AI acceptable use policy: what to allow, restrict and check

What is an AI acceptable use policy?

An AI acceptable use policy, often shortened to AI AUP, sets out the specific behaviour expected of anyone using AI tools for the company. It is narrower than a full AI policy. It does not explain who approves vendors or how the register is reviewed; it tells an employee, at the moment they open a tool, what they can and cannot do.

Many companies already have an IT acceptable use policy covering email, devices and the internet. An AI AUP can be a new section of that document or a separate page that links to it. What matters is that it speaks in tasks and examples, not abstractions. "Use AI responsibly" gives nobody a decision; "do not paste client contracts into any AI tool" does.

This page is general information, not legal advice. Have a qualified adviser review the version you adopt, particularly if you handle personal data or work under client confidentiality terms.

How do you decide which AI uses to allow?

Sort uses into three tiers based on the information involved and the consequence of an error: allowed, allowed with approval, and not allowed. A use is not risky because AI is involved; it is risky because of what goes in and what the output is used for. The same tool can be fine for rewriting a public blog post and unacceptable for summarising a disciplinary file.

Build the tiers from a list of real tasks. Ask each team lead for five things people already use AI for, or want to. Place each one, and the tiers will have concrete examples from the first day.

What do good acceptable use clauses look like?

Good clauses name the tool, the task and the information, and they say what to do instead. Below are example clauses for a fictional 40-person software company, Northgate Apps, that has a business workspace of a chat assistant and uses the AI features in its code host and office suite.

Each clause is short enough to remember and specific enough to act on. Notice that none of them says "be careful"; each one names a behaviour.

What information should an AI acceptable use policy restrict?

Restrict information by class rather than by listing every document type, and tie each class to the accounts where it is allowed. The usual classes are public, internal, confidential or client, personal information and secrets such as passwords and keys. Secrets are the easiest rule to write: never, in any tool.

The UK National Cyber Security Centre's guidance on large language models makes the general point that a query is visible to the organisation providing the model, and recommends not including sensitive information in queries to public models. A company business workspace may have different contractual terms, so record what your plan says. Microsoft's documentation, for example, says Copilot Chat used by organisations falls under its data protection addendum and that prompts and responses are not used to train foundation models, while noting that controls vary by subscription plan. Your AUP can then say which classes are allowed in that workspace and which are not allowed anywhere.

For UK businesses handling personal data, the ICO's guidance on AI and data protection is the official reference. Its pages currently say the guidance is under review following the Data (Use and Access) Act, so check what is live before relying on a specific point.

What should employees check before using AI output?

Employees should check every factual claim, figure, quote and citation that will leave their hands, and remove anything they cannot trace to a source. The NIST Generative AI Profile describes confabulation as the production of confidently stated but erroneous or false content; a fluent answer is not evidence. The AUP should make the person using the output responsible for it.

A worked example: a Northgate marketer asks the assistant for a short paragraph on why small teams adopt project tools. The draft includes a line saying a majority of small businesses switched tools last year, with a source name attached. The marketer cannot find that figure in the named source. Under the AUP clause on claims, the line comes out and the paragraph is rewritten to describe the reasons customers give in Northgate's own support tickets, without a number.

Weak version: "Most small teams switched project tools last year, according to a recent industry survey." Reviewed version: "Teams that contact us usually mention two reasons for switching: missed handovers and too many tools." The second is less dramatic and entirely defensible.

Should employees sign an AI acceptable use policy?

A signature or acknowledgement is reasonable, but it only shows that someone received the policy. Pair it with a ten-minute walkthrough of three cases from your own work, one per tier, and ask each person to place them and explain why. That gives you evidence that people can apply the rules, which is worth more than a signature if something goes wrong.

The European Commission's AI literacy Q&A, which concerns Article 4 of the EU AI Act, says certificates are not required and that an internal record of training or other initiatives is enough. If EU rules apply to you, a short record of who completed the walkthrough and when is a sensible habit. Take advice on whether and how the Act applies to your business.

How do you enforce an AI acceptable use policy without surveillance?

Most small companies enforce an AI AUP through approved accounts, sensible defaults and a culture where people report mistakes, not through monitoring every prompt. Give people an approved business account so the convenient route is the allowed one. Turn on the administrator settings your plan offers. Make reporting safe, and treat a pattern of ignored rules as a normal performance conversation.

If you do review usage logs available in a business plan, say so in the policy and check the employment and privacy rules that apply. Hidden monitoring tends to damage the reporting culture the policy depends on.

Tosheo.ai takes the same line in its planned design: managers see skill progress and completion, never learners' raw chat logs.

When is an acceptable use policy not enough?

An AUP covers employee behaviour; it does not decide which vendors are safe, whether a workflow needs a data protection impact assessment, or how AI fits into a product you sell. If your company builds AI features, automates decisions about people or works in a regulated profession, you also need the governance layer and specialist review.

It is also not training. A rule like "check every claim" is easy to agree with and hard to do well under deadline. People need practice on realistic tasks to build that habit, which is what role-based practice is for.

Sources and scope

NCSC, ChatGPT and large language models: what's the risk?, 14 March 2023 ↗

Microsoft Learn, Enterprise data protection in Microsoft Copilot and Copilot Chat, checked 16 October 2026 ↗

ICO, Guidance on AI and data protection (under review), checked 16 October 2026 ↗

NIST, Generative AI Profile (NIST AI 600-1), July 2024 ↗

European Commission, AI Literacy - Questions and Answers, last updated 27 July 2026 ↗