TL;DR
Your team is probably already using AI, and some of it is on personal accounts nobody approved. You can fix that with a one-page policy written in about four hours: find out what's in use, buy one business-tier tool, write green, yellow and red data rules using examples from your own work, and walk the team through it in 20 minutes. Banning AI tends to backfire. In a 2026 TrustedTech survey, 37% of decision-makers said they'd keep using AI tools even if their workplace banned them.
You can write a usable AI policy for a 10-to-250-person company in about four hours. It should fit on one page and answer four questions: which AI tools are approved, what data can go into them, what needs a human check before it leaves the building, and who to ask when something isn't covered. Start by finding out what your team already uses, because odds are they're using something. Then pick the approved tools, write the data rules with examples from your own business, and walk everyone through it in a 20-minute meeting.
Does a small business really need an AI policy?
Your team is already using AI. What you don't know yet is which tools, and what's going into them.
In the U.S. Chamber of Commerce Foundation's Main Street AI Monitor (Ipsos, May 2026, 1,070 people who work at small businesses), 19% said AI showed up at their company mostly because employees started exploring tools on their own. Only 11% said it came from company guidance. About one in 10 had been offered any formal AI training.
Bigger surveys show the same pattern. TrustedTech's 2026 study of 2,001 U.S. and U.K. employees, run by Censuswide in March 2026, found 48% use AI tools their employer hasn't approved. It's higher at the top: 65% of decision-makers do it, compared with 31% of the people below them.
Then there's what gets pasted in. In BlackFog's January 2026 survey of 2,000 workers at companies with more than 500 employees, respondents reported sharing research and data sets (33%), employee data like payroll and performance information (27%), and financial statements or sales data (23%) on AI tools their employer hadn't approved. Of the people using unapproved tools, 58% were on free versions.
Those companies are bigger than most of the ones we work with, but the mechanics are the same at 40 people. Someone opens a free account, pastes in a client spreadsheet and asks for a summary. Nobody meant any harm. Nobody knew there was a rule, because there wasn't one.
What should a one-page AI policy cover?
Six sections. If one doesn't apply to you yet, leave a one-line placeholder so people know it's coming.
- Approved tools. The specific products and plans the company pays for, like "our ChatGPT Business workspace" or "Claude Team," and how to log in. Personal accounts are out for company work.
- Data rules. What can go in, what needs care, and what never goes in. Use examples from your own work, not abstract categories.
- Human review. Which AI output a person has to check before it reaches a client, a candidate or the public.
- Client disclosure. When you tell clients AI was involved, and how. Some client contracts already have something to say about this, so check yours.
- Connectors and agents. Who can connect an AI tool to email, the CRM, the shared drive or the accounting system, and who approves it.
- Questions and new tools. One named person who answers "can I use this?" within a day, and a simple way to request a new tool.
How do you write it in one afternoon?
Block four hours. Bring whoever runs operations and the one person on the team who uses AI the most. They'll know things you don't.
Hour 1: Find out what's already in use
Send a five-question anonymous form at the start of the week so the answers are in by your afternoon. Ask which AI tools people use for work, whether they're free or paid, whose account they're on, what tasks they use them for, and what they'd use more if the company paid for it. Keep it anonymous or you'll get the answers people think you want.
Then check the paper trail. Search expense reports and company card statements for OpenAI, Anthropic, Perplexity, Midjourney, Otter and Fireflies. In the Google Workspace admin console, go to Security, then Access and data control, then API controls, and look at which third-party apps have access to company accounts. In Microsoft 365, check Enterprise applications in Entra ID for apps users have consented to. Meeting note-takers and email add-ins that nobody remembers installing tend to show up here.
Hour 2: Pick the approved tools and lock the settings
Pick one general assistant for the whole company and buy the business tier. ChatGPT Business, Claude Team, Gemini in Google Workspace and Microsoft 365 Copilot all come with an admin console, and their business plans keep company data out of model training by default. Free and personal plans don't give the company that control, and some use chats for training unless the user turns it off.
Pick the one that fits where your files live. If the company runs on Microsoft 365, Copilot's access to Outlook and SharePoint counts for a lot. If your team writes long proposals, reports or analysis, test Claude and ChatGPT side by side on three real tasks before you decide.
Then go through the admin console. Turn on single sign-on if your plan includes it, set data retention, and decide which connectors (Google Drive, SharePoint, Slack, your CRM) are on. Switch off anything you haven't made a decision about yet. You can turn a connector on later. It's much harder to explain why a client's folder was searchable from every seat.
Hour 3: Write the data rules with your own examples
This is the part people actually read, so make it concrete. Three tiers is enough:
- Green: fine in any approved tool. Public information, your own marketing copy, internal process docs with no client names, generic templates.
- Yellow: approved company tools only, and a person checks the output. Client names and project details, proposals, internal financials, job descriptions, and anything that becomes a client deliverable.
- Red: never goes into any AI tool without written sign-off from the policy owner. Social Security numbers, bank and card details, health information, passwords and API keys, employee payroll and performance records, and anything a client contract says stays confidential.
Then write five examples from your actual business under each tier. "Client financial data" is abstract. "The QuickBooks P&L export we send a client every month" is something people recognize when it's sitting in their clipboard.
Hour 4: Draft it, check it, ship it
Use AI for the first draft. Paste the prompt in the next section into your approved tool along with your notes from the first three hours, then edit hard. Keep it to one page. If it runs to three, nobody reads past the first.
Then hold a 20-minute team meeting. Walk through the green, yellow and red examples, show people where to log in, and introduce the person who answers questions. Give everyone a date to move company work out of personal AI accounts, and tell them the company is paying for the tool so they don't have to.
Rolling AI out to a team right now? We're writing more step-by-step guides like this one, starting with how to turn your SOPs into an assistant people use, and they go out first in the NextGen Insights newsletter.
What prompt should you use to draft the policy?
Here's the prompt we'd start with. Fill in the brackets.
You're helping a [number]-person [industry] company write a one-page AI use policy for employees. Use plain language at an 8th-grade reading level. Use these sections: Approved tools, Data rules (green, yellow and red, with examples), Human review, Client disclosure, Connectors and agents, Questions and new tools. Here are our notes: [paste your inventory results, approved tools, data examples, and the name of the policy owner]. Don't add legal language. Don't invent rules we didn't give you. List any gaps you notice as questions at the end instead of filling them in.
That last line matters. Models will happily write a clause that sounds official and that nobody at your company ever decided. Asking for gaps as questions keeps those hallucinations out of the document and hands you a to-do list instead. Before it goes out, have whoever handles your HR or legal questions read the parts about client confidentiality and employee data.
Should you just ban AI tools instead?
We wouldn't. A ban mostly moves the usage somewhere you can't see it.
In the TrustedTech study, 37% of decision-makers said they'd keep using AI tools even if their workplace banned them and they faced discipline. Among decision-makers using unapproved tools, two of the top reasons were limited access to approved tools (29%) and the unapproved tools simply working better (28%). A paid seat and a clear rule fix both.
Here's a hypothetical. Picture a 40-person accounting firm that bans AI in January. By March, staff accountants are drafting client emails in free chatbots on their phones, a couple of partners are summarizing engagement letters the same way, and nobody's mentioning it. The firm has the same data risk it started with, and now people are hiding it. Put the same firm on one paid workspace, with client tax documents on the red list and a named person to ask, and leadership can at least see what's happening.
What does it cost, and what does it save?
The policy itself costs an afternoon of two people's time. The real cost is the seats.
At the time of writing, the business tiers of ChatGPT and Claude list at roughly $25 to $30 per user per month, depending on annual or monthly billing. Check the current pricing pages before you budget, because these change. For a 30-person company where 20 people use AI every week, that's about $500 to $600 a month if you only buy seats for the people who use it.
Some of that money is probably being spent already. Look for $20-a-month personal subscriptions on expense reports. Moving those people to a company workspace often costs only a few dollars more per person, and the company gets admin controls, ownership of the data, and the ability to shut off access the day someone leaves.
We're not going to put a savings number on this, because it would be a guess. What you're buying is the ability to answer two questions fast: which AI tools have your client and employee data in them, and what's connected to your systems. When a client asks, "Do you use AI on our work, and how?", you'll have a one-page answer ready.
How do you keep the policy current?
AI tools change every month. Review the policy every quarter.
- Put a review date 90 days out at the top, with the owner's name next to it.
- Keep a running list of tool requests. When the same request shows up a third time, make a decision on it.
- Re-check connected apps in your Google or Microsoft admin console each quarter. In BlackFog's survey, 51% of respondents said they'd connected AI tools to other work systems or apps without IT approval.
- Add to the agents section as soon as anyone wants an AI tool to send email, update the CRM or move money by itself. The rule we'd start with: agents can draft and prepare, and a person approves anything that leaves the building or costs money.
Not sure where AI is already running in your business, or where it should be? Take the AI Systems Assessment and we'll show you which tools and workflows are worth standardizing first.
