Here’s the fact every AI policy has to start from: your team is already using these tools. Survey after survey lands in the same place — a large share of employees use AI assistants for work, most without mentioning it, many by pasting company content into personal accounts. So the policy question was never “should we allow AI?” That ship sailed quietly, one browser tab at a time. The real question is whether the use that’s already happening gets governed — or stays in the shadows, where you keep all the risk and none of the visibility. The answer is a one-page acceptable use policy with three rules, and to make it concrete, here’s a sample you can download, adapt, and sign (PDF) — the rest of this article explains what’s on it and why.
Rule one: what goes in — the data line
The central risk of casual AI use is data leaving your control: content pasted into a personal-account chatbot has left the building — to a vendor, under consumer terms, possibly retained, outside every protection you’ve built. The first rule maps your data classification onto prompts: public and general content — always fine (drafting, brainstorming, explaining: the everyday value, unobstructed) · internal business content — approved tools only (rule three) · Restricted content — never in a prompt, period: client records, financials, health and HR information, credentials, anything under a confidentiality agreement. One clarifying sentence in the policy prevents the most common violation: “summarize this client contract” is pasting Restricted data, even though the request sounds innocent. Most employees breaking this rule today don’t know they’re doing it — which is the whole argument for a written policy over an unspoken hope.
Rule two: what comes out — the verification line
AI output is confident by design and correct by coincidence often enough to be dangerous: fabricated citations, plausible-but-wrong figures, code that almost works. The second rule: a human verifies anything AI-assisted before it leaves the building or gets relied on — client deliverables, published content, financial or legal anything, code headed for production. The framing that makes it stick: AI is a fast first-draft colleague, not a source; you sign what goes out, same as always. Businesses with confidentiality or regulatory obligations should add their industry’s line here with counsel’s blessing — the sample leaves a spot for it.
Rule three: which tools — the approved path
The escape valve that makes rules one and two followable: a short list of sanctioned tools on business accounts and business terms — where “your data isn’t training material” is contractual rather than assumed, sign-in runs through company identity, and the account survives into offboarding instead of vanishing with a personal login. The sample includes a blank approved-tools table for exactly this, and one process promise that makes the whole policy work: want a tool that isn’t listed? Ask — answered within two business days. A slow yes is how shadow usage was born; evaluating the requests is ordinary vendor diligence with two extra questions — what happens to prompt data, and can training on it be switched off?
The rollout: the meeting matters more than the memo
Print the page, then introduce it in a meeting — because two things must be said out loud to be believed. This is permission, not prohibition: here are the approved tools, please use them, we expect them to help. And amnesty for the past: nobody’s in trouble for the pasting that already happened; from today, here’s the line. Add the acknowledgment signature (the sample has the block — the same habit as every policy worth having), fold the two-minute version into the awareness rotation and new-hire onboarding, and revisit twice a year — this field moves faster than anything else on the policy shelf, and the approved-tools table will earn its edits.
The bottom line
An AI acceptable use policy is three rules on one page: Restricted data never goes in, a human verifies what comes out, and sanctioned business accounts are the only path — introduced as permission with guardrails, signed, and refreshed twice a year. Download the sample, swap in your company name and your tools, and the AI use that was already happening finally happens where you can see it — with the client data staying home and the productivity staying yours.