Whether AI assistants are used in your business has usually been decided already. Sales have quote texts drafted for them, accounts summarise long emails, someone translates an enquiry from a customer abroad. The only open question is whether this happens on accounts the business is responsible for, and under rules everyone knows.
This article shows what to settle before you officially introduce ChatGPT, Microsoft Copilot or another assistant. It does not replace a legal assessment. Clarify that with your data protection officer and, where needed, a lawyer (as of September 2026).
Why the question comes up now
If you do not give your staff a tool, you still get AI in the business, just on personal accounts. That is not misconduct by your staff but a gap in what the business provides. They want to do their work well and quickly, and they reach for what is available. More on this in Shadow IT and shadow AI.
The good news: make the use official and you get both. Staff keep a tool that helps them, and the business knows where its data goes.
Personal account or business account
The most important step is also the simplest: move away from personal accounts to access that the business signs up for and manages. The large providers have business offerings for this, OpenAI for ChatGPT, Microsoft for Copilot as part of Microsoft 365, Anthropic for Claude.
What a business account changes compared with a personal one:
- The business has a contract with the provider. This usually includes a data processing agreement. With a personal account the business has none at all.
- Use of your input for training. With business offerings, providers generally commit not to use your input to train their models. Check this in the contract, not in the marketing.
- The business manages access. When someone leaves, their access is blocked. Their history does not stay behind on a personal account.
- You know where the data is stored. Some offerings let you keep the data in Europe. Where the computers that generate the answers are located is a separate question.
For particularly sensitive data there is another route: a model running in-house. What that takes is described in Running AI in-house.
Which data may go in
Staff need a simple answer to the question of what they may type in. A split into three levels, agreed with your data protection officer, has proven useful:
| Level | Examples | Rule |
|---|---|---|
| Harmless | your own drafts without names, public texts, product descriptions | may go in |
| Internal | quotes, minutes, customer enquiries with names | business account only, leave names out where possible |
| Off limits | personnel files, health data, passwords, login details, confidential contracts | must not go in |
What matters is not the perfect list, but that everyone knows it and can ask someone when in doubt.
Copilot sees what the employee may see
With Microsoft 365 Copilot there is one more point that is often overlooked. Copilot searches the files, emails and Teams chats that the employee in question has access to. That is intended and useful. It becomes a problem when sharing in the business has grown too generous over the years.
If the salary list sits in a SharePoint folder that everyone can open by mistake, nobody has found it so far. Copilot finds it as soon as someone asks about salaries. So tidy up the sharing before you introduce it: who may see which folders, and does that match reality? Why permissions have to be checked in the system itself is explained in Roles and permissions.
Rules for your staff
A good set of rules fits on one page. At the very least it should answer:
- Which tool is approved, and which account is it used with?
- Which data may go in, according to the levels above?
- What is it for? Drafting texts, summarising, translating, collecting ideas. Not for decisions about people, for example on job applications.
- Who checks the result? Always the person who uses it.
- Who do I turn to with questions or when something has gone wrong?
Write the rules as instructions, not as a list of bans. Anyone who only reads what they must not do goes back to their personal account.
Checking remains a human job
An assistant writes well, but it does not know whether something is true. It can copy figures wrongly, invent a deadline or get the tone of a translation wrong. So the rule is: whatever the assistant writes is a draft. Whoever sends it has read it and is responsible for it.
This applies above all to anything that leaves the building and to anything with figures in it. The limits of language models are described in What AI cannot do for your business.
Briefing staff is part of it
Since February 2025 the European AI Act has required companies that use AI to do something for the AI literacy of their staff. In practice this means: whoever approves a tool also explains how to use it sensibly and where its limits are. What exactly this involves and what changed in 2026 is described in AI literacy under the AI Act.
How to proceed
- Ask in each department who uses which AI tools today, and for what.
- Choose a tool that fits the most common tasks and sign up for a business account.
- Agree with your data protection officer which data may go in.
- Tidy up the sharing in your file storage before an assistant searches it.
- Put the rules on one page and brief your staff.
- After a while, ask what helps and what is missing.
An approved tool with clear rules is safer than any ban. And it is often the first step towards more: an assistant that can also query your own systems. How we go about this is shown on our AI Consulting page.
