What to Put in an AI Acceptable Use Policy
The sections a short, enforceable AI use policy needs: approved tools, data rules, human review, disclosure, agent permissions and how to ask for something new.
Answer first
An AI acceptable use policy tells staff which AI tools they may use, with which data, for what, and what they must check before relying on the output. A good one fits on two pages, names a fast way to request new tools, and is backed by technical controls where the risk is high. A policy that only says "do not put confidential data into AI" without naming approved tools pushes people toward unapproved ones.
The sections to include
1. Purpose and scope
State who the policy covers (employees, contractors, agents acting for them) and what counts as AI use: chat assistants, AI features inside existing software, coding assistants, and agents that act on systems.
2. Approved tools
List approved tools by name and the account type required (company workspace accounts, not personal ones). Link to the live list rather than embedding it, because it will change often.
3. Data rules by class
Map your existing data classification to AI use. For example:
- Public: any approved tool.
- Internal: approved tools with company accounts and data retention settings confirmed.
- Confidential or personal data: only tools approved for that class, with a recorded reason.
- Restricted (secrets, credentials, regulated data): never, unless a specific system is approved for it.
The NIST Generative AI Profile lists data privacy and information security among the risks specific to or heightened by generative AI; tie your rules to your own classification rather than inventing a new one.
4. Human review and accountability
The person who uses AI output is accountable for it. Require review in proportion to impact: light for internal drafts, careful for anything sent to customers, published, used in a decision about a person, or committed as code. Say plainly that AI output can be wrong, including confident and well-formatted errors; the OWASP list for LLM applications calls this risk "Misinformation" (OWASP Top 10 for LLM Applications 2025).
5. Prohibited uses
Keep this short and specific to your business. Typical entries: decisions about employment, credit or access to services without the approved process; impersonating people; generating content that breaches law or company policy; bypassing security controls. If you operate in the EU, align this list with the EU AI Act prohibited practices: eight have applied since 2 February 2025, and a ninth, on AI that generates non-consensual sexual content or child sexual abuse material, applies from December 2026.
6. Disclosure
Say when staff must tell recipients that content or an interaction involves AI, for example customer-facing chat or synthetic images. Check the transparency obligations that apply in your markets.
7. Agents and automation
Agents need rules people do not: who may connect an agent to company systems, which permissions it may hold, which actions need approval, and who owns it. Point to the technical standard that enforces this, such as access control patterns for enterprise agents.
8. Requesting a new tool or use
Give a single intake form and a published turnaround target. A slow or opaque request path is a common cause of shadow AI.
9. Reporting problems
Tell people how to report a harmful output, a data leak or a suspected misuse, and promise no penalty for good-faith reports.
10. Ownership and review
Name the policy owner and a review date. Review at least yearly and when a major tool or regulation changes.
Step-by-step: from draft to working policy
- Collect the AI tools already in use from SSO, expense and browser-extension data, so the approved list reflects reality.
- Draft the ten sections above in plain language and keep the whole policy to about two pages.
- Map each data class to allowed tools and confirm the retention and training settings for each tool.
- Back the highest-risk rules with controls: data loss prevention for restricted data, blocked personal accounts, agent permission scopes.
- Publish the intake form and turnaround target at the same time as the policy.
- Train people with short examples from their own work, then measure.
How to know it is working
- Share of AI use on approved tools (from SSO and network data).
- Number of intake requests and median turnaround.
- Reported incidents and near misses, which should rise at first as people learn to report.
- Exceptions granted and their age.
Decision framing
The policy's value is that it lets people use AI with confidence instead of guessing. Measure success by approved-tool adoption and request turnaround rather than by how many uses were blocked.
Related reading
- AI acceptable use policy starter: a free tool that builds a Markdown draft from these ten sections
- Enterprise AI governance that unblocks delivery
- AI governance FAQ
- Procurement intake for models, agents and MCP connectors
- AI training and enablement services