AI acceptable use policy template for your business

This is an AI acceptable use policy you can adopt as it stands: 15 numbered clauses, four annexes, a Word file and a PDF. There’s no email address to give and no watermark. I wrote it for organisations that use AI tools made by someone else, not for teams that build models.

The clauses don’t lean on one country’s law. Where a rule depends on local legislation, the text says “the privacy law applying to [Organisation]” and leaves that part to you. If the worry behind your search is a specific paste, whether client data in ChatGPT counts as a breach is the companion piece.

AI policy template: how to use it, and what it does not cover

Read it end to end once before you edit a word. Then fill in Annex A, because the never-paste list for your own data is the one part nobody outside your organisation can write. Clause 3 comes next: which tools are approved, and who approves the next one.

The template sets rules for how people behave. It isn’t a data protection impact assessment, it doesn’t tell you whether a particular tool is lawful for you, and it says nothing about building models. Have whoever advises you on employment and data law read your edited version before you adopt it.

Nor will it stop a paste. A signed policy lets you respond once a rule is broken, no more. Anyone who says a document keeps data from leaving the building is passing paper off as a control. I prefer a first version with one named owner and no clause you won’t enforce.

The 15 clauses of the AI acceptable use policy

Each clause is short on purpose: a rule staff cannot recite is one they will not follow. Below each clause is a line on why it’s there, which stays on this page and isn’t in the download. I’d resist adding clauses before the first review: every extra line is one more thing people forget.

1. Purpose and scope

This policy governs the use of artificial intelligence tools in the course of work for [Organisation]. It applies to all employees, contractors, temporary staff, volunteers and directors, on any device including their own, and whether or not [Organisation] provided the tool.

It covers tools reached through a browser, an application, a plugin, an extension, a feature embedded in another product, or an API. Where it conflicts with a client engagement term or a professional obligation, the stricter requirement applies.

2. Definitions

AI tool means any system that generates, summarises, classifies, transcribes or predicts using a model, including features embedded in products you already use. Input means anything a user types, pastes, uploads, dictates or connects to one. Output means what it returns.

Client information means any information relating to a client’s affairs, whether or not the client provided it. Personal information has the meaning given by the privacy law applying to [Organisation].

3. Approved tools, and how a tool gets approved

Only tools on the approved list in Annex B may be used for work, and only with input that clause 5 allows. The list names the tool, the plan or tier, the account type and the approver.

A tool is added only after a named person has checked, in writing: what the vendor does with input, whether input trains models and whether that can be switched off, where data is stored, how long it is kept, and whether the contract is with [Organisation] rather than an individual.

Free and personal tiers are not approved.

4. What must never go into an AI tool

The data types listed in Annex A must never be entered into any AI tool, approved or not, in any form, including inside a screenshot, a pasted table, an uploaded file or a connected mailbox. This prohibition has no approval path.

If a task seems to require it, stop and escalate to the person named in clause 15.

5. What may go in, and on what condition

Input not listed in Annex A may be entered into an approved tool once it has been cut to the minimum the task needs and direct identifiers removed or replaced.

Removing a name is not removing an identity: a matter number, a rare job title, or a date plus a location can identify a person on its own. Replacing an identifier with a code does not take the information outside privacy law for [Organisation], which holds the mapping.

Keep the mapping between codes and real values outside the AI tool.

6. Client confidentiality is a separate duty

Confidentiality obligations apply to AI tools as to any other third party. Entering client information into a tool run by another organisation is a disclosure to a third party.

Where an engagement letter, retainer, professional rule or contract requires client permission before such a disclosure, that permission is obtained before the information is entered, and recorded.

7. Personal information and privacy law

Personal information may be entered into an approved tool only where that is consistent with the purpose it was collected for and permitted by the privacy law applying to [Organisation]. [Organisation] remains responsible for it after it enters a tool, including for its security.

Where the regulator recommends against a category of input, that recommendation is followed unless a named person records a reason not to.

8. Consent, engagement letters and notices

Where clause 6 requires client permission, it is obtained in the engagement letter, a signed consent or an equivalent written record, and states what will be disclosed, to whom, where the data will be stored, and that AI tools may be used.

Privacy notices and the privacy policy describe the use of AI tools in terms an individual can understand.

9. Human review before reliance

Output is a draft. No output may be sent to a client, filed, published, relied on in advice, or used to decide something about a person until a competent human has checked it against a source. The reviewer is accountable for the content as if they had written it.

Citations, figures, quotations, legislative references and calculations are checked one by one against the primary source, never against another AI tool.

10. Disclosure and marking of AI assisted work

Where content generated or materially altered by an AI tool is published, given to a client, or filed with a court, tribunal or regulator, [Organisation] discloses that fact where a rule, contract or professional obligation requires it, and wherever a reader would otherwise be misled.

Staff do not remove or defeat any marking or metadata a tool attaches to its output.

11. Records: the use case register

[Organisation] maintains the register in Annex D. Each approved use case records the tool, the purpose, the categories of input, who approved it, the date, and the next review. Prompts and outputs on client matters are kept in the matter file under normal record keeping rules, not only inside the tool’s own history.

12. Accounts, plans and vendor terms

Work input is entered only through accounts held in [Organisation]‘s name, on the plan named in the approved list. Personal accounts, personal email addresses and consumer tiers are not used for work input.

Where a tool offers a setting controlling whether input trains models, or how long it is retained, the setting is recorded in the approved list and checked at each review.

13. Devices, browsers and extensions

Browser extensions, plugins, desktop assistants and integrations that can read a page, a document or a mailbox are installed only from the approved list. Staff do not connect an AI tool to a mailbox, a document store, a practice management system or a code repository without approval under clause 3.

14. Incidents

If prohibited data has been entered into an AI tool, or anything into an unapproved one, the user reports it to the person named in clause 15 the same day. Nobody is disciplined for reporting promptly.

[Organisation] records what was entered, when, into which tool and under which account; asks the vendor to delete it and records the answer; and assesses whether the incident is notifiable. Under Australia’s Notifiable Data Breaches scheme, the test includes whether a reasonable person would conclude it would be likely to result in serious harm.

15. Owner, review cycle and changes

This policy is owned by [named role]. It is reviewed at least every twelve months, and whenever a tool is added to the approved list, a regulator publishes guidance that changes a clause, or an incident under clause 14 exposes a gap. Every version carries a date and a change log.

Staff are notified of material changes and acknowledge again using Annex C.

Annex A: data that must never be entered

The two filled columns use Australia and Spain as worked examples. Swap them for the identifiers your own country issues, and add any sector numbers your staff handle, such as patient or policy numbers.

I’m wary of an Annex A longer than one screen: people remember a short list and skim a long one. If you’re unsure where client data hides beyond the obvious numbers, what counts as client data once names are gone covers matter details and combinations.

Annex B: the one page sheet

Print it and put it where people work.

This is the part most staff will actually read, so it carries no clause numbers. If it were my firm, this sheet would sit beside every printer while the full policy stayed on the intranet. Before you print, compare the right-hand column with our fuller list of what to redact in a document.

Approved tools, plans and accounts: _______________________ (this list is clause 3; keep it current).

Annex C: acknowledgement

A signature doesn’t make anyone careful, but it fixes which version each person saw. Collect it at onboarding and again whenever clause 15 brings a material change.

I wouldn’t accept a reply-all “noted” as the acknowledgement. Insurers and clients increasingly ask whether staff were told the rules, and the AI section of an insurance questionnaire is easier to answer with a dated signature per person.

Annex D: use case register

One row per approved use: its purpose, which tool on which plan, the categories of input allowed, the approver, and the date it comes up for review. Clause 11 keeps it current.

For a five-person team, a shared spreadsheet with these six columns is enough. The column I care most about is the review date: an overdue row is a tool nobody has rechecked. Larger organisations sometimes route AI traffic through a gateway that logs every call, which shows which uses are missing from the register.

Which AI tools to approve: Copilot, Gemini and Claude business terms

Clause 3 asks a named person to check what the vendor does with input before any tool is added. For the large assistants, the answer depends far more on the plan than on the product name.

An entry that names a product but no plan is one I’d send back, because the plan and its settings decide the answer. Microsoft’s commitments, for instance, apply to work accounts, which is why whether Copilot trains on your data depends on how people sign in.

ChatGPT has the same split between consumer and business plans; how OpenAI treats business data sets out which settings to record under clause 12.

Does the EU AI Act require an AI policy? Article 4 on AI literacy

Not by name. Article 4 has applied since 2 February 2025 and originally asked providers and deployers to ensure, to their best extent, “a sufficient level of AI literacy” among staff using AI on their behalf. Regulation (EU) 2026/1744, adopted on 8 July 2026, replaced that wording.

The new text asks them to “take measures to support the development of AI literacy”, without guaranteeing “any specific level” for anyone. That seems more honest to me: no employer can promise what an employee knows. Whether the Act reaches you at all depends on where your AI output ends up.

The European Commission’s question and answer page says no certificate is needed; an internal record of training or other guidance can serve instead. Its own example is staff using ChatGPT to draft advertising text, who should be told about risks such as hallucination. Supervision and enforcement rules apply from 2 August 2026.

NIST AI RMF or a company AI policy: which do you need first?

Where does the NIST AI Risk Management Framework fit in? NIST released it on 26 January 2023, says it is intended for voluntary use, and is revising version 1.0. It helps you manage AI risk across an organisation; it isn’t a set of rules for staff to sign.

So the two answer different people. A framework guides whoever governs AI use and how decisions are escalated; an acceptable use policy tells the person at the keyboard what they may paste. For a firm under fifty people, I’d get the policy signed before opening any framework.

If you’re in the US, our American version maps these clauses to NIST’s functions and adds the state rules that change a line. Federal law already reaches AI use: announcing Operation AI Comply in September 2024, the FTC’s then chair, Lina Khan, said there is no AI exemption from the laws on the books.

Rolling out your AI usage policy: where paper stops

Clause 4 carries the rest, and a busy person breaks it without noticing: a client’s card number inside a pasted email thread, an identity number in a screenshot. A policy deals with that afterwards; it can’t step in. The Word and PDF files wait in the download box above the clauses.

That gap is why I’d pair Annex A with something that catches the identifier at the moment of pasting, and why the cheapest controls to add before the next paste go beyond writing things down.

Nonimo’s signed apps for Mac and Windows run that check on the computer, and with a browser extension installed by policy they also read what’s pasted into the AI tool. Our security page explains why your text never reaches us.

To try it on a prompt of your own, start from the Nonimo home page.

Sources

• Regulation (EU) 2024/1689, the AI Act: Article 4 as first adopted, Article 50(2) and 50(4), and the dates in Article 113.
• Regulation (EU) 2026/1744 of 8 July 2026, in the Official Journal of 24 July 2026 on EUR-Lex: the new Article 4 and the 2 December 2026 deadline it adds as Article 111(4).
• European Commission, AI literacy questions and answers: no certificate needed, internal records, supervision from 2 August 2026.
• NIST, AI Risk Management Framework: released 26 January 2023, voluntary, version 1.0 under revision.
• FTC, crackdown on deceptive AI claims and schemes: Operation AI Comply, 25 September 2024; “no AI exemption from the laws on the books”.
• Microsoft Learn, data, privacy and security for Microsoft Copilot: prompts, responses and Graph data kept out of foundation model training; retention set by admins in Purview.
• Google, Generative AI in Google Workspace Privacy Hub: customer data not used for training without prior permission or instruction.
• Anthropic Privacy Center, commercial products: inputs and outputs not used for training by default.
• Anthropic Privacy Center, consumer plans: the Model Improvement setting.

Accessed 10 November 2026.