Acceptable use policy for AI at work
A full AI acceptable use policy you can fill in today: approved tools, which data each tool may receive, banned uses, disclosure, output review and new-tool requests.
Most small companies already have staff using AI tools at work. Some use a company account, some use a personal one, and nobody has written down which data is allowed where. This template gives you a complete policy to fill in, plus notes on the decisions that take the most thought.
It is written for an owner, ops lead or IT lead at a company of roughly 10 to 500 people. It assumes no dedicated security team. Treat it as a starting point and review the final version with counsel before you issue it, especially the monitoring and disciplinary sections, which depend on where your staff work.
Before you fill it in
Do three things first, because the policy is only as good as the list in Section 3.
- Find out what people use now. Check expense reports and card statements for AI subscriptions, ask each team lead for the tools their people open daily, and look at which AI features are switched on inside software you already pay for (email, docs, CRM, help desk). Expect to find more than you thought.
- Decide your data classes. The template uses four: Public, Internal, Confidential and Restricted. If you already have a data classification scheme, use yours and change the names in Section 4.
- Name one owner. One person approves tools, answers questions and keeps the list current. Without an owner the approved list goes stale within a quarter and people stop checking it.
The policy
Copy everything in the block below into your handbook or document system and replace each [bracketed] field.
[COMPANY NAME]
ACCEPTABLE USE POLICY FOR AI TOOLS
Version: [1.0] Effective: [DATE] Next review: [DATE + 6 MONTHS]
Policy owner: [NAME, ROLE] Questions: [EMAIL OR CHAT CHANNEL]
1. PURPOSE AND SCOPE
This policy sets the rules for using AI tools in [COMPANY] work. It covers
chat assistants, coding assistants, meeting note-takers, writing and image
tools, AI features inside software we already use, and AI agents that can
act in company systems. It applies to employees, contractors and interns,
on company or personal devices, whenever the work or the data belongs to
[COMPANY] or its customers.
2. DEFINITIONS
AI tool: software that produces text, code, images, audio, summaries or
decisions from a prompt or from data you give it.
Approved tool: a tool listed in Section 3, used through the account type
listed there.
Company account: an account provisioned by [COMPANY], signed in with
company single sign-on (SSO) where the tool supports it.
Output: anything an AI tool produces.
AI agent: an AI tool that can take actions in other systems (send email,
edit records, run code, make purchases) rather than only reply to you.
3. APPROVED TOOLS
Only the tools and account types below may be used with company data.
A free or personal account of an approved tool is NOT approved.
| Tool | Account type | Approved for | Highest data class allowed | Owner |
|-------------------|---------------------|----------------------------------|----------------------------|--------------|
| [Chat assistant] | [Business plan, SSO]| Drafting, summarising, analysis | Confidential | [IT lead] |
| [Coding assistant]| [Company org seat] | Code in company repositories | Confidential (no secrets) | [Eng lead] |
| [Meeting notes] | [Company workspace] | Internal and customer meetings | Confidential, with notice | [Ops lead] |
| [Image tool] | [Company account] | Marketing drafts | Public | [Mktg lead] |
| [AI in CRM/help desk] | [Existing licence] | Features listed at [LINK] | Confidential | [Ops lead] |
The current list lives at [LINK]. If a tool is not on it, it is not approved.
4. DATA CLASSES
Public: already published or meant for anyone. Website copy, published
prices, public job ads.
Internal: everyday company information not meant for outsiders. Internal
process docs, meeting notes without customer details, draft plans.
Confidential: customer personal data, customer files, contracts, pricing
agreements, financial records, source code, employee records.
Restricted: passwords, API keys and other credentials; payment card
numbers; bank account numbers; government ID numbers; health
information; anything a contract or law says must not leave a named
system.
Rules:
a) Restricted data must never be entered into any AI tool.
b) Confidential data may be entered only into tools whose row in Section 3
allows Confidential.
c) If you are unsure of the class, treat it as the higher one and ask
[POLICY OWNER].
5. BANNED USES
You must not:
a) Enter Restricted data into any AI tool, or Confidential data into a tool
not approved for it.
b) Use a free or personal AI account for company or customer data.
c) Connect an AI tool or agent to company email, files, calendars, CRM,
code repositories, finance or HR systems without approval under
Section 8.
d) Let AI output make a final decision about hiring, firing, pay,
promotion, discipline, credit or pricing for a named person. A named
human makes these decisions and checks the facts behind them.
e) Create content that impersonates a real person, or produce fake
reviews, testimonials, endorsements or customer quotes.
f) Record or transcribe a meeting with an AI tool without telling the
participants first.
g) Use an AI tool to get around a security control, licence term or
access restriction.
h) Send AI output to a customer while presenting it as checked by a person
when nobody has checked it.
6. DISCLOSURE
a) Internal: tell your reviewer or manager when a deliverable was mostly
produced by AI. Routine drafting help needs no label.
b) Customers: follow [CUSTOMER DISCLOSURE GUIDELINES, LINK]. Automated
chat or email agents identify themselves as automated. Meeting
note-takers are announced at the start of the call.
c) Contracts: if a customer contract limits AI use on their work, the
contract wins. Account owners keep a list of these customers at [LINK].
7. REVIEWING OUTPUTS
a) You are responsible for any output you use, as if you wrote it.
b) Before output leaves your hands, check names, numbers, dates, quotes,
links and citations against the source. Open every source an AI tool
cites before you repeat it.
c) Code produced with AI goes through the same review and tests as any
other code.
d) Customer-facing, legal, financial and HR documents produced with AI get
a second-person review under [EXISTING REVIEW PROCESS].
8. REQUESTING A NEW TOOL OR INTEGRATION
a) Submit a request at [FORM LINK] with: tool name and website, what you
want to use it for, the highest data class it will receive, number of
users, cost, and whether it connects to other company systems.
b) [POLICY OWNER] reviews it against the AI tool approval checklist and
replies within [10] business days.
c) Until approved, you may trial the tool with Public data only.
d) AI agents that act in company systems also need a named owner, their
own login, and the access controls in [AGENT ACCESS STANDARD, LINK].
9. INCIDENTS
Report any of the following to [CHANNEL/EMAIL] within [1 HOUR] of noticing:
- Confidential or Restricted data entered into an unapproved tool or account.
- AI output with a material error sent to a customer or published.
- An AI agent taking an action outside what it was approved to do.
Prompt reporting will not, by itself, lead to disciplinary action.
10. MONITORING
[COMPANY] may review usage and admin logs from company AI accounts to
enforce this policy and investigate incidents, as allowed by law.
11. BREACHES OF THIS POLICY
Breaches are handled under [DISCIPLINARY POLICY, LINK].
ACKNOWLEDGEMENT
I have read this policy and will follow it.
Name: ____________ Role: ____________ Signature: ____________ Date: ______
Notes on adapting
Section 3 is the policy. Everything else points back to it. Name the account type in every row, not only the product. The same assistant can be fine on a business plan with SSO and training turned off, and unacceptable on a personal free account. Staff need to see that difference written down.
The "highest data class" column is where you will argue. Set it from what the vendor commits to in writing (retention period, whether your data trains their models, where it is stored) and not from the product's reputation. Run each tool through a fixed approval checklist so the decision is the same for every tool.
Keep banned uses specific. "Don't misuse AI" gives nobody anything to follow. Each item in Section 5 describes something a person could do on a Tuesday afternoon without thinking. Add items from your own sector: a clinic might ban entering any patient detail into chat tools, a law firm might ban AI drafting of advice letters without partner review.
Decide what "mostly produced by AI" means for your team. A workable rule: if you would be uncomfortable with your manager learning later that AI wrote it, say so now.
Contractors and agencies. If agencies produce work for you, add a line to their contracts requiring them to follow Sections 4, 5 and 7 when they handle your data. Your policy does not bind them otherwise.
Regulated data. If you handle health records, payment cards or data under a sector rule, your existing compliance obligations already restrict where that data can go. The Restricted class is meant to keep it out of AI tools entirely until your compliance owner approves a specific tool and setup.
Worked example: Pinegate Supply
Pinegate Supply is a fictional 45-person distributor. The ops lead found eleven AI tools in card statements, six of them personal accounts expensed by staff. She approved three on company accounts: a chat assistant on a business plan with SSO, the AI features already inside their help desk, and a meeting note-taker for sales calls. Everything else went to "not approved, request it through Section 8."
Two staff requested an image tool for product photos. She approved it at Public only, since product photos are published anyway. A request for an AI bookkeeping assistant that connects to the accounting system went through the agent access standard and was approved read-only for one quarter.
She issued the policy with a 15-minute walkthrough at the monthly all-hands and collected signed acknowledgements through the HR system. The first review six months later added two tools and moved one from Internal to Confidential after the vendor signed a data processing agreement.
Rolling it out
- Publish the policy and the approved tool list at stable links.
- Walk through Sections 3, 4 and 5 at a team meeting. Those are the parts people need to remember; the rest they can look up.
- Collect acknowledgements within two weeks.
- Cancel or transfer personal subscriptions that are now replaced by company accounts, so people have no reason to keep using them.
- Put the review date in the owner's calendar.
Do it with Merit AC: the AI Skills Academy at /academy/course trains staff to use approved tools well, which makes the rules easier to follow.
This piece is free. 35 more like it, and each new one, come with AI Skills Academy membership.