A one-page AI use policy for a 50-300 person company
Most employees already use AI at work, and most companies still haven't told them the rules. A plain one-page AI use policy you can adapt this week: approved tools, green, yellow and red data, who reviews AI work before a customer sees it, and what Texas law does and doesn't require.

In brief
A one-page AI use policy for a 50-300 person company needs five things: the approved tools (on company accounts), a green, yellow and red rule for what data can go into them, who reviews AI work before it reaches a customer, a rule that nobody connects an AI tool to company email, files or the ERP without IT's OK, and one person to ask. Surveys keep finding that only about a third of companies or fewer have a formal policy. Texas's AI law took effect Jan. 1, 2026, but it doesn't require a private company to write one. You write it because your people are already using AI, and a short rule they read beats a long one they don't.
- Only about 28-36% of companies have a formal AI policy in 2025 surveys, while 46% of U.S. workers admit uploading sensitive company data to public AI tools.
- One page covers it: approved tools, green, yellow and red data, who checks AI work, what can be connected, and who to ask.
- Company accounts are the biggest single fix: business data terms, and the history stays when people leave.
- Courts have held companies and professionals responsible for what their AI produced, so a person checks prices, dates and part numbers.
- Texas's AI law (effective Jan. 1, 2026) doesn't require a written policy, but following the NIST profile gives a defense.
Your people already decided. The policy is catching up.
In most 50-300 person companies, nobody held a meeting about AI. People just started using it. Someone in purchasing rewrites supplier emails with ChatGPT. Someone in accounting asks it to fix a formula. A sales rep pastes a customer's spec sheet in and asks for a summary.
Most of that is fine. Some of it isn't. The trouble is that nobody told them which is which.
The surveys line up on this. Each one asked a different question, so read them as separate findings:
Few companies have told their people the rules
Three 2025 surveys, three different questions. Each bar is the share that said yes.
Sources: ISACA 2025 AI Pulse Poll, as published by iTWire; Gallup, June 2025; EisnerAmper, 2025. All are surveys.
Meanwhile, the risky habits are already there. In a 2025 KPMG and University of Melbourne study, 46% of U.S. workers admitted uploading sensitive company information or intellectual property to public AI tools. And 58% said they rely on AI output without checking it. That's a survey of what people admit to, so the real numbers may be higher.
We wrote about shadow AI at work: why people go around you and why bans rarely work. This guide is the next step. It's the actual page you hand out.
Why one page
A 30-page AI policy written by a law firm has a place at a bank. In a 120-person distributor, nobody on the order desk will read it. They'll sign the acknowledgment and keep doing what they were doing.
One page forces you to decide what matters. If a rule can't be said in a sentence, the people it's for won't remember it on a busy Tuesday. You can keep a longer version for IT and legal. The one page is what the staff actually reads.
The policy that works is the one people can repeat back to you.
The five things it has to say
Every rule on the page answers one of these questions. If yours answers all five, it's done.
- Which tools can I use? Name them. Company accounts only for work data.
- What can I put in them? Green, yellow and red data.
- Who checks the work? Anything AI helps write that reaches a customer, a supplier, a price or a person's job gets a human review.
- What can I connect? Nothing gets access to company email, files or the ERP without IT's OK.
- Who do I ask? One name, and a promise that asking doesn't get anyone in trouble.
The template
Copy this, change the names in brackets, and cut anything that doesn't fit your company. It's written for the staff, not for lawyers.
[Company] AI use policy (one page)
| Section | What it says |
|---|---|
| Why we have this | AI tools can save you time. We want you to use them. This page says which ones, what you can put in them, and when a person has to check the result. |
| Approved tools | [Tool 1, e.g. Microsoft 365 Copilot] and [Tool 2], signed in with your company account. Personal accounts and free tools are fine for general questions, never for company data. Want a new tool? Ask [name]. |
| Green: use freely | Public information, general questions, your own rough drafts, formulas, rewording an email that has no customer details in it. |
| Yellow: approved tools only | Customer emails and POs, price lists, quotes, supplier documents, internal reports, meeting notes. |
| Red: never | Passwords and logins, bank and card numbers, Social Security numbers, health information, personnel files, and anything a customer contract or NDA says can't leave our systems. |
| A person checks | Anything AI helped write gets read by you before it goes to a customer or supplier. Prices, quantities, dates and part numbers get checked against the system, every time. Hiring, firing, pay and discipline decisions are made by people, not by AI. |
| Connecting things | Don't give any AI tool, add-on or app access to company email, files, the ERP or the CRM without [IT name]'s OK, even if it asks nicely. |
| Mistakes | If you put something in the wrong place, tell [name] the same day. Nobody gets in trouble for reporting it. We only get in trouble when nobody knows. |
| Questions | Ask [name]. We review this page every [six] months. Last updated [date]. |
That's the whole thing. The rest of this guide explains the choices behind it, so you can defend them or change them.
Why "company accounts" is the rule that matters most
The biggest single fix is moving work onto business accounts. Business versions of the major tools come with different data terms than the free ones. For example, Microsoft says in its Microsoft 365 Copilot privacy documentation that prompts, responses and data accessed through Microsoft Graph aren't used to train its foundation models. It also says Copilot only shows a user data they already have permission to see.
Two cautions. First, read the terms of the specific plan you buy; don't assume. Second, "only shows what you already have permission to see" cuts both ways. If your file permissions are a mess, Copilot will find the mess. Clean up who can see the payroll folder before you turn it on.
Company accounts also fix the quiet problem: when someone leaves, you shut off their access and their chat history stays with the company.
How common is the other way? A 2025 LayerX report, as reported by eSecurity Planet, found that 71.6% of generative AI access in the companies it studied went through non-corporate accounts. LayerX sells browser security, so treat it as a vendor's data, but it matches what we see.
Why "a person checks" is not optional
Two court cases make the point better than any policy memo.
Air Canada, 2024. A customer asked the airline's website chatbot about bereavement fares. The chatbot gave him wrong information. When he sued, Air Canada argued the chatbot was responsible for its own actions. A British Columbia tribunal rejected that. As summarized by the law firm DWW, the tribunal said it makes no difference whether the information comes from a static page or a chatbot. Air Canada paid $812.02 in damages, interest and fees. The amount was small. The rule wasn't: what your AI tells a customer, your company told the customer.
The lawyers and ChatGPT, 2023. In Mata v. Avianca, lawyers filed a brief citing court cases that ChatGPT had made up. A federal judge in New York ordered a $5,000 penalty. The tool didn't sign the brief. The people did.
Your version is smaller and more likely: an AI-drafted quote with the wrong unit price, or an order confirmation with a ship date nobody checked. That's why the template singles out prices, quantities, dates and part numbers. Those are the details AI gets wrong with confidence, and the ones customers hold you to. For how to build that check into a real process instead of hoping people remember, see human-in-the-loop automation.
What Texas law does and doesn't require
Texas passed the Texas Responsible Artificial Intelligence Governance Act (HB 149), which took effect Jan. 1, 2026. It applies to anyone who does business in Texas. Here's what an owner should know, based on the enrolled bill:
- It doesn't require a written AI policy. Nothing in it says a private company must have one.
- Its bans are about intent. For example, it bars using AI with the intent to unlawfully discriminate. The bill says a disparate impact alone isn't enough to show that intent.
- The disclosure rules mostly fall elsewhere. The duty to tell people they're dealing with AI falls on government agencies and health care providers, not ordinary private businesses.
- Only the Attorney General enforces it. There's no private lawsuit under the act, and there's a 60-day window to fix a problem after notice. Penalties for violations that aren't fixed can be much larger.
- There's a safe harbor. A company that finds a violation through its own internal review, or that substantially follows the federal NIST generative AI risk profile, has a defense.
That last point is a good reason to keep a short written record of how you review AI use, even though the law doesn't require a policy. This is a summary, not legal advice. If you're in health care, lending, insurance or hiring at scale, have your lawyer read the act and any federal or other state rules that apply to you.
The NIST AI Risk Management Framework it refers to was released in January 2023 and is voluntary. NIST added a Generative AI Profile in July 2024. Both are long. You don't need to adopt them to write a one-page policy, but they're where a longer version would come from.
Where the other side has a point
Some owners say a policy is paperwork, and that the real control is training and a good IT setup. They're partly right. A policy no one enforces does nothing, and the best protection is technical: company accounts, single sign-on, and admin settings that block unapproved apps from reading mail.
Others worry a policy will scare people off AI entirely. That's a real risk if it reads like a list of punishments. That's why the template starts with "we want you to use them" and says nobody gets in trouble for reporting a mistake. If the page makes people hide their AI use, it has failed.
And some will say one page is too short. For a 50-300 person company, we disagree. One page that people follow beats a complete document that sits in a shared drive.
How to roll it out in two weeks
- Day 1-3: Ask each department, without blame, which AI tools they use and for what. The answers tell you which tools to approve.
- Day 4-5: Pick one or two tools and buy business accounts. Turn off or restrict the rest where you can.
- Day 6-8: Fill in the template. Have one person from the order desk and one from accounting read it and tell you what's unclear.
- Day 9-10: Have IT review which AI apps and add-ons already have access to email and files. Our guide on third-party apps with access to Microsoft 365 mail walks through it.
- Day 11-14: Hand it out in a 15-minute team meeting, not by email. Show one green, one yellow and one red example from your own company.
Then put a date on the calendar six months out to read it again. AI tools change fast, and a policy that names last year's tools is a policy people stop trusting.
Where this leads
Once the policy is in place, you'll notice a pattern in the yellow column. The same few jobs show up again and again: reading customer POs, answering order-status emails, pulling numbers into a weekly report. When three people are pasting the same kind of document into a chat window every day, that's not a policy question anymore. It's a process that should be built properly, with the data staying in systems you control and a person approving the result.
Common questions
What should a one-page AI use policy include?
Five things: the approved tools on company accounts, a green, yellow and red rule for what data can go into them, who reviews AI work before it reaches a customer, a rule against connecting AI tools to company email, files or the ERP without IT's OK, and one person to ask.
How many companies have an AI use policy?
Fewer than half, in the 2025 surveys we found. ISACA reported 28% of organizations have a formal, comprehensive AI policy; Gallup found 30% of employees say their organization has guidelines or policies; EisnerAmper found 36% of U.S. desk workers say their company has a formal policy.
Does Texas law require a company to have an AI policy?
No. The Texas Responsible Artificial Intelligence Governance Act (HB 149), effective Jan. 1, 2026, does not require private companies to write one. It bans certain intentional misuses, is enforced only by the Attorney General, and gives a defense to companies that substantially follow the NIST generative AI risk profile. This is not legal advice.
Is a company responsible for what its AI chatbot tells customers?
In the Air Canada case in 2024, a British Columbia tribunal said yes: it made no difference whether wrong information came from a web page or a chatbot. Air Canada paid $812.02 in damages, interest and fees.
Should employees use personal ChatGPT accounts for work?
Not for company data. Business accounts come with different data terms, and the company keeps the history when someone leaves. Personal accounts are fine for general questions with no company information in them.
What data should never go into an AI tool?
Passwords and logins, bank and card numbers, Social Security numbers, health information, personnel files, and anything a customer contract or NDA says can't leave your systems.
How often should we update our AI policy?
At least every six months. Tools change quickly, and a policy that names outdated tools stops being taken seriously.
Sources
- ISACA 2025 AI Pulse Poll, as published by iTWire, June 2025
- Gallup, AI use at work has nearly doubled in two years, June 2025
- EisnerAmper, Survey of employees using AI at work, 2025
- KPMG and University of Melbourne, Trust in AI 2025 (U.S. findings)
- eSecurity Planet, LayerX report on shadow AI, Oct. 2025
- Microsoft Learn, Data, privacy and security for Microsoft 365 Copilot
- DWW, BC tribunal finds Air Canada liable for inaccurate advice given by website chatbot
- Barry Sookman, Moffatt v. Air Canada: a misrepresentation by an AI chatbot, Feb. 2024
- Legal Dive, Lawyers fined over fake ChatGPT cases, June 2023
- Texas Legislature, HB 149 enrolled text (Texas Responsible Artificial Intelligence Governance Act)
- NIST, AI Risk Management Framework
Shadow AI: your employees already use ChatGPT at work. What to do about it
Human-in-the-loop automation: why a person should approve every step
Is AI safe for customer and order data? Questions to ask any vendor