The Complete AI Acceptable Use Policy Guide (2026)

Define who the policy covers, what it covers, and why it exists. Scope ambiguity is the most common reason a policy fails to bind contractors, subsidiaries, or AI features inside existing tools.

"This policy governs the use of all artificial intelligence and generative AI tools - including but not limited to large language models, AI assistants, AI features embedded in approved software, and AI coding tools - by all employees, contractors, consultants, and third parties acting on behalf of [Company]. It applies to use on any device, network, or account, whether company-provided or personal, when used for [Company] business. Its purpose is to enable safe, productive, and compliant use of AI while protecting [Company] data, customers, and legal obligations."

List the AI tools employees may use and the conditions attached. A tiered approach - approved for general use, approved with restrictions, prohibited - is clearer than a binary list and easier to maintain.

"Only AI tools on the [Company] Approved AI Tools Register may be used for business purposes. Tools are classified into three tiers: (a) Approved - may be used for the data classes listed in the register; (b) Restricted - may be used only for the specified use cases and data classes, subject to the controls noted; (c) Prohibited - must not be used for any business purpose. The register is maintained by [the AI Governance function] and reviewed [quarterly]. Use of any AI tool not on the register requires prior approval under Section 4. Personal consumer AI accounts must not be used for [Company] business."

Pair the register with technical enforcement. A policy that lists approved tools but cannot detect or block unapproved ones is aspirational; our review of ChatGPT safety shows why unapproved-tool blocking is the control that actually changes behaviour.

This is the most important operational section. State plainly what data must never be entered into AI tools, and what may be entered only into which tier. Use your existing data classification scheme so the rules are unambiguous.

"The following data classes must never be entered into any AI tool except a [Company]-approved private deployment: personal information and sensitive personal information; protected health information; payment card and financial account data; authentication credentials, API keys, and secrets; source code and proprietary algorithms; legally privileged material; trade secrets; and any data classified [Confidential] or above. Data classified [Internal] may be entered only into Approved-tier tools that operate under [Company]'s no-training agreement. Public data may be entered into any Approved-tier tool. When in doubt, do not enter the data and consult [the AI Governance function]."

Tie each prohibited class to your data classification framework so the policy and the classification scheme stay consistent as both evolve.

Define how someone requests a new tool or a new high-risk use case, who decides, and how fast. A workflow that is too slow drives shadow AI; one with no gate drives uncontrolled risk. Name the owner and the service-level target.

"To request approval for a new AI tool or a new use case involving Restricted data, submit a request to [the AI Governance function] including the tool, the intended use case, the data classes involved, and the business justification. Requests are triaged within [five business days]. Low-risk requests may be approved by [the AI Governance lead]; requests involving Restricted data or high-risk use cases require review by [the AI Governance committee] and, where personal data is involved, a privacy impact assessment. Approved tools and use cases are added to the register; rejected requests receive a documented rationale and, where possible, an approved alternative."

Tell employees what is monitored and why. Transparent monitoring is both an ethical obligation and, in many jurisdictions, a legal one; it also improves voluntary compliance. Hidden monitoring erodes trust and can be unlawful.

"To protect [Company] data and meet regulatory obligations, [Company] monitors the use of AI tools on company systems and networks. Monitoring may include logging of AI tool access, inspection of prompts and responses by data-loss-prevention systems for sensitive data, and review of usage analytics. Monitoring is conducted for security, compliance, and policy-enforcement purposes only, in accordance with applicable law and [Company]'s [Privacy Notice]. Employees should have no expectation of privacy in business use of AI tools on company systems."

Align the disclosure with your privacy notice and local employment and surveillance law, which varies by jurisdiction.

Make it easy and blameless to report a mistake. Most AI data exposure is accidental; the difference between a near-miss and a breach is often whether someone reported it quickly. State the channel and the response commitment.

"If you believe sensitive data has been entered into an unapproved AI tool, or into an approved tool in breach of this policy, report it immediately to [security@company] or via [the incident channel]. Prompt good-faith reporting of an AI data-exposure incident will not result in disciplinary action against the reporter, even where the exposure resulted from their own error. [Company] will assess each report, contain the exposure, and take remediation steps. Failure to report a known exposure is itself a breach of this policy."

Connect this to your broader incident response plan so AI exposures route into an existing, tested process rather than an ad hoc one.

Set the bar a tool must clear before it joins the register. AI vendors differ sharply on training, retention, residency, and sub-processing; this section makes those questions mandatory rather than optional.

"Before any AI tool is added to the Approved AI Tools Register, [the AI Governance function] must assess the vendor against [Company]'s AI Vendor Risk standard, including: whether the vendor trains on customer data by default; data retention and deletion terms; data residency and sub-processor locations; security certifications (such as SOC 2 or ISO 27001); the availability of a data processing agreement; and contractual breach-notification commitments. Tools that train on input data by default, or that cannot meet [Company]'s data residency requirements, must not be approved for use with Restricted data."

Require training and tie it to access. The EU AI Act explicitly obliges organisations to ensure staff AI literacy, and training is the cheapest control with the highest behavioural return.

"All employees who use AI tools for business purposes must complete [Company]'s AI literacy and acceptable use training before being granted access, and must complete refresher training [annually]. Training covers approved tools, prohibited data, how to recognise and report incidents, and the risks of inaccurate or biased AI output. Completion is tracked by [the AI Governance function], and access to approved AI tools may be suspended for employees who do not complete required training."

State the consequences of breach proportionately. Enforcement language gives the policy teeth, but it must be paired with the blameless reporting in Section 6 so people report mistakes rather than hide them.

"Breaches of this policy will be handled in accordance with [Company]'s disciplinary procedures, proportionate to the severity and intent of the breach. Inadvertent breaches that are promptly self-reported will be treated as learning and remediation opportunities. Deliberate breaches - such as knowingly entering prohibited data into an unapproved tool, circumventing technical controls, or failing to report a known exposure - may result in disciplinary action up to and including termination, and may be reported to regulators or authorities where required by law."

Provide a controlled path for legitimate exceptions so people do not route around the policy entirely. An exception process with an owner and an expiry is far safer than a rigid policy that drives covert workarounds.

"Exceptions to this policy may be granted only by [the AI Governance committee] on documented business justification, for a defined period not exceeding [six months], with compensating controls specified. All exceptions are recorded in the exceptions register with the approver, rationale, controls, and expiry date, and are reviewed at expiry. An exception does not set a precedent and may be revoked if the underlying risk changes."

Set how often the policy is reviewed and who owns it. AI moves quickly; a policy reviewed annually will lag the tools and regulations it governs. Version control prevents stale copies circulating.

"This policy is owned by [the AI Governance function] and reviewed at least [every six months], and additionally whenever a material change occurs in AI tools in use, applicable regulation, or [Company]'s risk profile. Each version carries a version number, effective date, and summary of changes. The current approved version is published at [location]; all prior versions are archived. Material changes are communicated to all affected staff."

Define the key terms so the rest of the policy is unambiguous. Disputes about whether a tool or a data class is in scope usually trace back to undefined terms.

"For the purposes of this policy: 'AI tool' means any software or service that uses artificial intelligence or machine learning to generate, analyse, or transform content, including generative AI and embedded AI features. 'Generative AI' means AI that produces new content such as text, code, images, audio, or video. 'Prompt' means any input submitted to an AI tool. 'Restricted data' means the data classes listed in Section 3. 'Approved AI Tools Register' means the maintained list of permitted tools and their conditions. 'AI Governance function' means [the named team or role accountable for this policy]."

With these 12 sections in place, you have a complete, defensible AI AUP. The fastest way to assemble them is to start from the free downloadable template, which provides all 12 as editable provisions, or to generate a first draft tailored to your profile with the AI policy generator.