Executive and vCISO Insights
AI Acceptable-Use Policy for Employees and Administrators: What Can Be Entered, Shared, Automated, and Published
Give employees and administrators clear AI rules based on data, authority, validation, external impact, records, and reporting.
An AI policy should help people complete useful work safely—not merely prohibit a technology that may already be present in browsers, productivity suites, mobile devices, software-development tools, and business applications.
The clearest policy answers four questions: which services are approved, what information may enter them, what authority may be delegated, and which outputs require verification or approval before use.
Write the policy around risk decisions
A policy that says “use AI responsibly” is too vague to guide a person who needs to summarize a customer document, analyze a spreadsheet, draft code, create a public image, or connect an agent to email. A policy that bans every AI feature may be impossible to enforce because AI is embedded in common applications.
Use a risk-based structure. Define approved services and accounts; classify data; distinguish assistance from action; require verification; identify prohibited outcomes; and explain how to request an exception or report a problem. Assign a business owner, security owner, privacy or compliance reviewer where relevant, and technical administrator.
Define prohibited, conditional, and approved use
Prohibited without a separately approved path
- Entering passwords, API keys, private keys, authentication tokens, recovery codes, or other secrets.
- Submitting protected health, payment-card, taxpayer, legal, employee, customer-confidential, export-controlled, or other restricted data to an unapproved service or account.
- Using AI to impersonate a person, fabricate evidence, misrepresent authorship, bypass a security control, or make an unlawful discriminatory decision.
- Allowing an agent to send money, change privileged access, delete production data, publish externally, or execute disruptive security actions without explicit authorization and safeguards.
- Publishing unverified allegations, incident claims, legal conclusions, medical advice, compliance determinations, or copyrighted material presented as original.
Conditional use
Conditional use may include customer communications, regulated information, source code, security findings, recruitment, performance decisions, legal documents, financial analysis, clinical support, public content, images of real people, connectors to internal systems, or action-capable agents. Require the appropriate approved platform, purpose, data minimization, documented review, and professional approval.
Generally approved use
Examples may include brainstorming, formatting, rewriting an organization’s own non-confidential draft, summarizing approved public information, creating internal training questions, or explaining common technical concepts—provided the output is checked and the account is approved. Each organization must define its own boundary.
Set rules for information entered into AI
- Use only organization-approved services and managed accounts for company work.
- Minimize data to what the task actually needs.
- Remove direct identifiers when the business purpose does not require them.
- Do not paste credentials, secrets, private keys, or recovery material.
- Follow existing classification, privacy, records, customer, and contract requirements.
- Verify that attachments, images, transcripts, and connector sources are permitted—not only typed prompts.
- Do not move data from a restricted system into AI merely because access is technically possible.
- When uncertain, stop and use the exception or review process.
The organization should maintain a table mapping each approved service to account type, business owner, permitted data classifications, prohibited data, connectors, retention, publication rules, and support contact. The enterprise AI data-privacy guide explains why consumer and business paths must be evaluated separately.
Require output verification based on consequence
Generative AI can produce convincing but false or incomplete output. Verification should be proportionate to the decision. A private brainstorming note needs less review than a customer notice, regulatory statement, software change, security response, financial decision, clinical communication, or public allegation.
| Output | Minimum review | Additional control |
|---|---|---|
| Internal low-risk draft | Author checks accuracy, relevance, and confidentiality | Remove unsupported claims and retain sources when needed |
| Customer or public content | Qualified human fact, privacy, copyright, and brand review | Approval before publication; preserve source and asset rights |
| Code or configuration | Peer review, secure testing, dependency and secret checks | Staging, change approval, rollback, and production monitoring |
| Security analysis | Validate against original telemetry and authoritative sources | Do not act on attribution or severity solely from generated output |
| Regulated or high-impact decision | Qualified professional review under applicable procedure | Document evidence, rationale, approvals, and appeal or correction path |
Separate assistance from delegated action
An assistant produces a suggestion. An agent or integration may send email, change a record, operate a browser, call an API, create code, or schedule later work. The policy should require a documented owner and authorization before any AI system receives tool access or persistent credentials.
Define actions that always require human confirmation. Use separate identities, least privilege, approved destinations, transaction and volume limits, time-bound access, logging, and a tested stop procedure. Prohibit agents from weakening security controls or approving their own elevated access.
For technical depth, see Securing AI Agents for Business.
Cover records, copyright, disclosure, and incidents
Records and evidence
State whether prompts, files, outputs, approvals, model versions, or evaluation results are business records and how they should be retained. Avoid logging sensitive content without a lawful and necessary purpose. Preserve enough evidence to investigate important actions and correct published material.
Copyright and rights
Do not assume generated content is free of third-party rights. Public-facing text, images, code, presentations, and reports require originality, source, license, trademark, privacy, and publicity-rights review appropriate to the use. The U.S. Copyright Office’s AI initiative explains that copyright questions depend on the nature of human authorship and the material involved; qualified counsel should address legal conclusions.
Disclosure
Define when customers, partners, employees, or the public should be told that AI contributed to content or a process. Do not use disclosure as a substitute for accuracy or review.
Incident reporting
Require prompt reporting of restricted-data entry, unexpected disclosure, suspicious connector behavior, prompt injection, harmful or discriminatory output, incorrect public content, unauthorized action, lost credentials, or evidence that an account was compromised. Provide a contact path outside the affected AI service.
Implementation checklist for leadership and IT
- Discover current AI use before finalizing the rule.
- Approve a small set of managed platforms and publish the permitted-use matrix.
- Configure identity, provisioning, sharing, connectors, retention, logs, and administrative roles.
- Train by role with examples from the organization’s real work.
- Create a rapid exception path so legitimate needs do not become shadow AI.
- Monitor for unmanaged applications, browser extensions, API keys, and agent connections.
- Review a sample of important outputs and actions for evidence that the policy works.
- Correct mistakes without hiding them, and update the policy after material platform or risk changes.
The policy should be approved through the organization’s normal governance process and reviewed by qualified legal, privacy, human-resources, compliance, and labor professionals when their requirements apply. This article is an educational framework, not a ready-to-sign legal policy.
Assign responsibilities and an exception path
| Role | Policy responsibility |
|---|---|
| Executive sponsor | Approves risk direction, resources, accountability, and high-impact exceptions |
| Business owner | Defines purpose, validates benefit, owns output and process consequences |
| IT administrator | Operates identity, licensing, tenant settings, connectors, devices, logs, and offboarding |
| Security | Threat-models use, reviews access and monitoring, tests high-risk workflows, supports incidents |
| Privacy, legal, compliance, HR, records | Reviews matters within professional responsibility and applicable procedure |
| User | Uses approved paths, minimizes data, verifies output, obtains required approval, reports concerns |
The exception process should ask for the business need, proposed service and account, data, users, connectors, actions, duration, alternatives, evidence, owner, and compensating controls. Set an expiration. Emergency approval should not become permanent through inattention.
Enforcement should be proportionate and consistent with employment agreements, labor rules, policy, and law. Corrective action should consider whether the organization provided a usable approved alternative, clear training, and an effective reporting path. The aim is safer behavior and accountability, not hidden experimentation.
Policy language to tailor through formal review
Approved services: Organization information may be processed only through AI services, account types, models, connectors, and features approved for that use. Personal accounts and unapproved browser extensions may not be used for organization-restricted information.
Human responsibility: The person using or approving AI-assisted work remains responsible for verifying accuracy, protecting information, following professional and organizational requirements, and obtaining required approval before an output creates an external effect.
Agents and automation: AI systems may not receive persistent credentials, tool access, or authority to send, publish, purchase, change access, alter production, or delete data unless the use has a named owner, documented authorization, least privilege, limits, logging, human approval where required, and a tested stop procedure.
Reporting: Users must promptly report suspected restricted-data entry, unauthorized disclosure, compromised access, unsafe action, manipulated instruction, materially incorrect publication, or other AI-related security or compliance concern through the designated incident channel.
These examples are starting language only. The organization should reconcile them with its information security, privacy, acceptable use, records, intellectual property, incident response, human resources, procurement, change-management, and professional-practice policies.
Questions employees should be able to answer
Can I use AI if the material is already on the Internet?
Public availability does not prove accuracy, permission, currentness, or suitability for the business purpose. Use approved services, respect licenses and terms, verify the authoritative source, and follow publication review.
What should I do after entering restricted information by mistake?
Stop further use, do not conceal or repeatedly copy the content, and report the event through the designated channel. Provide the service, account, time, information type, sharing or connector details, and actions taken so the response team can contain and evaluate it.
Sources
- NIST: AI Risk Management Framework
- NIST AI 600-1: Generative AI Profile
- U.S. Copyright Office: Copyright and Artificial Intelligence
- CISA and UK NCSC: Guidelines for Secure AI System Development
- OpenAI: Enterprise privacy and business data
- Microsoft Learn: Enterprise data protection for Microsoft 365 Copilot and Copilot Chat
Update and correction history
- August 2026: Initial policy framework prepared from NIST, CISA, U.S. Copyright Office, and official provider guidance. It is not a substitute for organization-specific legal or compliance review.
Informational Use and Verification Notice
The OC Security Audit Cybersecurity Intelligence Center is provided for general educational and security-awareness purposes only. Its content is based on publicly available sources believed to be reliable at the time of review; however, product capabilities, technical conditions, laws, regulations, and other facts may be incomplete, disputed, corrected, or changed after publication. OC Security Audit does not represent or warrant that every statement is complete, current, or error-free.
Do not rely on this content as the sole basis for cybersecurity, legal, compliance, financial, operational, procurement, or other decisions. Independently verify material information, review the cited primary sources and current contract terms, evaluate how the subject applies to your environment, and consult qualified professionals when appropriate. To the fullest extent permitted by applicable law, OC Security Audit is not responsible for loss or damage arising from decisions or actions taken solely in reliance on this content without appropriate independent verification or professional review.
Use of the Intelligence Center does not create a professional-client, auditor-client, attorney-client, fiduciary, or other advisory relationship. Nothing in this section is a guarantee of security, compliance, prevention, accuracy, or outcome, and this notice does not replace the website’s governing terms or any written engagement agreement.
Turn AI policy into operating controls
OC Security Audit can help align AI acceptable use with data classification, identity, Microsoft 365, vendor risk, security monitoring, compliance readiness, and executive governance.