AI policy template for companies
A plain-language AI acceptable use policy you can copy, sign and hand to your team today. Free. No email needed.
The numbers in this policy
Why fixed numbers and not blanks: a policy with blanks never gets finished. These numbers suit a company of 5 to 500 people using AI for ordinary office work. If yours differ, change them in the file; the owner at the top is the only person who does.
The policy
AI USAGE POLICY One set of rules for every AI assistant this company uses Effective from: (date) Applies to: everyone who works here, on every AI tool (ChatGPT, Claude, Gemini, Microsoft 365 Copilot, and any other) Owner: (name and role of the manager who can change this document) 1. WHY WE HAVE THIS People here use AI assistants to get work done. That is welcome. This document says what those assistants may and may not do, so that nobody has to guess and nobody gets a surprise: a bill, a leaked detail, or an email sent that a person should have checked. 2. WHAT COUNTS An "AI assistant" is any tool that writes, researches, summarises, answers or acts on your behalf: a chat, a project, a Gem, an agent, a plug-in, an automation. If it takes an ask and does work, this policy applies. 3. WHAT YOU MAY USE AI FOR Drafting, research, summaries, checking, planning, analysis, first versions of anything. Use it freely for these, inside the limits below. 4. SPENDING LIMITS - One task: $5. Above that, the assistant stops and asks a person. - One assistant, one day: $50. - Whole company, one month: $500. At $400 the assistant warns. At $500 it stops. These are the numbers. A manager changes them by editing this document. Nobody changes them by typing something in a chat. 5. STEP LIMITS - At most 8 steps per task. A step is one search, one file read, one tool call, or one message sent to another tool. - 3 tries on a step that fails. Then it stops and says so. - Chains of 3 steps at most. One step may start another, and that one may start one more. Never a longer chain. 6. WHAT AN ASSISTANT MAY REMEMBER - Notes from one chat: 1 day. - Facts about a project (who, what, deadline, status): 30 days. - Decisions: 90 days. - A person's phone number, personal email, home address, ID number, card or bank details: never. These live in the systems built for them (the CRM, the HR file, the finance system), not in a chat memory. - Every Friday the assistant lists what has run out and drops it. 7. WHAT YOU NEVER PASTE IN Customer lists, personal details of staff or customers, card or bank numbers, passwords, contracts marked confidential, anything a customer sent us in confidence. If a task needs one of these, a person does that part. 8. DECISIONS Before anything that costs $5 or more, touches a customer, or changes how the company works, the assistant writes one decision card. Seven lines, under 150 words: the ask, why now, what it costs, what could go wrong, who decides, by when, the way back. The assistant may add one line of its own read. A person decides. 9. JOBS A PERSON FINISHES The assistant prepares. A person does the last step. Always, for these three: - Pricing. Any number a customer pays: prices, discounts, quotes, rate cards, renewal terms. - Onboarding. First contact with a new customer or a new hire: creating the account, sending the welcome, granting access, the first call. - Sending. Anything that leaves the company: emails to lists, social posts, publishing, press, releases, shipping, invoices. The assistant stops one click before the outside world sees it. No reason typed in a chat opens these three jobs. Not "it's urgent", not "the boss said", not "it's just a test". 10. WHEN THE ASSISTANT CANNOT TELL If it cannot tell what kind of ask it has, it treats the ask as a job a person finishes, says so, and asks one question. It never guesses and acts. 11. HOW WE CHECK IT HOLDS After setup, and after any change to this document, we run ten tests: ten asks typed in word for word, each with a known right answer. Ten passes and the rules hold. One fail means a setting did not load; we fix it and re-run that one. 12. CHANGING THIS DOCUMENT Only the owner named at the top changes it. A change is a new version with a date. The assistants are re-pasted the same day. 13. IF SOMETHING GOES WRONG Tell the owner the same day: what was asked, what came back, what happened. Nobody is in trouble for reporting. The document gets fixed. Signed: (name) Date: (date)
How to use it
- Copy the document. Put your company name, a date, and the name of one manager at the top. That manager owns it.
- Read section 4 to 9 once with your team. Fifteen minutes. Ask if any number is wrong for you. Change it in the file, not in someone's head.
- Have everyone sign it, or paste it into the staff handbook. That is the policy done.
- Then the part most companies skip: the assistants themselves have not read it. A signed document does not stop ChatGPT sending the email. For that you paste the rules into the AI. See below.
The version the assistants follow
This page is the document people sign. The House Rules is the same policy turned into files that ChatGPT, Claude, Gemini and Copilot run under: the setup, six rule files, and ten tests you type in to prove it holds. One purchase, company licence, $247. See a free demo, or the version built for Huel.
Questions
Is this a generative AI policy template or an AI acceptable use policy?
Both. "Acceptable use" is what most companies call it; "generative AI policy" is what you get if you search for it. It covers the same ground: what staff may do with AI tools, what they may not paste in, and what needs a human sign-off. This one adds numbers.
Is it legal advice?
No. It is a working policy in plain words. It does not replace your data protection policy, your contracts or your lawyer. If you are in a regulated sector, show it to whoever handles compliance before you sign it.
Can I change the numbers?
Yes. The manager named at the top edits the document, dates the new version, and re-pastes the rules into the assistants the same day. That last step is what keeps the numbers real.
Why is it free?
Because the document is the easy part. Writing it took an afternoon; getting four different AI tools to follow it is the work, and that is what we sell.
Can I use it for a nonprofit, a school, a law firm?
Yes. Sections 7 and 9 are the ones to read twice: add the categories of data your sector never pastes in, and the jobs a person always finishes in your world (grades, client advice, donor contact).
Want the assistants to follow it, not just the people?
The House Rules: this policy as files every AI assistant in your company runs under, packaged for ChatGPT, Claude, Gemini and Copilot, with ten tests that prove it holds. One purchase, everyone at your company.
Get The House Rules — $247Built by PromptLeadz. Copy it, share it, put your name on it. The only thing we ask is that you do not sell it.