AI Acceptable Use Policy: Scope, Rules and Enforcement

An AI acceptable use policy is a formal document that states which generative AI tools employees may use, which data may enter them, which uses are prohibited, and who answers for the output. An adequate policy in the United States covers eight components (purpose and scope, user responsibilities, prohibited uses, data governance, security, legal compliance, reporting and consequences, and review), assigns a named owner, and is reinforced by training and a defined review cycle. This article sets out each component, the evidence behind it, and a template outline that an organization can adapt to its own risk profile.

TL;DR

• An AI acceptable use policy defines approved tools, permitted data, prohibited uses, accountability for output, and the consequences of violation.
• The central control is data classification: public information may generally be entered into external tools, whereas confidential, customer, personal and regulated data may not.
• Employees remain responsible for reviewing and fact-checking all AI output, and material use should be documented and disclosed to a supervisor where the employer expects it.
• A policy without a named owner, a tool inventory, annual training and a scheduled review tends to become obsolete within a single planning cycle.
• The NIST AI Risk Management Framework offers a voluntary structure for connecting the policy to existing governance and risk controls.

Table of Contents

• What an AI acceptable use policy is and why organizations need one
• The components of an effective policy
• Acceptable and prohibited uses of AI tools at work
• Data rules: what may and may not be entered
• Review, disclosure and accountability for AI output
• Incident reporting
• Approving tools and connecting the policy to governance
• Enforcement, training and review cycles
• Template outline for an employee AI policy
• Sequencing the policy within an adoption plan
• Perspective
• Support in drafting and implementing the policy
• FAQ
• Sources

What an AI acceptable use policy is and why organizations need one

An acceptable use policy (AUP) is a long-standing instrument in information security: it specifies what users may and may not do with an organization’s systems. The AI variant applies the same logic to generative tools such as chat assistants, image generators, coding assistants and AI features embedded in existing software. Its object is not the technology itself but the behavior of the people using it, which is why it is distinct from a broader AI governance program that addresses model development, procurement and oversight.

The need arises from a simple asymmetry. Employees can adopt a consumer AI tool in seconds, without procurement, security review or legal sign-off, and the data they paste into it leaves the organization’s control. The American Employers Alliance notes that many AI tools use input data for training and that even tools claiming otherwise may change their terms of service, and it therefore advises treating any data entered into an external AI tool as potentially public. A written policy converts that cautionary principle into an enforceable expectation.

A policy also serves a documentary function. When a client, auditor or insurer asks how the organization governs AI use, a dated, approved, communicated policy is the first artifact requested. Organizations that have not yet inventoried their AI exposure may find it useful to begin with an AI risk assessment, since the policy’s prohibitions should reflect the risks actually present.

The components of an effective policy

TechTarget’s guidance on drafting such a policy identifies eight elements: purpose and scope, user responsibilities, prohibited uses, data governance, security requirements, compliance and legal obligations, reporting and consequences, and a review and update process. Published sample policies from universities, professional associations and law firms vary in length but converge on essentially this set.

The order matters in practice. Scope and data governance determine nearly every other provision, so they should be drafted first and reviewed by the people who understand how information actually moves through the organization.

Acceptable and prohibited uses of AI tools at work

Permitted uses are typically defined by task type: drafting and editing non-sensitive text, summarizing public material, brainstorming, generating code scaffolding that will be reviewed, and analyzing data that has been de-identified or is already public. The policy should list examples without claiming the list is exhaustive, since new use cases appear faster than revisions.

Prohibitions deserve more precision. A widely circulated law-firm template prohibits employees from using generative AI to make or assist with employment decisions, including recruitment, hiring, promotion, performance monitoring, discipline and termination. The rationale is that such decisions carry discrimination exposure and require human judgment and documentation that an opaque tool cannot supply.

A second category concerns professional advice. OpenAI’s own workplace guidance stresses that ChatGPT is not a licensed professional and advises that, for important health, legal or financial matters, employees should consult a qualified expert or defer to their organization’s policies before acting. A workplace policy for ChatGPT and comparable assistants can adopt this principle directly: output informs a professional’s work and never replaces it.

A practical drafting device is a three-column list: permitted without approval, permitted with approval or on approved tools only, and prohibited. Employees consult such a list far more readily than continuous prose.

Data rules: what may and may not be entered

Data governance is the provision with the greatest effect on actual risk. The American Employers Alliance guidance states that employees may enter publicly available information and original prompts without sensitive content, but should never enter confidential business information, trade secrets, customer data, employee personal information, financial data, legal matters, or information subject to regulatory protection such as HIPAA or FERPA.

Organizations that already classify data can map the rule onto their existing scheme. The University of California Agriculture and Natural Resources guidelines, for example, permit information at Protection Level P1 (public) to be used freely in ChatGPT, while P2, P3 and P4 information should not be used. Reusing a classification employees already know avoids creating a parallel vocabulary and reduces ambiguity at the moment of decision.

The NCSEA program adds a corollary worth adopting: employees should use the least sensitive data necessary to accomplish a task. This minimization principle is valuable because it applies even to approved tools, where the question is not whether data may be entered but how much is needed.

Policies for independent healthcare practices and other regulated small organizations should name the applicable regime explicitly and state that regulated data is excluded from any tool lacking an appropriate contractual arrangement. Such determinations require legal review specific to the organization, and the policy should designate who makes them.

Review, disclosure and accountability for AI output

A defensible policy places responsibility for output on the human user. Three duties recur across sample policies.

Verification. A law-firm policy requires employees to review all AI output before using or forwarding it, checking that it contains no biased, offensive or discriminatory content, does not improperly disclose personal or confidential information, and is accurate when compared against other trusted sources. The NCSEA program similarly makes employees responsible for reviewing and fact-checking all AI-produced material.

Disclosure. Expectations differ by employer and client. OpenAI’s guidance suggests employees keep conversation links or logs where their employer expects disclosure of AI use. A policy should state whether disclosure to a supervisor, to a client, or both is required, and for which categories of work.

Documentation. For consequential work, the NCSEA program asks that material AI use in engineering analysis, client work product or consequential decisions be documented to show the source material, validation performed, human review, and the approved system or feature used. This record is what allows an organization to reconstruct how a deliverable was produced if it is later questioned.

Incident reporting

The policy should treat accidental disclosure as an incident rather than a disciplinary event alone, because prompt reporting limits harm. The NCSEA program provides a useful model: if restricted or confidential information is entered into an unapproved tool, or exposed through an AI feature, connector, agent, memory store or log, it is to be treated as a potential data incident, with prompt notice to the designated IT or information-security contact, the supervisor and the appropriate technology or AI governance contact.

The drafting implication is that the reporting channel must be named, monitored and simple. A policy that obliges employees to report but does not identify a recipient produces silence.

Approving tools and connecting the policy to governance

A policy that merely prohibits unapproved tools invites workarounds unless an approved alternative exists. The usual solution is an inventory of authorized tools with a documented review process. Georgia Tech’s policy illustrates the pattern: its AI governance program reviews the policy annually, or more frequently as needed, and maintains inventories of authorized AI tools that are themselves reviewed at least annually.

The NIST AI RMF Playbook offers a framework for the surrounding structure. It indicates that organizational AI risk policies should define key terms, connect AI governance to existing governance and risk controls, align with data governance policies, and establish the frequency and detail of monitoring, auditing and review. The underlying framework, released on January 26, 2023, is intended for voluntary use to improve the incorporation of trustworthiness considerations into the design, development, use and evaluation of AI systems.

Tool approval should draw on vendor due diligence, including data retention, training use of inputs, access controls and contractual terms. The same discipline applies as for any third party, and the process described in the guidance on vendor risk assessments transfers directly. Where AI tools authenticate through existing accounts, identity governance determines who can reach them and with what privileges.

Enforcement, training and review cycles

Enforcement is the point at which many policies fail. A written rule that is neither taught nor monitored does not alter behavior. Three mechanisms address this.

1. Training. One law-firm policy requires generative AI training within a reasonable or specified time after hire and at least annually, and permits the employer to treat failure to participate as a violation. Linking training to the policy converts it from a document into a practice.
2. Consequences. The policy should state that violations may lead to discipline, up to and including termination, in line with existing human resources procedures, and should be reviewed by employment counsel.
3. Monitoring. Technical controls, such as restricting access to unapproved tools or inspecting outbound data flows, should be described in general terms, with the legal basis for monitoring confirmed for the relevant state.

Review frequency should reflect the pace of change. TechTarget recommends assessments at least annually, with quarterly reviews advisable given the rapid rate of change and deployment of AI. A reasonable compromise for a smaller organization is an annual formal review with a trigger-based review whenever a new tool category, regulation or incident arises.

Template outline for an employee AI policy

The following outline can be adapted to most organizations. Bracketed items require organizational decisions.

1. Purpose. State the objective: enabling productive use of AI while protecting data, clients, staff and the organization’s reputation.
2. Scope. Employees, contractors, interns and any person using organizational data, on any device or account, including AI features embedded in approved software.
3. Definitions. Generative AI, AI tool, approved tool, restricted data, confidential data, public data.
4. Approved tools. [List and owner of the inventory; process for requesting additions.]
5. Permitted uses. Representative examples, labeled non-exhaustive.
6. Prohibited uses. Employment decisions, regulated data entry, professional advice without expert review, [other organization-specific items].
7. Data rules. Classification mapping; minimum-necessary principle; treatment of inputs as potentially public.
8. Review and accountability. Mandatory verification; documentation for consequential work; disclosure expectations.
9. Incident reporting. Named contact; timeframe; no-blame reporting of accidental disclosure.
10. Training. At hire and annually.
11. Enforcement. Consequences consistent with existing policies.
12. Review. [Owner, cadence, triggers]; version history.
13. Acknowledgment. Signature or electronic attestation.

A first draft of this length is typically sufficient for organizations under a few hundred employees. Larger or regulated organizations will append sector-specific schedules.

Sequencing the policy within an adoption plan

A policy is most effective when it is one component of a broader adoption plan rather than a standalone reaction to employee behavior. Readiness work, which clarifies current tool use, data flows and risk tolerance, should precede drafting; a 12 to 24 month AI adoption roadmap places that readiness phase first for this reason. Gaps between the policy and actual practice can then be tracked in a compliance gap register with named owners, so that each deviation has an accountable party and a closure date.

Perspective

Our experience in technical advisory work is that an AI acceptable use policy succeeds or fails according to whether a person under time pressure can apply it in seconds. Policies written as legal prose are seldom consulted at the moment a project manager decides whether a client spreadsheet may be pasted into an assistant. For that reason we favor short, tiered rules, a classification scheme the organization already uses, and one named contact for questions.

We also hold that the policy is a risk instrument before it is a compliance artifact. It should follow from a documented assessment of where AI is already in use, which data it touches, and which vendors hold that data. Our work combines technical review with leadership practice because a rule that leaders cannot explain calmly, and staff cannot recall under pressure, does not reduce risk.

Finally, we regard review as a standing duty and not an event. A policy with an owner, an inventory and a scheduled revision remains accurate as tools change, and it gives leadership a defensible position when clients and auditors ask how AI is governed.

Jesse Hart

Support in drafting and implementing the policy

PROJECT-JTH provides technical advisory services, including systems and risk review, for organizations that need a policy grounded in how their systems and data actually operate. Engagements can cover the assessment that precedes the policy, the technical controls that support it, and the communication of expectations to staff and leaders. Details of the consulting offering are available on the technical consulting and risk review page.

Independent consultants who deliver assessments and reports to clients can also review Project Praxis, priced at $99 per month, and teams managing production records and contracts can review Project Relay, priced at $49 per month. Further articles on risk and governance are collected in the PROJECT-JTH Insights weblog.

FAQ

What is the difference between an AI acceptable use policy and an AI governance policy?

An acceptable use policy regulates the conduct of employees who use AI tools: permitted tools, data and tasks. A governance policy is broader and addresses accountability structures, risk management, procurement and monitoring across the AI lifecycle. The NIST AI RMF Playbook suggests that AI risk policies should connect to existing governance and risk controls, so the two documents should reference one another.

Can employees enter customer data into ChatGPT?

Guidance from the American Employers Alliance advises against entering customer data, employee personal information, financial data or regulated information into AI tools. Any exception should rest on an approved tool, an appropriate contractual arrangement and a decision by the person designated in the policy, with legal review where regulations such as HIPAA apply.

How often should an AI usage policy be updated?

At least annually, according to TechTarget’s guidance, which adds that quarterly reviews are advisable given how quickly AI changes. Many organizations also schedule an out-of-cycle review when a new tool category, regulation or incident arises.

Should the policy require employees to disclose AI use?

That depends on the employer’s and clients’ expectations, and the policy should say so explicitly. OpenAI’s guidance advises employees to keep conversation links or logs where disclosure is expected, and the NCSEA program asks for documentation of material AI use in consequential work.

Who should own the policy?

A single named role should own it, commonly within IT, information security or a cross-functional governance group. Georgia Tech assigns annual review to its AI governance program, which illustrates the value of a standing body with defined responsibility for the policy and the tool inventory.

Sources

• The Employer’s Guide to AI Workplace Policies, American Employers Alliance
• How to create an AI acceptable use policy, plus template, TechTarget
• Acceptable Use of Generative AI Tools, Fisher & Phillips
• Responsible use of ChatGPT at work, OpenAI Academy
• Guidelines for AI Tools in the Workplace, University of California ANR
• Generative Artificial Intelligence Program, NCSEA
• Generative AI Use in the Workplace Policy, Goosmann Law Firm
• Artificial Intelligence in Administrative and Employment Contexts, Georgia Tech Policy Library
• AI RMF Playbook, NIST
• AI Risk Management Framework, NIST

Recommended

• Audit Ready AI Risk Assessment in 6 Steps for Teams
• Start with MUR: Identity Governance Basics for Enterprise Leaders
• DevOps for Compliance: SBOMs & Signed Artifacts for U.S. Regulated Orgs
• Audit Ready Compliance Gap Analysis: One Row Register, Named Owners