use cases by impact (data sensitivity, decision consequence, regulatory exposure, reputational effect). This drives the permitted/restricted/prohibited tiers.

Step 4: Draft the policy with cross-functional input. Legal, privacy, security, HR, procurement, business leaders, and ethics. ISACA emphasizes that policies developed in isolation by IT are rarely effective.

Step 5: Approve at the appropriate level. Senior management or board-level committee approval gives the policy authority.

Step 6: Communicate and train. Role-based training, practical examples, FAQs, and attestation. Awareness is the control that makes the policy real.

Step 7: Implement supporting technical controls. Approved tool catalog, DLP rules for prompts, web filtering of unapproved AI sites, enterprise tenant configuration with training opt-out, logging.

Step 8: Monitor and measure. KPIs/KRIs such as: percentage of staff trained, number of unapproved tool attempts blocked, policy exceptions granted, AI-related incidents, audit findings.

Step 9: Enforce consistently. Investigate violations, apply consequences uniformly, and feed lessons into training.

Step 10: Review and update. Regulatory change (e.g., EU AI Act obligations), new model capabilities (agents, multimodal), incidents and audit results trigger revision.

5. Framework Integration

The AAIR exam expects you to connect the AI AUP to recognized frameworks:

• NIST AI RMF – The Govern function (GOVERN 1: policies, processes and procedures are in place; GOVERN 4: organizational culture and accountability) directly calls for documented policies on acceptable use. Map and Manage functions rely on the use-case categorization an AUP establishes.
• ISO/IEC 42001 (AI Management System) – Clause 5.2 requires an AI policy; Annex A controls include acceptable use of AI systems, responsible use objectives, and communication to interested parties.
• ISO/IEC 23894 (AI Risk Management) – the AUP is a risk treatment measure.
• EU AI Act – prohibits certain practices (Article 5), imposes obligations for high-risk systems, and requires AI literacy of staff (Article 4); an AUP is a primary mechanism to meet these.
• COBIT – APO01 (Managed I&T Management Framework), APO12 (Managed Risk), APO13 (Managed Security), and the governance component 'Principles, Policies and Frameworks'.
• ISACA Risk IT / Digital Trust Ecosystem Framework – policies as enablers that translate risk appetite into practice.
• Privacy frameworks (GDPR, NIST Privacy Framework) – lawful basis, purpose limitation and data minimization constraints on prompts.

6. Common Weaknesses Auditors Find

• Policy exists but no approved tool list – users cannot tell what is allowed.
• No linkage to data classification.
• Vague prohibitions ('use responsibly') with no examples.
• No exception process, so staff bypass the policy entirely.
• No ownership or review date.
• Not communicated to contractors and third parties.
• No technical enforcement, so compliance is unverifiable.
• Does not address embedded AI in existing SaaS tools.
• Fails to address output accountability and human review.

7. Exam Tips: Answering Questions on AI Acceptable Use Policies

Tip 1 – Identify the control type. An AI AUP is a preventive and directive/administrative control. If a question asks what type of control a policy is, do not choose 'detective' or 'technical'. Monitoring of policy compliance is detective; the policy itself is preventive.

Tip 2 – Think 'governance first'. ISACA answers almost always favor establishing governance, ownership, and risk appetite before technical fixes. If asked for the first step in addressing uncontrolled generative AI use, prefer 'define and approve an acceptable use policy aligned to risk appetite' or 'perform an inventory and risk assessment' over 'block all AI sites'.

Tip 3 – Senior management approval is key. Questions on what gives a policy authority or what is the most important factor for effectiveness usually point to executive/board approval and visible sponsorship, followed by communication and training.

Tip 4 – Link to data classification. When a scenario involves employees entering data into public AI tools, the best answer typically connects the AUP to the data classification scheme and DLP controls, rather than banning AI outright. ISACA favors risk-based, proportionate responses over blanket prohibition.

Tip 5 – Human accountability never transfers to the model. If a choice suggests the AI vendor or the tool is accountable for outputs, it is wrong. The policy should state users remain accountable and must verify outputs.

Tip 6 – Policies without enforcement are incomplete. If asked how to ensure compliance, look for answers combining training/attestation with monitoring and technical enforcement (approved tool catalog, CASB/DLP, logging). Policy alone is 'necessary but not sufficient'.

Tip 7 – Exceptions must be formal. Watch for answers that mention a documented, risk-assessed exception process. 'Manager verbal approval' is a wrong answer.

Tip 8 – Scope includes third parties and embedded AI. Questions that test completeness often hinge on whether the policy covers contractors, vendors, and AI features embedded in existing software, not just standalone chatbots.

Tip 9 – Know the framework mappings. Be ready to associate the AUP with NIST AI RMF Govern, ISO/IEC 42001 AI policy requirements and Annex A acceptable-use controls, EU AI Act Article 4 AI literacy and Article 5 prohibited practices, and COBIT APO01.

Tip 10 – Watch the qualifiers. 'MOST important', 'BEST', 'FIRST', 'PRIMARY' change the answer. For 'FIRST', think inventory/risk assessment or governance mandate. For 'MOST important' for effectiveness, think management support and communication. For 'BEST' to reduce data leakage, think classification-linked rules plus DLP.

Tip 11 – Eliminate extremes. Options such as 'prohibit all AI use', 'allow unrestricted use', or 'rely solely on vendor terms of service' are distractors. The exam rewards balanced, risk-based governance.

Tip 12 – Distinguish the policy from related documents. If a question describes model validation, bias testing, or training data lineage, it is about development/model risk standards, not the AUP. If it describes employee behavior with AI tools, it is the AUP.

Tip 13 – Review cadence. Policies must be reviewed at least annually and upon significant change (new regulation, major incident, new technology). Answers that imply a one-time effort are wrong.

8. Practice Questions

Q1. An organization discovers employees are entering client contracts into a public generative AI tool. Which is the risk practitioner's BEST recommendation?
A. Immediately block all generative AI websites.
B. Develop an AI acceptable use policy that links permitted tools to the data classification scheme and implement DLP controls.
C. Ask the vendor to confirm it does not store data.
D. Require employees to sign a nondisclosure agreement.
Answer: B. It is risk-based, addresses root cause, and combines directive and technical controls. A is an extreme response that drives shadow AI; C and D do not address the behavior.

Q2. Which of the following MOST increases the effectiveness of an AI acceptable use policy?
A. Detailed technical definitions of machine learning.
B. Senior management approval, communication and role-based training.
C. A list of every AI vendor on the market.
D. Publishing the policy on the intranet.
Answer: B. Authority and awareness drive compliance. D alone is passive; A and C do not address behavior.

Q3. An AI acceptable use policy is primarily which type of control?
A. Detective technical control.
B. Corrective control.
C. Preventive administrative control.
D. Compensating control.
Answer: C.

Q4. Who should remain accountable for decisions made using output from a generative AI tool, according to a well-designed AI AUP?
A. The AI vendor.
B. The IT department.
C. The human user and their management.
D. The model.
Answer: C. Accountability cannot be delegated to a system or vendor.

Q5. An auditor reviewing an AI AUP notes it does not address AI features embedded in existing productivity software. What is the PRIMARY concern?
A. The policy is too long.
B. The policy scope is incomplete, leaving significant AI use ungoverned.
C. Users will need more training.
D. Vendor contracts must be renegotiated.
Answer: B. Scope completeness is a fundamental policy requirement.

Q6. Which NIST AI RMF function MOST directly addresses the establishment of AI acceptable use policies?
A. Map
B. Measure
C. Manage
D. Govern
Answer: D. Govern covers policies, processes, accountability and culture.

9. Quick-Reference Summary

• What: Mandatory, management-approved policy defining permitted, restricted and prohibited AI use, data rules, output accountability, disclosure, enforcement and exceptions.
• Why: Controls shadow AI, data leakage, IP exposure, regulatory breach and reputational harm; demonstrates due care.
• How: Inventory → risk assess → cross-functional drafting → executive approval → training/attestation → technical enforcement → monitoring → review.
• Control type: Preventive, administrative/directive.
• Frameworks: NIST AI RMF Govern; ISO/IEC 42001 AI policy and Annex A acceptable-use controls; EU AI Act Articles 4 and 5; COBIT APO01/APO12.
• Exam mindset: Governance first, risk-based not absolute, human accountability, policy plus enforcement, formal exceptions, complete scope, periodic review.