A client sends a tense email: they complain about a delay, demand an explanation and invoke a contractual clause. You want to reply quickly. You paste it into a chatbot, ask for “a calm, professional message”, then send the generated version.
In those few seconds, three different problems may arise: you may have used an unapproved tool, sent information that the organisation has not authorised you to enter there, or sent a message that invents an explanation, promises something the company cannot offer, or omits a contractual detail.
AI literacy starts less with the prompt and more with this pause. You do not have to become a specialist to avoid the most common mistake: putting real data into a workflow you do not understand.
For more information on artificial intelligence, see also After the Chatbot Closes: What Students Should Be Able to Do Without AI, which offers complementary details on this topic.
The InfoBuz first-paste test: green, yellow or red
This is a practical InfoBuz framework, not a test written verbatim into the AI Act. Its purpose is to turn a vague rule — “use AI responsibly” — into an easy-to-apply work decision.
| Outcome | When you are here | What you do |
|---|---|---|
| Green: you can proceed | The task is permitted, the content is already public or created specifically without identifiers, and the tool is approved. | Use AI for a first draft, then check the result before using it. |
| Yellow: proceed only in a controlled way | You have non-confidential internal content or a summary reduced to what is strictly necessary, and the organisation has approved a work environment for that workflow. | Use only the designated tool, reduce the data and carefully validate the response. |
| Red: stop and ask for approval | Personal data, sensitive information, contracts, prices, passwords, source code, security incidents or a decision about people are involved. | Do not paste it into a chatbot of your choosing. Ask your manager, IT/security, DPO or legal team for the correct route, as appropriate. |
The test has three questions, in this order: is the tool approved? is this version of the data permitted? who checks the result before it has consequences? If any of them does not have a clear answer, do not improvise with real data.
Applying it to the client email: what changes in practice
The original email is generally a red flag. It may contain names, contact details, an order number, the history of the commercial relationship, negotiated terms or clues that can identify the client even if their name is removed.
One possible working version may become yellow, not automatically green: manually formulate the issue as a minimal summary — for example, “business client complains about a delivery delay and requests a diplomatic response; do not promise compensation” — and use it only in a work environment approved for that use. If no such environment or clear internal rule exists, stop the workflow.
Once AI proposes a draft, checking it is not merely a matter of style editing. Compare the message with the real situation: the delivery deadline, confirmed facts, contractual obligations, the person who can approve compensation, the tone and any new commitment. A polite text can still be wrong.
Not all work information follows the same route
1. Public content or content created for practice. A press release that is already public, a generic product description or text invented to test tone can usually fall into the green zone if the tool is approved. Checking the result remains mandatory: AI may introduce names, figures or claims that are not in the source material.
2. Non-confidential internal content. A presentation outline, a general meeting summary or a list of ideas may fall into the yellow zone. The question is not whether the document appears harmless, but whether the organisation has allowed that type of material in that environment and whether you can remove details that are unnecessary for the task.
3. Sensitive or confidential content. This includes, as a guide, CVs, performance evaluations, salaries, medical records, complaints, contracts, negotiations, non-public prices, passwords, access keys, source code, configurations and details of security incidents. They are not all the same legal category, but they call for the same operational reflex: do not move them into a chatbot you have chosen on your own.
“I removed the name” does not solve the problem on its own
Removing a name can reduce risk, but it does not automatically turn a document into anonymous material. Context may indirectly identify a person or company: their role, city, order number, a unique project, the timeline of an incident or the combination of several details.
The European Data Protection Board states that anonymisation must be assessed on a case-by-case basis. For an AI model to be considered anonymous, both the direct or indirect identification of individuals and the extraction of personal data through queries must be very unlikely. That is why “Client A” or initials are not automatic permission to upload material to an AI service. edpb.europa.eu
In practice, separate two questions that employees often confuse: can I reduce the data? and do I have permission to use this reduced version in this tool? The first does not always resolve the second.
Personal account, public tool, enterprise environment: what matters is approval of the workflow
The product name does not decide the issue. The same service may be accessed through a personal account, a free account or a company environment configured and contracted for particular activities. These situations are not equivalent simply because the interface looks the same.
Before using it, an employee should be able to get straightforward answers: is there a work account? Is the product approved for my task? What types of data are allowed? Has the organisation set rules on retention, the use of data for training, administrator access, logging, transfers and agreements with the provider? If nobody can answer, do not assume that a personal account is an acceptable alternative.
What the law requires — and what it does not
Article 4 of the AI Act has applied since 2 February 2025. It requires providers and deployers of AI systems to take measures to support the development of AI literacy among staff and other persons using or operating AI systems on their behalf, taking account of their role, experience, training and context. The Regulation does not require a single certificate, nor does it require guaranteeing an identical level of knowledge for every person. ancom.ro
Regulation (EU) 2026/1744, adopted in July 2026, amended the wording of Article 4 accordingly: the emphasis is on measures taken by the organisation to support the development of AI literacy, rather than on a promise that every user has reached a uniform threshold. ancom.ro
For employees, the practical translation is simple: AI literacy does not just mean getting a convincing answer. It means recognising what should not be entered, when you need the approved route and why a fluent result does not automatically become correct.
Romania: do not wait for a penalty to ask for clear rules
ANCOM recommends that organisations document their AI literacy measures and states that a specific certificate is not necessary. The authority even gives the example of employees using ChatGPT for advertising copy or translations: they must be informed of the risk of invented or inaccurate results and of the need for verification before publication or transmission. ancom.ro
ANCOM also states that national market-surveillance authorities may begin monitoring Article 4 from 2 August 2026, and that in Romania supervision and enforcement depend on the entry into force of the relevant provisions of the national implementing law. This nuance does not suspend the usefulness of internal rules: it shows why it is wrong both to announce automatic penalties and to postpone all measures until external clarification appears. ancom.ro
If your company has not told you what you are allowed to do
Do not ask for “an AI course” in general terms. Ask for a decision that applies to your task. You can send this message:
“For this task, what is the approved AI tool? What version of the data may I use in it? Who validates the result before it is sent or used externally?”
If necessary, add four pieces of information: the tool you want to use, the type of data, the purpose and the audience for the result. For example: “I want to summarise a client complaint for an external response.” This clarification helps the manager, IT/security, DPO or legal team decide on the appropriate route without unnecessary email exchanges.
The rule to remember: a good prompt saves time. A timely stop can prevent exposure of a client, an incorrect contractual promise or a security incident. The tool, the data, the verification — only then comes the prompt.
Frequently asked questions
Does the AI Act require every employee to obtain an AI certificate?
No. Article 4 requires providers and deployers of AI systems to take tailored measures to support AI literacy. It does not require a single certificate, nor does it require guaranteeing the same level of knowledge for every person.
Can I put a document into a chatbot after removing the name?
Do not assume that it is automatically permitted. The document may indirectly identify a person or contain contractual, commercial, technical or security information. Check both whether the data have been reduced sufficiently and whether the workflow is approved for that particular tool.
Is every public chatbot prohibited at work?
There is no automatic prohibition that applies to every situation. But a product being accessible does not mean it is approved for work data. The task, the data, the internal policy, the service configuration and the organisation’s relationship with the provider all matter.
What should I check before sending an AI-drafted email?
Check the facts, figures, names, sources, deadlines, commercial commitments, relevant clauses and tone. For an external message, verification must confirm above all that the draft does not promise something the organisation has not decided or cannot deliver.
Who is responsible if an AI result is wrong?
The precise legal liability depends on the situation and the applicable procedures. As a working rule, AI is not an authority that assumes responsibility: the organisation must establish tools and checks, and an employee should not send an unverified result on its behalf.
