An AI acceptable use policy (AI AUP) is an internal document that defines which AI tools your staff can use, what data they can upload, when a human must verify AI output, and who to contact when a case is unclear. The policy should be an operational guide for how to use AI safely at work, avoiding the risks of data loss and compliance violations that can result when employees make their own judgment calls about which tools to use.
Your employees are already using these tools, and the data they upload can include confidential internal or customer information, creating a security risk known as “shadow AI.” To mitigate the risks associated with shadow AI, creating and enforcing an AI policy is essential. But simply trying to ban these tools has become impractical, as AI is now embedded by default in most online platforms such as Microsoft 365, Google Workspace, Zoom, etc.
Instead, the goal of a strong AI policy is to establish practical rules that people will follow instead of bans they try to work around. A successful rollout requires operational clarity across several key areas. This guide covers the critical decisions leaders must make regarding approved AI tools, data rules and categories, human review processes, account management, incident response, and training.
Key takeaways for creating an effective AI acceptable use policy:
- Make your workforce aware of the risks of shadow AI.
- Set data rules with real-world examples.
- Organize your approved AI tools into tiers.
- Keep the AI policy document to one or two pages.
- Assign an operational owner for ongoing AI policy management.
What shadow AI looks like in a small business
Data security today means knowing what your employees are actually doing with AI: what data is getting uploaded to these tools and who gets access to it. Your staff may be inputting company or client details into ChatGPT or Copilot to create internal documents. Or they may install unvetted AI browser extensions without considering the data privacy consequences. This “shadow AI” trend has become a leading security risk for small and mid-sized organizations.
Everyday examples of behavior contributing to shadow AI:
- Pasting customer or internal details into a free chatbot
- Opening work files in personal accounts with AI features
- Relying on unverified AI output for deliverables
- Installing unvetted third-party plugins that access internal workspaces
- Allowing AI features to remain switched on in apps the organization uses
Four ways shadow AI can go wrong:
- Your data is handled under the AI tool provider’s terms, bypassing your network restrictions and compliance obligations.
- Incorrect, unverified AI output is delivered to a customer, damaging your reputation and possibly invalidating a contract.
- A regulatory obligation is breached without anyone noticing, possibly leading to financial or legal penalties.
- An ad hoc AI tool ecosystem lacks oversight, making it difficult to determine which tools need to be reviewed or deactivated.
One of the most frequent situations we are encountering at Fidelis is our clients failing to recognize the security level needed to implement AI. Once we have demonstrated what safe AI use entails, we can review their technology licenses and implementation, and then train their end users.
What an AI acceptable use policy should include
A usable AI policy is a practical operational roadmap, not a lengthy legal document. To ensure employees actually read and follow it, keep the policy under two pages. It should clearly outline six core operational decisions, structured in order from tool approval to long-term governance.
1. Approved tools, sorted by tier
Your staff need to know which AI tools are safe for business tasks and they are allowed to use. The simplest way to manage this is by grouping the tools into approval tiers, similar to the white, gray, and black designations given to other software.
| Tier | What qualifies | Data allowed | Examples |
|---|---|---|---|
| Tier 1: Approved business tools | Company-managed accounts with data protections and secure sign-in | Internal business data, and sensitive data where the tool's terms and your obligations allow it | Microsoft Copilot Chat included in many Microsoft 365 plans |
| Tier 2: Approved with limits | Vetted business tools with acceptable privacy terms but not cleared for sensitive data | Internal, nonsensitive data | Paid AI tools such as ChatGPT, Claude, or Gemini |
| Tier 3: Approved for public information only | Free or consumer AI tools reviewed and added to the company's approved list | Public information only; no customer, employee or financial data | Consumer or free versions of ChatGPT, Claude, Gemini, etc. |
| Not yet approved | Anything not on the approved list | None; employees must ask for a review before using it for work | Free consumer versions of public chatbots and unvetted browser extensions |
To maintain security, employees should be trained to distinguish an enterprise-ready, company-managed business account from a personal or free account. The distinction comes down to data governance: an enterprise account guarantees your inputs are never used to train public AI models, defines explicit data retention timelines, provides administrative oversight, and requires users to sign in using corporate identity credentials.
It’s important to use caution with the frontier AI providers such as OpenAI (ChatGPT), Anthropic (Claude), or Google (Gemini). Never assume one of their plans is safe or that its security features remain permanent. You need to actively configure account settings to opt out of their AI model training. And like any software provider, their terms of service change frequently, meaning your operations team must regularly verify and recheck data handling clauses to ensure your proprietary inputs are genuinely being excluded from the vendors’ training pools.
For organizations running or migrating to Microsoft 365, your approved list should naturally begin with tools covered directly by your own tenant and corporate licensing.
2. Data rules and categories
Vague directives like “don’t share sensitive information” fail because employees define sensitivity differently. Your AI policy must replace ambiguity with concrete, real-world examples. Categorize data three ways: data that never enters any AI tool, data restricted to approved accounts only, and public information safe for any platform.
For the first category, certain data classes are considered confidential or personally identifiable information (PII) and are entirely off-limits from public AI models. Examples include:
|
|
Specific industries have special considerations. Nonprofit IT management requires the protection of donor databases and client case histories to safeguard community trust and maintain nonprofit status. Manufacturers and defense suppliers must ensure their technical data never leaks. Similarly, financial services firms must protect client wealth profiles and tax histories. Beyond these industry specifics, general constraints apply to all contracts under NDA, corporate client lists, and protected health information.
3. Human review and accountability
An AI tool can assist with production, but a human must own the final result. Your AI policy must state explicitly that the individual who sends or publishes a deliverable bears total accountability for its accuracy and compliance.
Automated outputs cannot bypass human oversight, especially when errors carry severe consequences. Review is mandatory for all customer-facing messages, reports, financial figures, legal or compliance submissions, software scripts, and any documentation affecting an employee’s status or an applicant’s hiring process.
Consider a real-world scenario from the QuickBooks accounting software environment. A Fidelis client used an AI assistant to troubleshoot a QuickBooks software issue. The AI generated a confident, well-written answer, but based it on a different software version. Following the AI’s advice led to a major configuration mistake that took hours for our team to untangle. If a person had cross-checked the AI’s steps against the actual system version before applying them, the disruption would have been avoided completely.
4. Accounts, identity, and access
A policy is only as strong as the access controls enforcing it. Employees must access approved platforms only with company-managed sign-ins protected by multifactor authentication (MFA). Personal accounts must never touch company data.
To keep your digital perimeter secure, staff must obtain explicit IT approval before activating any AI plugins, browser extensions, or software connectors. Left unchecked, these utilities can quietly scrape data from internal workspaces. IT administrators must place strict boundaries on what an AI tool can connect to within the corporate network. Furthermore, security teams must track AI licenses continuously and ensure all access is immediately revoked during employee offboarding.
Managing these endpoints is a critical layer of modern cybersecurity defense. It binds identity tracking to your broader operational security, a process deeply integrated into identity threat detection and response (ITDR) frameworks.
5. Requesting new tools and exceptions
If technology adoption is slow or confusing, employees inevitably resort to behaviors that contribute to shadow AI, such as installing AI plugins or working from personal devices. To prevent staff from bypassing your AI security measures, establish a fast, simple decision-making process for approving or rejecting the adoption of new AI tools. The goal is to encourage innovation without compromising data security.
Provide a short request form, managed by a named reviewer, and commit to a clear, predictable decision window (such as three business days). The review process should evaluate standardized criteria: how the AI vendor handles data, their terms of service regarding training models, their supported sign-in methods, and exactly which internal systems the tool will connect to. A consistent review process allows employees to explore new AI tools through an established approval pathway while providing visibility into potential risks and maintaining appropriate oversight.
6. Incidents, training, and reviews
Your AI policy will only protect your business if it lives actively within your daily culture. To help your teams understand the policy and their responsibilities, define clearly what constitutes an AI security incident, such as accidentally pasting client data into a public chatbot. Then establish a clear, no-blame reporting path so employees know who to contact and feel safe disclosing mistakes immediately.
Once your policy is created, train your staff on what is acceptable and what is restricted. Pair the document launch with a concise, 30-minute role-based training session, and require a signed acknowledgment from all users. As teams begin using approved tools, hold open Q&A sessions to share successful use cases and discuss what didn’t work. This cross-trains the entire organization and keeps your approved AI tool list relevant.
At Fidelis, we regularly facilitate and host these training sessions and interactive Q&As to ensure our clients stay secure.
Approved and prohibited AI use by role
This table shows how an AI acceptable use policy impacts individual teams using AI tools within an organization. It references the AI tool-approval tiers defined earlier.
| Role | Public-information tasks (Tier 1, 2, or 3) | Approved business tools only (Tier 1) | Prohibited |
|---|---|---|---|
| Sales | Draft a follow-up email using public information about a prospect | Summarize notes from a call with a named prospect | Paste a CRM export or prospect list into a free tool |
| Finance and accounting | Explain a general accounting concept | Draft a question about a software issue, naming the exact software and version | Paste payroll, bank details or full financial statements into a tool not approved for sensitive data |
| HR | Draft a generic job description | Rewrite policy wording using internal documents | Upload resumes, performance reviews or medical information to an unapproved tool |
| Customer service | Draft general reply templates | Draft a careful reply to a specific customer issue, without account numbers | Include account numbers, payment details or personal data in a prompt |
| Operations and projects | Outline a meeting agenda | Summarize meeting notes that include client names | Upload client contracts or technical drawings to an unapproved tool |
| IT | Explain a general error message | Review a script that contains no credentials | Paste passwords, keys or internal configurations |
How to roll out your AI policy without slowing people down
If you currently have no documented AI policy, your immediate goal is to establish operational guardrails without damaging employee productivity. Implementing an AI acceptable use policy successfully requires a practical, step-by-step rollout that balances data security with your team’s workflow.
1. Find out what is already in use
Begin by executing a transparent, no-blame assessment to find out which AI platforms each department currently uses and how they use them. Simultaneously, have your IT team or managed service provider audit your environment for unauthorized browser extensions, desktop software, and cloud sign-ins. This baseline discovery ensures you are addressing the actual state of shadow AI within your daily operations.
2. Choose and license the tools you will approve
Consolidate your technology stack by selecting a minimal number of effective platforms to officially sanction. Transition your staff away from free, personal accounts and move them onto company-managed enterprise licenses that offer built-in privacy protections. Providing approved, secure tools naturally deters employees from seeking out unverified public alternatives.
3. Write the one-page version
Draft a concise, single-page version of your policy using the six structural headings detailed in Section 3 above. Keep the rules straightforward and focused on immediate data safety, so the document remains scannable and easy for staff to internalize. You can always expand the framework as your organization’s AI adoption matures.
Free AI Acceptable Use Policy Template (PDF)
Need a starting point? Download our AI Acceptable Use Policy Template, covering approved AI tools, data restrictions, human review, incident reporting, and policy ownership. Adapt it to your organization and have your leadership and IT team review it before implementation.
4. Train with real examples
Launch the policy alongside short, interactive, role-based training sessions tailored to specific department workflows. Use concrete, everyday scenarios to demonstrate how to use AI safely at work, and provide a clear open forum where staff can ask questions and feel supported.
5. Check and adjust
Treat your AI acceptable use policy as a living document by establishing a recurring schedule to review new software requests and log any data incidents. Use these operational insights to continuously refine your rules, remove risky applications, and add newly vetted ones to your approved list.
Who owns the AI policy day to day?
While leadership defines risk tolerance and signs off on the framework, an AI acceptable use policy requires an ongoing operational owner. This person manages the approved tool list and request queue, tracks licenses, and coordinates ongoing training.
If your organization has internal IT, they can gather data to help leadership make informed, strategic decisions, leveraging an outside provider for co-managed IT support if necessary. Conversely, if you operate without internal IT, your external IT partner may be able to take on this role.
Your IT provider will need to understand your specific workflows, compliance obligations, and security needs. They will also need to research appropriate AI tools, analyze vendor privacy changes, and deliver clear operational recommendations directly to your leadership team.
Fidelis provides this role through our IT Manager as a Service offering, guiding and maintaining the policy with internal leadership retaining the final word on operational risk.
Why AI policy needs security controls
Your AI policy will set expectations for employee behavior, but it cannot block them from making mistakes. Mistakes will happen, and effective security controls can limit the damage. To truly protect your organization, your policy guidelines should be backed by proactive technical safeguards that activate when someone inevitably ignores the rules.
A robust defense requires layered security controls across your entire digital environment. This begins with enforcing MFA for user access across all accounts and deploying advanced endpoint protection on every device.
Your IT team should also secure communications using intelligent email filtering and restrict the installation of unapproved software and risky browser extensions. Where your software licensing supports it, they should configure data loss prevention (DLP) rules to actively block sensitive information, such as credit card details or Social Security numbers, from being pasted into web forms. Finally, continuous sign-in and activity logging can be paired with the active monitoring of cloud accounts to flag early signs of credential takeover.
While technical controls like these cannot guarantee a complete prevention of data leaks, they can drastically reduce your attack surface. For answers and guidance on setting clear policy boundaries at your workplace, check out our Fidelis professional cybersecurity services in Seattle.
Frequently asked questions about AI acceptable use policies
Do small businesses need an AI acceptable use policy?
Yes. If your staff uses internet-connected tools, they are very likely already using AI. Without an AI acceptable use policy, your organization may be exposed to data leaks, compliance violations, accusations of plagiarism, and intellectual property risks that can threaten your operations.
Should we ban AI tools at work?
No. Banning AI typically drives its use underground, fueling security liabilities while stalling your team’s productivity. A successful approach replaces a blanket ban with clear, structured boundaries that encourage safe, efficient innovation.
Can employees use free AI tools for work?
Yes, but only if the organization has approved the tool for public-information (Tier 3) tasks. Free, consumer-grade AI tools usually retain the data users input to train their models, which can compromise proprietary or personal information. To maximize secure operations, use managed accounts with verified, strict data privacy guarantees.
What is shadow AI?
Shadow AI refers to the unauthorized use of AI applications or browser extensions within an organization without the knowledge, vetting, or oversight of the IT team or leadership.
How long should an AI acceptable use policy be?
One or two pages is ideal. A short, highly scannable document ensures that employees actually read, remember, and follow the guidelines in their daily workflows.
How often should we update it?
Review the framework at least every six months. AI tools, user terms, and regulatory compliance rules evolve rapidly, requiring continuous oversight to keep your approved tool list and security configurations accurate. Navigating these rapid shifts can often benefit from the guidance of an experienced IT professional.
Setting AI rules for your team? Start with what your staff already use
We saw at the outset that data security today means having visibility into your employees’ behavior with AI: what tools are they using, what data are they sharing with these tools, and how are they using AI to augment their work?
If you’re worried about uncontrolled AI use in your organization and the risks associated with “shadow AI”, reach out to us today. We can review your technology stack and answer your questions about AI. And if you already know you need to take control of AI in your organization, we can guide you toward creating an AI acceptable use policy that protects your data while allowing your team to leverage AI’s potential.
Is your team already using AI tools?
Tell us which tools your staff use and what data they handle. We can help you set clear rules, approve the right tools, and keep them current.
