Chances are someone at your company has already pasted a contract, a client list or a chunk of code into an AI assistant using their personal account. They do it because it saves them an hour and because nobody has told them how to do it properly.
Banning AI doesn't work: people use it anyway, in secret and with worse tools. What works is a short policy that people understand and that says what they can do.
What it should say
Which tools are approved
A closed list, with the name of each tool and always the business version. The difference matters: the free or personal accounts of many assistants may use what you type to train their models, and business plans usually rule this out in the contract. Anything that isn't on the list isn't used with company data.
What data you can give an AI
This is the section that clears up the most doubts. A classification with just a few levels works well:
- Public information: no restrictions.
- Internal information, such as procedures or templates: only in the approved tools.
- Personal data about clients or employees: only if the provider has signed a data processing agreement, as the GDPR requires. If in doubt, ask first.
- Passwords, keys and non-public financial data: never.
Personal accounts
Using a personal account on an AI assistant for company work has to be explicitly banned. It's the easiest data leak to commit and the hardest to spot.
Who is accountable for the output
The AI helps and the person signs off. Anything that goes to a client or out under the company's name is reviewed by whoever sends it, and they answer for it just as if they had written every word.
Uses that aren't allowed
The EU's Artificial Intelligence Act, Regulation (EU) 2024/1689, prohibits certain practices. They include manipulating people with subliminal techniques that cause them harm, social scoring, and inferring people's emotions at work, except for medical or safety reasons. One line in the policy reminds everyone that these uses are ruled out, whatever the tool.
Keys and access
AI accounts are set up with the corporate email address. API keys are kept in the password manager and never in a document, in code or in a chat.
How a new tool gets approved
New tools appear all the time. The policy says who to ask for one, what gets checked (that it has a business plan, that it doesn't train on your data and that the provider signs the data processing agreement) and who signs it off.
What to do if something goes wrong
Anyone who thinks they've given an AI data they shouldn't have needs to know who to tell, and to be sure that reporting it won't get them into trouble. It's handled like any other security incident.
Training is part of the rules
The same regulation requires companies that use AI systems to take steps so that their staff learn to use them. It doesn't set a level or say how to do it, but with a policy nobody has read it's hard to argue that you comply.
A session when people join and a reminder once a year is a reasonable minimum. Better still if the session uses your company's own cases: what you can ask the AI to do in administration, in sales or in support.
What a document can't do on its own
A policy doesn't stop someone opening a tool that hasn't been approved. That's what technical controls are for: limiting from the network which assistants can be reached, and monitoring what data leaves the company. That's the territory of data loss prevention.
Short enough to read in one sitting
Nobody reads a twenty-page policy. Ours fits in a few pages, and we review it when the regulations or the tools we use change.
If you'd like to start from a template and adapt it to your company, we cover that in Safe AI and training. If what you need is to fit your use of AI into the EU regulation, that's Responsible AI and the AI Act.
Shall we look at your case?
Tell us what you have today and what you would like to have.