Rolling out AI rules employees will actually follow

Why policies are not followed

When an employee pastes a customer's contract into a personal chatbot, it is rarely defiance. Usually one of two things is true. They did not know the rule, because the rule lives in a document they were never shown. Or they knew, and the approved way to get the job done was slower, unclear or did not exist.

A rollout that only announces the policy fixes neither. A rollout worth the name does five things, in this order: tells people what is happening, shows them what they may use, gives them a way to ask for more, asks them to confirm they have read the rules, and teaches the few things that matter.

It helps to have written a policy people can read. If yours runs to twenty pages of defined terms, start with how to write an AI acceptable use policy, or generate a draft with the free policy generator.

Step 1: Tell people before anything changes

Announce the policy before it takes effect, and separately from any monitoring. The announcement should fit on one screen and answer three questions: what is changing, what do I need to do, and where do I ask.

If you are going to monitor which AI tools are used, this is the moment to say so, in writing, and to say exactly what is and is not recorded. In several places it is also the law. New York, Connecticut and Delaware require employers to give notice of electronic monitoring; we cover the details in employee AI monitoring notices. In Germany and France, monitoring that can observe employee behaviour is a matter for the works council or the CSE before it starts.

Legal duty aside, notice is what makes monitoring tolerable. People accept "we record which AI tools are used, never what you type" when they hear it from their employer first. They do not accept discovering it. We go further into this in monitoring employee AI use without surveillance.

Step 2: Lead with what is allowed

The most-read part of any AI policy is the list of approved tools. Make it the first thing people see, and make each entry useful:

• The tool, by the name people know it by.
• What it is approved for: "drafting and research", "code completion for the engineering team".
• What must stay out of it: "no customer personal data".
• How to get access: "sign in with your work account", or who to ask.

An employee who can see that ChatGPT is approved in the company workspace, for drafting, signed in with single sign-on, has no reason to open a personal account. An employee facing a page of prohibitions and no alternatives will improvise.

Keep the list where people will look for it. A page they can open from a browser bookmark or from the tool that warned them beats a PDF attached to an email in March.

Step 3: Give people a route for everything else

No approved list is complete. New tools appear every month, and the person who finds one useful is usually right that it is useful. What matters is what happens next.

A request route needs three properties:

It is where the list is. The button to ask for a tool sits beside the list of tools you may use.

It asks for what a decision needs. Which tool, what for, and what kind of data would go into it. Not a business case.

It answers quickly, with a reason. Put a number on it in the policy: ten working days is defensible, three is better. When the answer is no, say why, and point to the approved alternative. A declined request with a reason and an alternative is a good outcome. A request that vanishes is how shadow AI starts.

On the reviewing side, count demand. Five people asking for the same transcription tool is one decision, and a signal about what the approved list is missing. Before you approve anything, read the vendor's terms for the plan you would hold: whether it trains on your data, how long it keeps it, and whether people can review it. Our vendor fact pages are a starting point.

Step 4: Ask for acknowledgement, and record it

Acknowledgement is a small act with two purposes. It makes it likely the person read the policy once. And it gives you a record: this person confirmed this version on this date.

Three details make the record worth having:

• Tie it to a version. When the policy changes materially, the old acknowledgement no longer covers it. Ask again.
• Make it one action. Read, then one button. Every extra step loses people.
• Track coverage, not culprits. "148 of 212 people have acknowledged version 2" is the number to manage and to report. Chase the gap by team and by reminder, not by naming individuals in a slide.

Customers ask about this directly. "Are employees required to acknowledge your AI policy?" is a standard questionnaire item, and a percentage with a date is a far better answer than "yes".

Step 5: Teach the four things that matter

Training on AI does not need to be long. It needs to cover the mistakes people actually make:

1. 01What generative AI is, and where it fails. It produces plausible text, not checked facts.
2. 02What must never go into an AI tool. Credentials, customer personal data, anything under a confidentiality agreement, unless the tool is approved for it.
3. 03Checking outputs, and who is accountable. The person who sends it owns it.
4. 04Spotting a risky tool and asking for approval. Including the AI features that appear inside software you already use.

Fifteen minutes, with a few questions after each part, is enough for most staff. People who build or operate AI systems need more.

What the EU AI Act asks for

If you operate in the European Union, Article 4 of the AI Act is the reason this step is not optional. The duty was softened in 2026, and it is worth being exact about what it now says. Our regulation tracker records it this way:

Two things follow. The duty is to take measures, so a course with completion records is a reasonable measure to point to. And no particular level, format or certificate is mandated, so nobody can sell you "Article 4 compliance" in a box. Read Article 4 and decide with counsel what is proportionate for your staff and the systems they use.

Outside the EU the same practice is simply good governance. The NIST AI Risk Management Framework puts workforce training and clear accountability in its GOVERN function.

Step 6: Show people what is recorded about them

If you monitor AI tool use, give each person a way to see what the monitoring holds on them. It is the strongest reassurance available, because it replaces a promise with something they can check. "Here is what we collect, here is what we never collect, and here is what was recorded about you in the last thirty days: ChatGPT, fourteen hours."

It also keeps you honest. A monitoring programme you would be uncomfortable showing to the people it covers is one to redesign.

Measuring whether it worked

Four numbers tell you, a quarter later, whether the rollout took:

• Acknowledgement coverage for the current policy version.
• Course completion.
• Requests received, and time to decision. No requests at all is a bad sign: it means people are not asking.
• Use of unapproved tools, as a share of AI use. If the first three are healthy this falls on its own.

Report them to whoever owns the policy. They are also most of what a customer's questionnaire will ask.

How GhostAudit helps

The employee AI hub in GhostAudit is one page for each employee: the approved tools, a request form, the policy to acknowledge, a short course and a plain account of what monitoring records about them. Acknowledgements and completions are recorded and can be cited as evidence. You can look at a live example for a fictional company.