While a company is still debating whether staff may use AI, staff are already using it: weekly notes, emails, a contract translation, meeting minutes. The useful question is not whether to allow it. It is how to use it without an incident. Did a customer file leave the company? If the draft is wrong, who is responsible? Does a public post need a label?
This is a one-page policy you can edit, with the reason for each line. It fits a company of tens or hundreds of people. You do not need a legal department to put it on the wall. It is management practice, not legal advice. For a specific compliance question, use the current text of the law and a lawyer.
Companies operating in China have a few extra lines, covered below. A team elsewhere can keep the same page and replace that section with its own local rules.
One page, not a manual
The longer the policy, the less it is read. People need four things: what may go in, which tools, who owns the output, and who to ask when they are unsure. Those four, printed beside a desk, beat a thirty-page AI regulation nobody opens.
A short policy is also easier to update. The tools change every few months. Look at the page once a quarter. Do not file it and freeze it.
What may go into an AI tool
This is the line that matters most. Split company material into three piles, with examples a person recognizes. “Sensitive information” alone is not a rule anyone can apply.
| Class | Examples | Rule |
|---|---|---|
| Allowed | Published product copy, public web pages, public industry material, writing and translation with no specific customer or price | Any tool the company has approved |
| Only after names and figures are removed | Internal process notes, meeting minutes, ordinary business email, analysis that does not name the customer | Remove customer names, amounts, personal names, and contact details first. Or use only a tool hosted inside the company |
| Not in an external tool | Customer records and quotations, contract text, drawings and formulas, financials, personnel files and pay, ID numbers and phone numbers | Only a tool the company hosts and has approved |
The examples are what make the page usable. People will not define “trade secret” at their desk. They know what a quotation is.
Which tools
A long ban list is weaker than a short allow list: these tools are approved, and each one is for these jobs. Anything else, ask first.
- Company accounts, not personal ones. Access can be removed when someone leaves, and a problem can be traced.
- Read the data terms once, centrally. Will inputs be used to train a model, how long are they kept, and can that be turned off? One owner checks this. Do not leave it to every employee.
- For work in China, prefer a model service that has completed the required filing. Where the material is sensitive, an open model on the company’s own network keeps the files inside.
Who is responsible for what the AI wrote
The person who used it. The model is an assistant, not an author of record. This has to be plain, because a model will state a wrong number, invent a source, or cite a clause that does not exist.
- Anything sent outside (email, quote, contract, public copy) is checked by that person for facts, figures, and promises before it is sent.
- Amounts, legal clauses, and technical parameters are checked against the original document, not against the summary.
- Images, video, and articles posted on a public platform follow that platform’s label. In China, the labeling rules below apply from 1 September 2025.
The China rules worth knowing
A smaller company does not need to read every regulation. The hard lines on the one-pager sit on a few texts. If you do not operate in China, keep the page and swap this section for your own.
- Interim Measures for the Management of Generative Artificial Intelligence Services, in force from 15 August 2023. They mainly regulate providers offering generative AI to the public inside China. Internal research and use that is not offered to the public is outside that scope. Article 4 still speaks to both providing and using the services: respect intellectual property and business ethics, protect trade secrets, and do not infringe privacy or personal information.
- Measures for Identifying Artificial Intelligence-Generated and Synthetic Content, in force from 1 September 2025. Article 10 says that when a user publishes synthetic content through an online information service, they should declare it and use the label the platform provides. Nobody may maliciously delete, alter, forge, or hide that label.
- Personal Information Protection Law. Sending a customer’s or employee’s personal information to a third-party tool is itself a processing activity. That is the main reason personal information stays out of external tools.
The Chinese texts are on the government site: Interim Measures for Generative AI Services and Measures for Identifying AI-Generated Content.
A page you can send
Replace the brackets and this is a policy you can issue.
[Company] — how we use AI (version 1, [date])
We want people to use AI. To protect customers, the company, and you, follow these lines.
1. Tools you may use: [Tool A] for writing, translation, and summaries. [Tool B] for questions against internal documents. Ask [owner] before using anything else. Use your company account.
2. Do not put these into an external AI tool: customer records and quotations, contract text, drawings and formulas, financials, personnel files and pay, ID numbers and phone numbers. If those need AI, use [internal tool].
3. Remove names and figures first: internal process notes, meeting minutes, ordinary business email. Delete customer names, amounts, personal names, and contact details before you use them.
4. You own the result: AI is wrong often. Before you send anything outside, check the facts, the numbers, and the promises. For amounts, clauses, and parameters, go back to the original document.
5. Label public posts: if you publish an AI-generated image, video, or article, declare it the way that platform asks.
6. If you are unsure, ask: [owner / channel]. One question is cheaper than a cleanup.
This page is reviewed every quarter.
How the page gets used
- A short session, with examples. Not a lecture on models. “Can this file go in? Does this sentence need a check?” Twenty minutes beats an email.
- Give a decent tool before you ban the others. If people have no safe tool that actually works, the ban moves the work onto personal phones. Issue the policy and the tool together.
- Name one owner. People know who to ask. When a tool changes, someone follows it.
- Collect the real questions, and update quarterly. The questions people actually ask are the next version of the page.
If you do not yet have an internal place to ask questions against your own files, that is a knowledge base, not another public chat account. A 3-minute readiness check shows which of those you are missing.
FAQ
Does a small company need a written AI policy?
Yes, and earlier rather than later. The smaller the company, the larger one leaked customer file is, relative to the business. The page above is an afternoon of work. Cleaning up after a leak is not.
Is a total ban the safest option?
Usually no. A ban pushes people onto personal accounts and phones, which are harder to see. A safer pattern is an allowed tool that works, and a short boundary around it.
What is the risk of personal AI accounts for work?
Three things. The company cannot see what was typed in. After someone leaves, the chats remain on their account. The data terms of a personal plan are often different from a company plan. That is why the page says company accounts.
Do we still need training after the policy is written?
Yes. The policy says what not to do. Training is how to do the real job inside that boundary, including how to run and edit a skill. The two together are what make people willing to use it and careful when they do. See how a skill differs from a chat.
If you do not have an internal tool yet
A policy works better beside a tool people are allowed to use. We can look at one job, such as questions against your own documents, and say whether a packaged tool is enough.
Talk through the first allowed tool →