Three Things to Put in a One-Page AI Acceptable Use Policy
Most AI policies fail by being too long to read. The one-page rule forces you to consider what is most important.
An SME needs an AI policy that is to-the-point. Something a new joiner can read on their first morning, understand, and apply to a real decision they have to make about a real tool by Wednesday. Anyone reading the policy should quickly know what can and cannot be done.
The format that does this is one A4 page. Not a 30-page document with appendices. Not a wiki maze of cross-referenced sub-policies. One page. Roughly 400-600 words. If it does not fit, something has to go.
If a one-page AI Acceptable Use Policy (AUP) covers these three things, the team can make the everyday decisions. If it covers anything else without these three, the AUP is not actually load-bearing.
1. Data Categorization
The single biggest source of AI-related risk for SMEs is not malicious actors. It is well-meaning employees pasting things into AI tools that should not have left the building. Customer data included in a ChatGPT "summarise this email thread" prompt; Draft commercial contracts inadvertently sent to Claude to "make this clearer." A column from the customer database into a code-assistant prompt for "give me a SQL query for this."
None of the users writing these prompts see themselves as exfiltrating data. They thought of themselves as being efficient. The AUP's job is to give them the language to recognise the line.
The minimum viable form is a three-tier classification:
- Public: content already on the website, in marketing materials, or in published documents. Pasting this into any AI tool is fine.
- Internal: day-to-day business content that is not public but is not sensitive — meeting notes, internal documents, draft language for review. May be pasted into approved AI tools (see Thing 2) but not into personal accounts on consumer AI products.
- Restricted: customer personal data, commercial contracts and pricing, employee personal data, financial data, security-relevant information (passwords, API keys, infrastructure topology), legal hold material. Must never be pasted into any AI tool, on any account, for any purpose, without an explicit data processing agreement and your data protection lead's sign-off.
That is the entire data section. Three labels and a one-line explanation each. A team that reads this once and refers to it once a month will catch the great majority of risky paste actions. Add a fourth tier only if the business actually has one (e.g., medical or financial-services-regulated content) and the addition is load-bearing.
2: Approved Tool List
The vague version of this clause says "use approved tools only." That clause does no work. The team does not know what is approved and the AUP did not tell them. The clause that does work is a five-row table of the actual tools approved as of today, the kind of work each one is for, and the kind of data each one is allowed to handle.
A typical SME starting position looks like this:
| Tool | Approved for | Data tier allowed | Owned account? |
|---|---|---|---|
| Microsoft Copilot in Office 365 | Document drafting, summarisation, email drafting | Internal | Yes — corporate tenant |
| ChatGPT (consumer) on personal accounts | Public-information research only | Public only | No — personal accounts |
| Claude Pro (corporate-owned account) | Long-form drafting, analysis | Internal | Yes — billed centrally |
| Notion AI (in our existing Notion workspace) | Notes, documentation, task summarisation | Internal | Yes — inside our workspace |
| Other (anything not on this list) | Ask {named person} before use | — | — |
Two things make this work. First, the table has named tools rather than categories. "Generative AI tools" is not a row anyone can apply on a Tuesday. "Microsoft Copilot in Office 365" is. Second, the last row provides an explicit escalation path — there is a named person who decides about anything not on the list, and the team knows who that is.
The list is a living document with periodic reviews to keep it current.
3: Output Review Obligation
The third thing is a single sentence that changes how the team treats AI output. Without it, AI tools are treated as oracles. With it, they are treated as draft generators.
The sentence: Every AI output is a draft. A named human is accountable for the final version that leaves the team.
Accountability is also one of the DSIT five principles for AI governance:
- A proposal drafted with AI assistance is reviewed and signed off by the proposal owner before it goes to the client.
- A summary of a customer call generated by AI is checked against the call recording before being locked in the CRM.
- A piece of code suggested by an AI assistant is reviewed and tested by the engineer before being committed.
The AUP does not need to enumerate every workflow. It needs to establish the principle, name the accountability, and let the team apply it. Over-trusting AI responses, that could contain hallucinated output is a significant risk to business that this pragmatic step helps mitigate.
AUP and Koallabs SAFE-AI
The AUP is the operational anchor of [[Stage 3 - Secure]]. Stage 3 is where governance becomes actual policy that the organisation operates against, and the AUP is likely to be the single most-read artifact.
For organisations operating in the UK, the AUP also discharges part of the obligation under UK GDPR Article 28 — controllers using AI vendors as processors must have in place documented instructions for processing. The data sensitivity section of the AUP is one component of those documented instructions. It is not the whole obligation, but it is the part that lives where the team actually works.
What to Skip on a One-Page AUP
Three things that do not earn their place at the SME stage:
- Lengthy AI ethics statements. While ethics are important, teams should be trained through other means such as dedicated courses and refreshers.
- Detailed vendor evaluation criteria. The Vendor Evaluation Framework is its own document, owned by the Data Protection Lead. The AUP says "use approved tools" and points at the list.
- Incident response procedures. Incident response belongs in the existing security incident playbook, with AI added as a possible incident type. Putting it in the AUP makes the AUP unreadable.
If the AUP grows past one page in your business, the question to ask is not "how do we shorten it" but "which of the three guiding topics became lost?" The answer is almost always output review obligation — the accountability sentence is the easiest to lose under "cargo-culted enterprise-policy", yet it is the most operationally important.
What Next?
If you do not have a one-page AUP and want a starting template, we can help you get started - why not book a 30-minute discovery call?
A one-page AUP that the team has read beats a 30-page AUP that the team has not. That is true even when the 30-page version is more complete on paper. The work happens where the team is, and the team is meeting on a Tuesday morning with seventeen other things to do.
Want to see where your organisation stands? Take the AI Readiness Scorecard to get a personalised AI adoption readiness report in just a few minutes.
Ready to scope an engagement? Book a 30-minute discovery call — no sales script, just a conversation.
