Can your employees use ChatGPT with client data? In many European SMEs it is already happening: someone pastes an email, a contract or a spreadsheet into ChatGPT, Copilot or Gemini. There is usually no bad faith. What is missing is a picture of real use. This guide explains what may be at stake and where to start — with no certificate and no legal classification.
The case (fictional, but everyday)
Marta handles client management in a company of 40 people.
A client sends a long, confusing email asking for changes to a contract. Marta copies the whole message — name, company, commercial terms and some personal contact details — and pastes it into ChatGPT with the prompt:
“Summarise this and tell me what they are asking for.”
It takes two minutes. She does not think twice. Anyone would do the same.
Nobody in the company told her she could not. Nobody told her she could. The topic was never discussed.
This scenario — or a variant of it — may be happening right now in many companies. Not out of bad faith. And not because employees want to break the rules.
Generative AI has entered daily work much faster than many organisations have been able to set policies, criteria and controls. That is where the problem starts. It has a common name: Shadow AI — AI use in the team that the organisation does not see, did not choose and has not documented.
What may be at stake, even if nothing shows
When Marta puts client information into an AI tool, several issues may be potentially applicable at once. This is not an automatic classification: it depends on the tool, the type of data and how the use is configured. They need review.
1. The data may leave the company environment
When an employee puts personal data into an external AI tool, that data may then be processed by the tool’s provider.
Questions many companies have not yet asked:
Which tool is the team using? Under what terms? Where does the processing take place? What happens to the information entered?
Depending on the tool, its settings and the terms of service, GDPR rules on processors — and, in some cases, international transfers — may be potentially applicable. That requires review case by case: a free personal account is not the same as an enterprise plan with a data processing agreement (DPA).
Usefulness is not enough. You need to know what happens to the information you put into the tool. If you buy ChatGPT Enterprise, Copilot or Gemini, the next step is usually the SaaS contract clause checklist.
2. The company is not off the hook because an employee did it
Marta made the decision on her own. That does not, by itself, take the organisation out of the picture.
If the organisation allows — or does not effectively prevent — the team from using AI tools with client information, a recommended measure is to have written down which uses are allowed, which information may be entered, and what to do if in doubt.
From a data-protection perspective, where personal data are processed the controller usually needs to be able to show that it has adopted appropriate measures (accountability). That is not a “compliance” certificate. It is being able to explain what was considered and what was documented.
Saying only “an employee did it” is usually not enough if there are no instructions, limits or visibility of use.
3. There may be no traceability at all
Imagine that six months later a client asks what happened to their data.
Can the company know whether any employee put that client’s information into ChatGPT? Which information? Who? With which tool? On which date?
In many organisations the answer would be: we do not know.
If the company cannot even see how the team uses AI tools, showing that appropriate measures were adopted is much harder. The first step is not an audit: it is a snapshot of real use in five questions.
4. The AI Act may also enter the conversation
Data protection is not the only piece. Regulation (EU) 2024/1689 (the AI Act) provides, in Article 4, for a duty to take measures to support AI literacy. That provision applies to providers and deployers of AI systems: you first need to determine whether the organisation falls into those categories. It is not an individual result obligation, and it does not require a specific course.
In practice, a company should not treat AI use as a purely personal decision of each employee. It is useful to know which tools the team uses, for what, and what risks that use may involve. Calendar and framework: AI Act for companies (2026–2028 guide). What a broader review includes: AI legal risk assessment.
Can a company ban employees from using ChatGPT?
Yes. A company can set limits on AI tools and even prohibit certain uses when there are reasons to do so.
Banning ChatGPT entirely is not always the most practical answer. The real problem is not an employee using ChatGPT to improve a text, summarise public information or generate ideas.
The problem appears when they enter information the company should not share with an external tool without having reviewed that use first.
The useful question is not “can my employees use ChatGPT?”. It is: for what, with which information, and under which rules?
What usually does not happen (and why that is misleading)
None of this usually produces an immediate consequence. No alarm goes off. No fine appears automatically. Marta’s client probably never finds out.
The lack of visible consequences creates a false sense that “nothing happens”. That nothing has gone wrong yet does not mean the use is under control or that the risk is not there.
It is enough for a client to ask about their data, for an incident to occur, for a third party to request guarantees, or for a new client to demand documentation. At that point, having no documented criteria can turn a one-off use into a larger problem. Six signals that an AI legal assessment is worth considering.
What a company should do (recommended measures)
The solution is not necessarily to tell Marta to stop using ChatGPT. AI tools can add real productivity; an absolute ban sometimes only pushes use out of sight.
What matters is a minimum of rules and of evidence that the use has been thought through.
1. An internal AI-use policy
It does not have to be a 50-page document. It should make clear, at least:
- Which AI tools the team may use.
- Which uses are allowed.
- Which information may be entered and which should not.
- What happens with client data.
- Who may use certain tools.
- What to do when there is doubt.
That is a recommended measure, not a certificate. It fits what an AI Legal Assessment documents when the five-question signal is critical or moderate.
2. See which tools the team actually uses
Picking an “official” tool is not enough. First you need to know what staff are using. There may be ChatGPT, Gemini, Copilot or others on personal accounts, with no inventory.
Before writing the policy, get a picture of real use. Start here: Does your company use ChatGPT? 5 questions.
3. Review provider settings and terms
AI tools do not all work under the same conditions. An enterprise plan is not a personal account. If the team uses AI routinely, review the plan, how data are processed and what the provider actually guarantees. Checklist: AI clauses in SaaS contracts.
4. Train the team (literacy adapted to context)
A policy nobody understands is of little use. Employees should at least know what information they may enter, what to avoid, and when to ask. You do not need to turn the whole workforce into AI specialists. You need them to know where the limits are, in the company’s real context.
5. Leave a documented criterion
The company should be able to answer a simple question: what have you done to see and bound AI use? And have an answer with evidence (policy, inventory, training, contracts). That is not bureaucracy for its own sake. It is being able to explain the documented measures.
The problem is not ChatGPT. It is not knowing how it is being used
Most companies do not need to stop using artificial intelligence. They need to know how they are actually using it: which tools, which information, which risks and which minimum rules.
That is Legal Stones’ starting point. The five questions give a snapshot of use. They are not a legal opinion and not a certificate. If the signal is critical or moderate, the next step is the AI Legal Assessment: identified use, needs and a documented plan.
This is not about banning ChatGPT. It is about knowing what your team may do with AI, what they should not do, and what is worth documenting.
Because the problem is not that your employees use AI. The problem is that nobody knows exactly how they are using it.
Frequently asked questions
Can your employees use ChatGPT with client data?
They may already be doing it, with or without permission. Pasting client data into ChatGPT, Copilot or Gemini may involve processing of personal data and, depending on the tool, other potentially applicable rules. There is no automatic yes or no for every company. The first step is to see real use: five free questions. Not a certificate.
Is it lawful to paste a client email into ChatGPT?
It depends on what data it contains, which tool and plan you use, and whether there are internal instructions. An email with a name, company and contact details usually includes personal data. It is better not to treat that as harmless by default. It requires review in your company’s context.
What is Shadow AI?
It is the use of AI tools (ChatGPT, Copilot, Gemini, personal accounts) that the company has not inventoried or governed. It is not a legal classification. It is a signal that visibility is missing. Legal Stones’ five questions are designed to surface it.
Should we ban ChatGPT at work?
Not necessarily. A total ban sometimes only pushes use onto personal accounts. The usual approach is to bound it: for what, with which information and with which tool. A short policy and a snapshot of real use are often more useful than an absolute prohibition.
Where do I start if I do not know what the team uses?
With the five questions (ChatGPT, client data, policy, inventory, personal tools). Then, if needed, an Assessment and, if there is a provider contract, the SaaS clause checklist.
Does your company use ChatGPT? 5 questions →
6 signals your company needs an AI Legal Assessment →