The One-Page AI Ethics Policy — How To Write One This Afternoon
Six sections, four steps and about three hours to give your business a written AI rulebook that people can actually follow.
Most small businesses have no written rule about AI. Staff paste client data into free tools, someone publishes unchecked AI copy, a freelancer quietly uses a chatbot on a confidential brief, and nobody knows who approved any of it. You do not need an ethics committee or a forty-page framework. You need one page that says what is allowed, what is not and who decides. This playbook walks you through writing it in an afternoon, section by section, with sample wording you can adapt.
Why Write It Down
An unwritten policy is not a neutral gap. It is a policy of 'everyone guesses,' and the guesses will differ. One person thinks pasting a client list into a free chatbot is fine because it saves time. Another would never dream of it. When something goes wrong, there is no standard to point to, and the person who guessed wrong takes the blame for a failure that was really yours.
A written page protects your customers, who deserve to know their data is handled consistently. It protects your staff, who should not have to improvise under deadline pressure. And it protects you, because it shows you took reasonable care.
Budget: about 2 to 3 hours for a first draft and 30 minutes for a team review. Estimated, not measured, and mostly spent making decisions rather than typing. AI can help you draft the wording, but every line has to be a choice you actually make. A policy written entirely by a chatbot, and adopted without thought, is worse than none, because it creates the illusion of control.
Before You Start: Find Out What Is Really Happening
Do not write the policy from imagination. Spend 30 minutes finding out how AI is used today. Ask each person, privately and without blame: which AI tools do you use for work, what do you put into them, and what do you use the output for? You will almost certainly discover uses you did not know about, including personal accounts on free plans.
Promise that nobody will be punished for answers given now. The goal is an accurate map. If people fear consequences, they will hide current use, and your policy will be built on fiction.
Section One: Approved Tools
List which AI tools staff may use for work, and on which plan. This matters because free consumer tiers and paid business tiers often have different terms on data retention and on whether your inputs may be used to improve the vendor's models. Read the terms, or have someone who can read them do it.
Sample wording: 'Staff may use only the AI tools on the approved list below, signed in with a company account. Any other tool requires written approval from [name] before first use.' Keep the list short, name an owner, and review it quarterly.
Section Two: Data Rules
State what must never be pasted into an unapproved tool: customer personal data, passwords and access keys, financial account details, health information, unpublished contracts, anything covered by a confidentiality agreement, and other people's private information.
Then say what counts as safe: public information, your own general writing, and material with names and identifiers removed. Give a concrete example of anonymising: replace 'Maria Gonzalez at Harbor Dental' with 'Client A, a dental group.'
Sample wording: 'Never enter personal data about customers or colleagues into any tool unless it is on the approved list and the task has been cleared for that data.' If you operate in a regulated field, such as health, finance or legal services, get professional advice on what additional restrictions apply.
A policy nobody has read protects nobody. If it does not fit on one page, it will not be followed.
Section Three: Human Review
Name the outputs that always need a human check before use: anything customer-facing, anything legal, financial or medical, anything that makes a claim about a product, and anything about a specific person. State who reviews and what reviewing means: reading the whole thing, checking every number and every factual claim against a source, and taking responsibility for it as if you had written it yourself.
The most important sentence in this section is: 'The person who sends it owns it.' AI-assisted work is not someone else's mistake. If a number is wrong or a claim is false, the author who published it is accountable.
Section Four: Disclosure
Decide when you tell customers that AI was involved. A sensible default: say so whenever a customer is talking directly to a bot, and never present AI-generated reviews, testimonials, 'customer photos' or staff biographies as real. Many platforms and some laws restrict deceptive uses, so check what applies to you.
Sample wording: 'Customers are told when they are chatting with an automated assistant and can reach a human on request. We never publish fabricated reviews, testimonials or images presented as genuine.' Also decide whether you will disclose AI use in client deliverables. Many clients now ask, and an honest answer prepared in advance beats an awkward one improvised.
Section Five: Decisions About People
State clearly that no AI system makes a final decision about hiring, firing, discipline, credit, a specific person's price or access to a service. It may inform or advise. A named human decides, and applies the four-question test from Day 52: can we name the inputs, give a two-sentence reason, review and override, and offer an appeal?
This is the section that most protects you from legal and reputational harm, and it is the one most often missing. Keep it short and absolute.
Section Six: Incidents And Questions
Say what to do when something goes wrong: stop using the output, tell [name] within 24 hours, and write down what happened, which tool and which data. Then say what happens next: you will assess whether anyone was affected, fix the process, and tell affected parties where required.
Make reporting safe. If people fear blame, they will hide mistakes until they become disasters. Sample wording: 'Reporting a mistake promptly is never grounds for discipline. Concealing one may be.' Also give people somewhere to ask questions, such as a named person or a channel, so that 'I wasn't sure' has a place to go.
Writing It This Afternoon
Step one, 30 minutes: map current use, as above. Step two, 45 minutes: for each use, mark it green (fine), amber (fine with review) or red (not allowed). Step three, 60 minutes: turn those marks into the six sections in plain language, with no jargon and no sentence longer than about 20 words. Step four, 30 minutes: have one or two staff members read it cold and point out anything unclear, unrealistic or missing.
Keep rules concrete. 'Never paste customer names or emails into a tool that is not on the approved list' beats 'Handle data responsibly.' Put the date and the owner's name at the top, and keep the whole thing to one page. If you need an appendix, the main page is not clear enough.
A Sample One-Page Skeleton
To make this concrete, here is the shape of a finished page. Title and date: 'Our AI Use Policy, reviewed [date], owner [name].' Purpose, one sentence: 'We use AI to work faster and better while protecting our customers, our staff and our reputation.' Then six short blocks, each with two to four bullet rules, covering tools, data, review, disclosure, decisions about people and incidents.
A typical finished page runs 350 to 450 words. If yours is past 600, cut. Every rule should be something you would actually enforce, and every rule should be something a new hire could follow without asking. Test it: hand the page to someone who has never seen it and ask them what they would do if a client sent a spreadsheet of customer emails and asked for a summary. If they answer correctly without help, the policy works.
Special Cases You Will Hit
Freelancers and contractors. They are often your biggest blind spot, because they use their own tools and you never see their settings. Put your data rules into the contract: which tools they may use on your material, what they must never paste, and that they must tell you when AI contributed to a deliverable.
Client-supplied data. Your clients may have their own rules, or contract terms that forbid putting their material into third-party AI. Check before you assume a tool is fine. If a client's contract is stricter than your policy, theirs wins.
Meeting recorders and transcription tools. These capture other people's voices and words. Recording consent laws differ by place, so check yours, and tell participants when a recorder is running. Decide where transcripts are stored and who may see them.
Images, voices and likenesses. Do not clone a real person's voice or generate their likeness without clear consent, and do not use AI-generated imagery in a way that misleads customers about what they are buying.
Rolling It Out
Introduce the policy in a 20-minute team conversation, not an email. Walk through each section with one real example, answer questions and ask what is unrealistic. If staff say a rule is impractical, take it seriously; they are telling you where the policy will fail. Fix the rule or provide a better tool, rather than insisting.
Then make compliance easy. Give people the approved tools on company accounts, a short list of green-light uses they can start on immediately and a named person to ask. A policy that only says no, and offers nothing in its place, produces secret workarounds.
Where Policies Fail
Four common failures. The policy is too long, so no one reads it. It is too strict, so staff ignore it and use personal accounts. It is never reviewed, so it describes last year's tools. And it has no owner, so nobody answers questions or updates it. Fix all four by keeping it short, making the approved tools actually good enough to use, putting a quarterly review in the calendar and naming one person responsible.
Another honest limit: a policy cannot stop a determined person from breaking it. What it does is set expectations, make safe behaviour the easy default and give you grounds to act when someone ignores it.
Keeping It Alive
Review it every quarter, because tools and laws change. Mention it at onboarding and again when you add a tool. When something goes wrong, update the page rather than only blaming a person. If you handle sensitive data or operate in a regulated field, have a qualified professional review the final version. This post is a starting template, not legal advice.
Draft Your Green, Amber And Red List
Ask your team how they use AI today, promise no blame, and sort each use into green, amber or red. That list is the skeleton of your policy. Do it before the end of the week.
