Why a ban doesn't work
A ban answers the wrong question. It asks how to stop the use. The question worth answering is how to bring it inside a frame you can stand behind.
The phone is next to the keyboard. A firewall blocks the office network. It does nothing about a private phone on mobile data. Anyone who has seen ChatGPT save them twenty minutes will photograph the email or retype it on their phone. The data still leaves the building, and now you see even less of it.
A ban ends the conversation. While the use is tolerated, people talk about it. They swap tricks at the coffee machine. After a ban, they stop telling you. You lose the exact information you'd need to write a sensible rule.
Your people solved a real problem. The order desk found a task where a language model helps. That's worth keeping. The account was the mistake, and the idea was fine. A good rule keeps the time saved and swaps out the account.
Find out what's actually in use
You can't write a rule for usage you haven't looked at. Management's gut feeling tends to be off here, in both directions.
Ask, don't hunt. Don't go through browser histories and don't have IT trawl the logs. It burns trust, and in Germany it can also get you into trouble with data protection rules and with the works council. Ask openly instead. Which AI tools do you use for work? What for, exactly? On which device and which account? What kind of data goes in?
Declare an amnesty. Put in writing that nobody gets in trouble for an honest answer, then stick to it. If the first person who owns up gets a formal warning, you'll never get an honest answer again.
Ask what it's good for. The most useful question goes last. Which task does the tool handle for you, and how much time does it save? The answers tell you what your sanctioned tool has to do well. If it can't draft that complaint reply as fast as the private account, the private account wins.
One tool, one contract, short rules
This is the part that actually ends shadow AI.
One sanctioned tool with a business contract. Pick one tool. One, not five. What matters is the contract behind it. As soon as staff enter personal data, the provider processes that data on your company's behalf, and Article 28 of the GDPR requires a contract for that, the data processing agreement. Also check what the contract says about using your inputs to train models and where the data is stored. A free private account gives you no such contract. Which ChatGPT plans come with one, and what your company still has to do afterwards, is covered in Is ChatGPT GDPR compliant for business?
The rule stays short. Nobody reads twelve pages. The core is three blocks. What may go in, such as your own drafts, public information and anonymised examples. What never goes in, such as passwords, health data, personnel files and any customer data on a private account. And what needs a question first, such as contracts, quotes with pricing terms and applicant data. Use examples from your own company in every block, not abstract categories. Clause 4 of the AI usage policy template has wording you can copy for what never goes in and what needs a question first. With the other nine clauses, the whole policy fits on two pages.
A name for questions. Under the rule goes a person with a phone number. In a 50-person company it's often whoever is already most interested in the topic. They don't need to be a lawyer. They need to know who they can ask when they're unsure themselves.
Better than the private account. The company tool has to be more convenient than the private one or you lose. Give licences to the people who actually use it and don't make them wait months for approval. Set up sign-in with their company account. Build ready-made templates for the most common tasks from the survey, like the complaint reply or a summary of a supplier contract. Once the company tool gets someone to a result faster, they stop opening the private app.
A short introduction to the tool also counts as one of the AI literacy measures that Article 4 of the EU AI Act asks for. That's one more reason to run it properly and keep a list of who attended.
Bring the works council in early
This section is for companies in Germany that have a Betriebsrat, the elected works council that German labour law gives co-determination rights. Austria and Switzerland have their own rules. It isn't legal advice. It marks where people trip.
What the law says. Under section 87, paragraph 1, no. 6 of the Works Constitution Act, the BetrVG, the works council has a right of co-determination over the introduction and use of technical devices designed to monitor employees' behaviour or performance. The Federal Labour Court reads "designed to" broadly. In its decision of 8 March 2022 on Microsoft Office 365, it held that it is enough for a device to be objectively suitable for collecting behaviour or performance data. Whether you intend to monitor anyone doesn't matter.
What that means for AI tools. On 16 January 2024, the Hamburg Labour Court decided in summary proceedings that merely allowing staff to use ChatGPT through their own private accounts in the browser did not trigger co-determination under no. 6. The employer had no access to the usage data. That changes once you roll out a company account with an admin console where usage, histories or per-person reports may be visible. At that point there's a strong case that the works council has a say. The Hamburg ruling is a single first-instance decision, and I wouldn't base a rollout on it alone.
Talk early instead of negotiating late. Bring the works council to the table before you buy the tool, and show them the survey results. Agree together which usage data the tool stores, who can see it, and that it won't be used to assess performance. That belongs in a Betriebsvereinbarung, a formal works agreement. Which rights the works council has and what such an agreement covers is in the article on AI and works councils in Germany. A works council that was involved from the start rarely turns into the brake. One that learns about it from the all-staff email can stop a rollout that needs its consent until the two sides agree.
Shadow AI tells you that your people want to work faster than your tools let them. A ban punishes them for it and pushes the customer data further onto their phones. Give them a company tool that beats their private account, and the usage comes back on its own to a place where you can see it and answer for it.