Short answer: Working with AI in line with the GDPR succeeds in everyday practice when you know which data flows where, process only what is necessary and give your team clear rules. What counts is daily practice: no personal or sensitive data in unvetted tools, prompts formulated without personal references wherever possible, and results checked before they are used. This guide shows what that looks like in concrete terms — the binding assessment in any individual case belongs with your data protection officer or your legal advisers.
Why data protection comes first in everyday AI use
AI tools are only a gain if their use is legally sound. Data protection is therefore not a brake but part of a clean way of working — and it builds trust with customers. The greatest risk in day-to-day work is not the tool itself but the way it is handled: the uncontrolled entry of sensitive data into unvetted systems, often out of convenience or ignorance.
This guide deals with ongoing operation: the question of how your team works day after day with AI tools that have already been introduced. How you select new tools, vet providers and conclude data processing agreements (Art. 28 GDPR) is covered separately in our article on introducing GDPR-compliant AI tools. Put briefly: that piece is about “which tool?”, this one about “how to work with it?”.
Which data may go into an AI tool — and which may not
The principle of data minimisation is the most important everyday rule: enter only the data that is necessary for the task. What that means in practice can be set out in three tiers:
- Uncritical, as a rule, are purely factual contents with no personal reference — general text drafts, product descriptions, code without access credentials, or publicly available information.
- Only into vetted tools belong personal data such as names, contact details or customer histories. “Vetted” means: the provider processes the data in a way that can be traced, a data processing agreement is in place, and your data is not used for training.
- Fundamentally out of place in general AI tools are particularly sensitive data: health data, salary information, job applications, trade secrets, access credentials or customers’ contract details — unless the tool has been expressly approved and secured for that purpose.
A simple rule of thumb helps the team here: what you would not email to an unknown external service provider does not belong in an AI tool unchecked either.
Formulating prompts without personal references
Many tasks can be completed without personal data leaving the company. The key lies in how the input is formulated.
Anonymise and use placeholders
Replace names, companies and identifiers with neutral labels before you put a text into an AI tool. “Complaint from Mr Müller of Company X about invoice 4711” becomes “complaint from a customer about an invoice”. For most writing and analysis tasks the result loses nothing in quality — you add the specific details back in only after the output.
Context instead of raw data
Often the model does not need the original text at all, only the task: “Draft a friendly reply to a complaint about a delayed delivery” works without any customer data whatsoever. Ask yourself before every input: does the tool really need this information in order to solve the task?
Check outputs before they leave the building
Everyday practice includes the other direction as well: AI supplies suggestions, and a human takes responsibility for them. Check results for factual accuracy — and for whether internal information has inadvertently found its way into a text that is going out.
Internal rules that work in everyday practice
Data processing agreements under Art. 28 GDPR and clear internal policies give your team confidence. For an AI policy to be lived day to day, it should be short and answer concrete questions: which tools are approved? Which data may go into them? Who checks results before they are used? Who do I turn to if in doubt?
This practical checklist is a suitable starting point:
- Is it clear where the data of the approved tools is processed?
- Is a data processing agreement in place for every tool that involves personal data?
- Is there an internal AI policy that everyone knows and can find?
- Has the team been trained — not just once, but refreshed as needed?
- Are sensitive data ruled out by technical or organisational means?
The culture behind it matters too: anyone who reports an accidental wrong entry should get support, not blame. That is the only way you will find out where things go wrong in daily work. And expect staff to fall back on private tools if the approved ones are too cumbersome — the best rule is the one that describes the simplest permitted route.
Typical breaches in everyday work — and how to avoid them
Most data protection problems with AI arise not from bad intent but from routine. These are the patterns we see again and again:
- The entire email thread in the prompt: instead of the actual question, the whole exchange is pasted in, names and contact details included. Remedy: state only the task, anonymise the thread.
- Private account instead of company access: staff use the free consumer version of a tool because they are familiar with it — without any contractual framework and often with training use. Remedy: provide approved company accounts and communicate them.
- Job applications and HR records: CVs or performance reviews are uploaded for a “summary”. Such data is particularly worthy of protection and as a rule does not belong in general AI tools.
- Confidential documents for translation or proofreading: contracts or costings end up in the tool because it is quick. Remedy: for such tasks use only expressly approved tools that are covered contractually.
- Unchecked adoption of results: AI outputs are sent to customers without review. That is a quality problem rather than a data protection one — but it undermines the same trust.
If you are unsure where your company stands on AI and data protection, a free AI check gives you an honest overview. For developing internal rules and training, feel free to get in touch.
Frequently asked questions
Am I allowed to use AI in my company at all?
Yes, with the right precautions. Tool selection, data minimisation and clear rules are decisive. With these building blocks AI becomes a safe tool — blanket bans, in our experience, only drive use into an uncontrolled grey area.
What is the greatest risk in everyday work?
The uncontrolled entry of sensitive data into unvetted tools. That is precisely what clear policies, approved company accounts and a team that understands why the rules apply will prevent.
Do I have to document every prompt?
No, there is generally no obligation to log individual inputs. What should be documented is the processing as such — that is, which tools are used for what — in the record of processing activities. For an individual case, have the matter checked legally if in doubt.
What should I do if personal data has been entered by mistake?
Report the incident internally — depending on the tool, inputs can be deleted or their deletion requested from the provider. Whether a notifiable personal data breach has occurred depends on the individual case and should be clarified with your data protection officer or legal advisers. More important than the individual response is eliminating the cause.
