Somewhere in your business today, someone is pasting a client file into a free AI tool to get a quick summary. They are not doing anything wrong. They work hard and they are in a hurry. And you do not know about it. That gap is what an AI policy template has to close.
That is shadow AI: AI use that happens without the organisation knowing about it or approving it. It is not sabotage and it is not carelessness. It is simply someone reaching for a tool because there is no other one.
In the Netherlands the term only showed up in search data at the start of 2026, and it has grown quickly since. Over the same period, search volume for AI-beleid (Dutch for AI policy) rose from about 110 to 170 a month. Those are two sides of the same moment: businesses notice that something is happening and look for a way to bring order to it.
Why this is not an IT problem
The reflex is a ban. It does not work, for the same reason it did not work with shadow IT: you ban the use, and the need stays.
An employee who spends four hours on work that could take twenty minutes will not go back to the slow way because a rulebook says so. They will just stop telling you. You have not removed the risk, you have made it invisible. And invisible risk is worse than visible risk, because you can no longer do anything about it.
So the useful question is how to make the right route as easy as the quick one.
What we are talking about, in numbers
1 in 6
Dutch businesses use AI, a doubling in two years. This is the use that gets measured; shadow AI is by definition not included.
95%
of business GenAI pilots deliver no measurable result on the profit and loss account. The researchers point to integration into work processes, not to the models.
Source: CBS (Statistics Netherlands), 2025 (AI use by Dutch businesses); MIT NANDA, The GenAI Divide: State of AI in Business 2025, August 2025, research covering 300 implementations and 150 executives.
That 95 percent figure is often misused
You come across it everywhere, usually as proof that AI does not work. That is not what it says.
- It is about generative AI pilots in general, not about AI agents and not about workflow automation.
- It measures no demonstrable return, not ‘failed’ or ‘stopped’. That is a real difference: many pilots were never measured.
- It is not a Gartner figure, although it is often mentioned in the same breath as Gartner’s forecast on agentic AI. Two studies, two subjects.
I mention it here because the explanation is what matters, more than the percentage. The researchers found that successful implementations sit deep inside existing work processes. Shadow AI is by definition the opposite: a loose tool next to the process, with no connection to the systems where the work actually happens.
What it actually costs you
Four costs, and the last two are rarely counted.
1. Data you cannot get back
As soon as a client file is pasted into an external service, that is processing of personal data. Under the GDPR you need a lawful basis for it, an entry in your record of processing activities and usually a data processing agreement. Without those, you cannot show that everything is in order when a client asks, or when the Dutch Data Protection Authority (Autoriteit Persoonsgegevens) does.
The point is not that something will necessarily go wrong. The point is that you do not know, and you cannot back it up.
2. Tenders and supplier assessments
More and more clients ask during procurement which AI systems you use and how you keep them under control. Answer ‘I’m not quite sure’ and you are out. Not because it is unsafe, but because you cannot show that it is safe, and buyers have to justify their choice.
3. Knowledge that leaves with the employee
This is the underestimated one. Someone who has spent months building their own way of working with AI, with prompts that work exactly for your type of file, takes that knowledge with them when they leave. Nothing is in a system, nothing is documented. What has been built in your own environment stays.
4. Twenty solutions to the same problem
Without an overview, every employee solves their own version of the same bottleneck. Five people all summarising quotes by hand with five different tools: that is paying to learn the same lesson five times, for something that could have been built properly once and then worked for everyone.
Example: a consultancy finds out what is already running
EXAMPLE 1
The situation. A consultancy with fourteen employees suspected ‘some AI use’ and wanted to draw up a policy. The management assumed it involved two or three people.
The approach. Instead of writing a rulebook, they took stock first. One question per employee, explicitly without consequences: which tools do you use, and what for? The answers were collected by someone without a management role.
The result. Eleven of the fourteen were using something. Nine different tools in total, three of them with client data. Two employees had developed a way of working that was so good it was then adopted across the company. Without the inventory it would never have come to light.
What was disappointing. The inventory took four hours and caused discomfort straight away: two people felt caught out, despite the promise made beforehand. If you do this, you have to keep that promise. One reprimand and the next inventory will only get you the answers people think you want to hear.
Illustrative example based on typical SME volumes, not a client case. In the free assessment we work through your own numbers.
An AI policy template that does not drive people away
Most AI acceptable use policies are legal documents that nobody reads. What works fits on two sides of A4 and answers four questions.
- What may I use? A list of approved tools. Not ‘everything IT allows’, but actual names. Short is fine; long is better than empty.
- What must never go in? This is the most important rule and the only one you make strict. For example: no personal data of clients, no medical or financial data, no source code, no third-party contracts. Be specific: ‘confidential information’ is not a usable boundary.
- Who decides on something new? One name, one way of working, and a commitment on response time. Anyone who has to wait a week for an answer will not ask next time.
- What do I do when in doubt? The answer has to be: ask, without consequences. A policy that punishes doubt does not produce care. It produces silence.
The sentence with the most effect
Put this in your policy word for word, at the top:
“Using something that isn’t listed here and it helps your work? Let me know. Then I’ll see whether I can add it. You won’t get any grief for it.”
This one paragraph decides whether your policy brings in information or silence. Every other provision comes second to the question of whether people still tell you things.
AI acceptable use policy: a template to copy
Here are the four questions above as a template you can paste into a document. Fill in the parts in square brackets and keep it this short.
AI acceptable use policy for [company name]
TEMPLATE
Before you write it. Ask every employee one question, explicitly without consequences: which tools do you use, and what for? Have someone without a management role collect the answers. Use what comes back to fill in section 1.
At the top, word for word. “Using something that isn’t listed here and it helps your work? Let me know. Then I’ll see whether I can add it. You won’t get any grief for it.”
1. Tools you may use.
- [Tool name]: for [what it is used for]
- [Tool name]: for [what it is used for]
- Something not on this list? See section 3.
2. What never goes into an AI tool. This is the one hard rule in this policy.
- Personal data of clients
- Medical or financial data
- Source code
- Contracts from third parties
- [Anything else specific to your business, named concretely]
3. Who decides on a new tool. [Name] decides. Send your question to [email address or channel]. You get an answer within [number] working days.
4. If you are in doubt. Ask [name] first. Asking never has consequences.
Two sides of A4 at most. A longer policy does not get read, so it does not get followed.
When your own environment is the answer
Self-hosting is often presented as the solution to shadow AI. It only is under one condition: your own environment has to be as easy to use as the free tool.
If it is slower, uglier or more cumbersome, people will not use it. Then you have an expensive environment and shadow AI as well, and that is the worst of all outcomes.
FLOW 1
Internal knowledge assistant in your own environment
Who it is for: Organisations where employees search their own documents every day.
What it gives you: Employees get answers from your own material, with a reference to the source document. The data does not leave the environment and every question can be traced.
Watch out: It stands or falls with the quality of your documents. An assistant that draws on outdated manuals gives wrong answers with great confidence, and that does more harm than having no assistant at all.
Note that the last step in that chain is not a technical one. Not found? Say so honestly. An assistant that always answers something will, within a month, undermine the trust of exactly the people who are meant to use it.
When you do not need to do anything with this yet
- You work alone. Then there is no shadow. Do write down for yourself what you put in where; that helps with the first client who asks about it.
- You process no personal data and no confidential information from third parties. The risk is then largely theoretical.
- You have no clients who check on this. Look at it again in a year; this question is shifting quickly.
For everyone else: the inventory takes half a day and gives you information you do not have now. That is the cheapest step in this whole subject.
In short
- Shadow AI is not disobedience. It is a sign that the alternative on offer is missing or too cumbersome.
- Banning moves the risk out of sight. Taking stock makes it manageable.
- Start with questions, not with a rulebook. Four hours of inventory gives you more than ten pages of policy.
- A policy only works if reporting stays free of consequences.
- Your own environment only solves it once it is as pleasant to use as the free tool.