Which data can employees enter into ChatGPT?
Summarising a payslip in ChatGPT takes a minute and can still be a data protection problem. A simple traffic light shows your team what may go in and what may not.
Using AI well · 7 October 2026 · 10 min read
Publicly known information you have deliberately released is usually safe to enter. Personal, confidential, or contractually protected data should stay out unless a reviewed solution explicitly allows it. The same goes for documents, images, audio, and code, not only text.
The name of the tool tells you little on its own. What matters is which account your team uses, what the contract says about storage and training, where the data is processed, and whether you actually know any of that. Without that clarity, the information stays out.
Where does the risk actually start?
Someone wants an email to sound friendlier. A salesperson asks for a quote to be summarised. Both take a minute and feel harmless.
A small accounting office shows how fast that turns risky. An administrator uploads a payslip to copy it into an email. The name, the social security number, and the salary are personal data. They do not belong in a public AI tool, even when the task feels quick and routine.
A blanket ban does not fix this. Usage simply moves into private accounts, outside the company's control. Instead, tell your team the name of the approved tool and exactly where the line sits.
Green: public and explicitly released
You can enter these contents with the least worry:
- text already published on your own website
- publicly accessible product information
- general ideas with no link to a real person or a specific customer
- sample data you made up yourself
- text explicitly released internally for this exact purpose
Is green automatically fine?
Even with green, a person stays responsible. The tool can distort statements, invent sources, or drop an important caveat. Always check the output against the original before you reuse it.
Yellow: check and prepare first
These contents need an owner and a clear rule first:
- internal work instructions
- unpublished marketing copy
- extracts from offers or project files
- business figures with no personal data
- internal templates, presentations, or code
- anonymised information
Is removing the name enough?
A location, a rare event, and an exact date can together reveal who is meant, even without a name. Remove every detail that makes a person directly or indirectly identifiable. If you are not sure it worked, treat the data as red.
Red: keep out of a public AI tool
These contents stay out until a professionally reviewed and contractually secured solution exists:
- customer and employee data
- applications, CVs, and appraisal notes
- passwords, credentials, and security keys
- confidential contracts and non-disclosure information
- unpublished prices, margins, or financial data
- health data and other particularly sensitive personal data
- private photos, voices, or videos of identifiable people
- trade secrets and security-relevant code
Five questions before every entry
Answer these before you copy or upload:
- Is the information already public or explicitly released?
- Does it involve a customer, an employee, or another identifiable person?
- Is it protected by a contract, a confidentiality duty, or an internal rule?
- Do you know how the provider stores and reuses inputs?
- Can someone check the result before it becomes a message or a decision?
What if one answer is unclear?
One unclear answer is reason enough to stop. Clarify the use before the data goes in, not after. That short pause costs less time than a data protection incident you have to clean up later.
Does a paid account make it safe?
A paid business account often gives you more control than a private one, but it does not replace a review. On many free or personal accounts, inputs can be used to train the model by default unless you turn that off in the settings. Business and API accounts often rule this out by contract, but do not rely on the plan name alone. Read the terms on data location, retention, access rights, logging, and training before you share anything sensitive.
Data protection law applies to AI-supported processing too, whether that is the GDPR in the EU and the UK or the rules where you operate. When a business hands processing to an external service, it remains responsible for adequate data protection. So before you roll a tool out, ask in writing that your inputs are not used for training, and clarify where the data is processed.
Sources and further reading
FAQ
Not automatically. An address, a single event, or an invoice amount can still make a person identifiable. Check the whole text, not just the name. If you are unsure, treat it as red.
No. A business account is a good starting point, nothing more. Check the contract, the account settings, the data location, and whether your inputs may be used for training. Without that check, the risk is much the same as with a private account.
Management decides what AI may be used for and who is accountable for it. Bring in a specialist for sensitive data protection or contract questions. The rule only works if every employee knows it and applies it day to day.
Mygenzy works with small and mid-sized businesses on processes, marketing, and AI. This article was created with the help of AI and reviewed by people.
Give your team a clear AI rule
In the Growth Check, we map out where your team already uses AI, which data is involved, and which simple rule is missing first. Legal and security-critical questions go straight to the right specialists.