Most companies already have an AI policy. Most employees already work around it. Here is how to write one people actually follow, with the tools named, the data rules spelled out, and leadership held to the same standard.
86% of office workers say their company has an AI policy. 66% say they have used AI tools at work that they believed the policy did not allow.
Both numbers come from the same survey. Wakefield Research polled 1,250 office professionals for PagerDuty this June, across the U.S., the U.K., Australia, and Japan. The gap between them is the whole story. Companies are writing AI policies. Employees are reading them, deciding they do not fit the work, and using whatever tool gets the job done.
That is not a discipline problem. It is a design problem. And it is fixable.
What the survey actually found
The PagerDuty data is worth reading closely, because it shows where the risk sits.
88% of respondents said they have shared work information with public AI tools. 43% shared emails or correspondence. 40% shared meeting notes. 34% entered customer data. 31% shared financial information or confidential company documents.
The motivation is not malicious. 89% of people who use AI at work first found the tools in their personal lives. They learned that ChatGPT or Claude or Gemini could draft an email in ten seconds at home, and they brought that habit to the office.
And they want more, not less. 77% said their company’s AI restrictions limit their professional growth. 75% said they would likely look for a new job offering better AI skills development.
One finding stands out above the rest. 81% believe leadership operates under different AI rules than everyone else.
A fair caveat: the survey covered companies with $500 million or more in annual revenue. Mid-market companies tend to have fewer formal policies, not more, so the gap between policy and practice is unlikely to be smaller there.
Why bans and vague rules fail
Most first-generation AI policies fall into one of two types.
The first is the ban. “Employees may not use generative AI tools for company work.” It is easy to write and almost impossible to enforce. The tools live on every phone and in every browser. A ban does not stop use. It just hides it.
The second is the vague warning. “Use AI responsibly and protect confidential information.” Everyone agrees with it. Nobody knows what it means on a Tuesday afternoon with a customer email open and a deadline in an hour. Can they paste the email into a chatbot? The policy does not say.
Both types share the same flaw. They describe what the company is worried about instead of telling employees what to do.
Wisconsin will not write this policy for you
Some leaders are waiting for regulation to settle the question. In Wisconsin, that wait could be long.
The state’s most visible AI law to date, 2023 Wisconsin Act 123, requires disclosures on AI-generated content in political ads. A Legislative Council study committee spent 2024 examining AI regulation. As of this fall, Wisconsin has no statute governing how private employers use AI tools internally.
That means the rules your team follows are the rules you write. Existing obligations still apply, from customer contracts to data protection laws to industry rules in healthcare and finance. But the day-to-day guidance on which tool to open and what to paste into it is up to you.
Name the tools, and know how they differ
The single most useful line in an AI policy is a list of approved tools, tied to company accounts. Employees reach for personal tools when the company has not given them a sanctioned option that works.
The major business offerings handle your data differently from their free consumer versions, and differently from each other.
Microsoft 365 Copilot works inside your Microsoft 365 tenant and respects the file permissions you already have. A strong fit for companies already running on Outlook, Teams, and SharePoint, with the tradeoff that it only sees what your permissions allow, so messy permissions become a Copilot problem.
OpenAI’s ChatGPT Business and Enterprise plans give administrators control over users and settings, and OpenAI states it does not train its models on business customer data by default.
Anthropic’s Claude Team and Enterprise plans offer similar administrative controls, and Anthropic states commercial customer data is not used for model training by default. Claude is a common choice for long documents and careful writing.
Google’s Gemini for Workspace runs inside Google Workspace with administrative controls managed in the same console as Gmail and Drive. A natural fit for companies already on Google.
The free and personal versions of these same tools come with different terms, and data handling depends on the settings each user chooses. That difference is exactly why the approved list matters. Vendor terms also change, so review them at least twice a year.
What a policy people follow includes
The best AI policies we see fit on one or two pages and answer practical questions.
Which tools are approved, and how do I get an account? Name them. Make access easy.
What data can go where? Use three tiers. Green data, such as public information and general writing, can go into any approved tool. Yellow data, such as internal documents and non-sensitive business information, goes only into approved business accounts. Red data, such as customer personal information, financial records, health data, and anything under a confidentiality agreement, stays out of AI tools unless a specific, reviewed system is built for it.
What work is encouraged? List real examples: first drafts, meeting summaries, spreadsheet formulas, research. Employees follow rules more willingly when the rules also tell them what they can do.
What needs human review? Anything going to a customer, anything with a number in it, and anything legal or contractual gets read by a person before it leaves the building.
How do I request a new tool? Give a named owner and a turnaround time. Two weeks is reasonable. Six months guarantees workarounds.
Who does this apply to? Everyone. Say so in writing, and say leadership is included.
Leadership has to go first
That 81% figure deserves its own section. When employees believe executives use whatever AI tools they like, the policy loses its authority overnight.
The fix is visible. Executives use the same approved accounts as everyone else. They follow the same data tiers. When a leader needs an exception, it goes through the same request process, and the outcome is shared. Policies spread through example faster than through email.
How this works at a real company
Take a 200-person distributor in Green Bay. Its current AI policy is one paragraph from 2024 that bans “unapproved generative AI.” Inside sales reps paste customer emails into free chatbots to draft replies. The controller uses a personal account to summarize contracts. Everyone knows the policy exists. Nobody follows it.
The company replaces it with a two-page policy. It buys Microsoft 365 Copilot licenses because the team already runs on Microsoft, plus a small number of Claude Team seats for the contracts and proposals group. It publishes the three data tiers with ten examples specific to the business. It names the IT manager as the tool request owner with a ten-day turnaround. The CEO announces the policy by walking through how she uses Copilot to prepare for board meetings.
Within a quarter, use of personal AI accounts drops, because the approved tools are easier and better. The company now knows which tools are in use, what data goes into them, and who to call when something goes wrong.
Review it every six months
An AI policy written today will be out of date by spring. Vendors change terms, new tools ship, and your team finds new uses. Put a review date on the policy itself, ask employees what is not working, and treat each revision as a sign the policy is alive rather than a sign it failed.
Where Earthling fits
Earthling Interactive helps mid-market companies in Madison, across Wisconsin, and throughout the continental U.S. put AI to work responsibly. As an AI consulting and AI development partner, we help leaders choose and configure the right tools, write policies employees actually follow, set data rules that match the business, and build custom AI systems for the work that needs more than an off-the-shelf tool. Reach out and let’s write yours.
