An AI acceptable use policy has to answer three questions clearly enough that someone can act without asking: which AI tools are approved, what information can go into them, and who to ask about anything not covered. Everything else is detail around those three answers.
You can download a template in two minutes. That isn’t the hard part. Open any of them and you’ll find blanks — those blanks are the policy. The rest is formatting.
First, find out what's in use
A policy written without knowing what your team already uses is a guess. Ask them which AI tools they use for work and say plainly there’s no penalty for answering honestly. The method, and why the list looks the way it does, is in Shadow AI at work.
The decisions only you can make
Six, and none are technical:
- Which tools and accounts are approved, and what each is approved for.
- What information is allowed into them, and what has to come out first.
- Who answers “can I use this?”, and how quickly.
- Who checks AI-assisted work before it is used or sent to a client.
- Who owns the policy and keeps it current.
- Who to tell when something goes wrong.
An IT provider can advise on all six and implement the result. If we decide them for you, the policy stops being yours.
The sections a working policy needs
1. Why it exists, and what counts as “AI”
Open with one short paragraph: we want you using these tools, here’s what we have to protect, here’s how to ask. People follow rules they understand the reason for — lead with a threat and it gets skimmed and filed.
Define scope broadly, or staff will assume the policy only covers the chat window. Name web-based tools, AI features inside software you already pay for, operating-system assistants, browser extensions and meeting recorders. Features switched on inside software you already bought are the category nobody thinks to ask about.
2. Approved tools, and how to ask for a new one
Keep the approved-tools list in an appendix so it can change without re-signing the policy. Record what each tool is approved for, plus limits on tasks, information or connected accounts — drafting public marketing copy is a different approval from summarising internal documents.
Then give people a route to ask: name the person, say what to include, give a realistic turnaround. A request process nobody uses is the same as none — the tool gets used anyway, quietly.
3. What information can go in
Three categories:
- Public — already published. An approved tool for an allowed task is fine.
- Internal — non-confidential business information, anonymised first. Spell out what “anonymised” means using your own examples — removing a name often isn’t enough, because other details in the file can still identify someone.
- Restricted — personal data, credentials, financial details, anything that would cause harm if it leaked. Stays out.
Two warnings worth writing down. Anonymising doesn’t automatically make a document suitable — the tool, the task and what’s left in the file all still have to pass. And whether a tool suits a class of information depends on the tool, the plan and the settings, so check the vendor’s current terms for the plan you’d actually be on.
4. Checking the output
AI can produce work that sounds right and isn’t. Four checks before anything is used or sent: facts, verified against original sources or your own records — including any source the tool cites; confidential details the recipient shouldn’t see; context, because accurate general information can still be wrong for this decision; and client permissions, since an approved tool doesn’t override a client’s contract. Say who is responsible for each, and when a second person reviews.
5. The specific traps worth naming
Short, direct clauses beat general principles:
- Meeting recorders — whether bots may join, who may invite one, what to do about an uninvited one. Recording-consent law varies by state, so it’s worth getting right, and worth confirming with your own counsel rather than from a blog post.
- Browser extensions — these generally need read access to whatever page is open, mailboxes and portals included. Deploy centrally or say no.
- Automations — nobody wires company data into an AI service through an automation platform or API without review. For mailbox connections, start with should you connect AI to your email?
- Impersonation and payments — no cloning anyone’s voice or likeness without written approval, and any urgent request for a transfer or credential change gets verified on a known number.
6. The clauses that need HR, legal or compliance eyes
Three need whoever handles HR or legal for you to look before staff see them: monitoring of company AI accounts, AI apps on personal devices, and what happens to AI history when someone leaves. None are IT defaults to set quietly.
If you operate under specific regulation — patient information, client funds, defence work — route that to your compliance officer or counsel instead of settling it in the policy. The policy’s job is to require a documented check and name who performs it.
7. Reporting mistakes, then reviewing the whole thing
Reporting is the most important section and usually the shortest. Someone will paste the wrong thing into the wrong window, and what happens next depends on whether your policy makes self-reporting safe. Name a real person and a real contact method; a placeholder nobody can reach is worse than nothing.
Set a review interval; six months is reasonable given how fast the tools move. Have people acknowledge the policy, but only after a chance to ask questions.
What makes the difference
Keep the rules to one page, use your own work as the examples, and train it once alongside the rest of your security awareness training. Written rules are cheap air cover for your staff — nobody should be in trouble for asking first. If you want the ongoing version, with the approved-tool list enforced in your systems and reviewed on a cadence, that’s AI governance and AI use policy services.
Frequently Asked Questions
How long should an AI acceptable use policy be?
The part your team acts on should fit on one page: approved tools, what information can go in, who to ask, how to report a problem. Definitions, scope and the tools appendix can run longer — people read those once, not weekly.
Can we just use a downloaded AI policy template?
Use one as a starting point, not a finished document. A template gives you the section list, but it arrives with the wrong company name, example tools you haven’t reviewed, and contact placeholders nobody can call — which is why borrowed policies sit unread in folders.
Who should own the AI policy?
Someone in the business, not your IT provider — usually the owner, an operations lead, or whoever owns HR policy. The decisions are business decisions, and staff need to put a face to the rules. Your IT provider’s job is to advise, implement the settings, and keep the approved-tools list accurate.
Does an AI policy make us compliant with our regulations?
No, and no policy can promise that. A written policy, an approved-tools list and the right access settings are controls that support a compliance position — but whether your use of a specific tool meets a specific regulation is a determination for your compliance officer or counsel, including whether an agreement such as a Business Associate Agreement is in place for that tool and plan. Write the policy so it requires that check and names who makes it.
What if we don't want to allow AI at all?
A legitimate position, and some regulated work calls for it — but write it down and explain why, because an unwritten ban is just an assumption. Give people a route to ask for exceptions too: prohibitions without one push the work onto personal accounts, leaving the same exposure with less visibility.
Want a template with the blanks explained?
We run a free 30-minute Shadow AI Audit call for businesses in Los Angeles, Burbank, Nashville and Franklin: current AI use, the safety points worth covering with your team, and a policy template to adapt — with the decision behind each blank talked through. No obligation, no charge.
Book the 30-minute Shadow AI Audit · or call (818) 284-4118.
Or call (818) 284-4118.
