ChatGPT and Copilot at work: six rules for a small business

Your staff already use it

Eurostat measures AI use in enterprises with 10 or more employees. In 2025, 34.54% of Belgian enterprises used at least one AI technology, against 19.95% across the EU. Among Belgian enterprises with 10 to 49 employees it was 28.82% (Eurostat, dataset isoc_eb_ai). The survey leaves out businesses with fewer than 10 employees, farming and the financial sector.

The six rules below are for the owner whose staff already use ChatGPT or Copilot. This is not legal advice; the texts on EUR-Lex are the reference.

BelgiumEU (27 countries)

2023

Belgium13.81%

EU (27 countries)8.06%

2024

Belgium24.71%

EU (27 countries)13.48%

2025

Belgium34.54%

EU (27 countries)19.95%

Source: Eurostat, table isoc_eb_ai (use of AI by enterprises), updated 15 June 2026

What the law asks: AI literacy

The EU AI Act calls a business that uses an AI system for its work a deployer (AI Act, article 3). Article 4, on AI literacy, has applied to providers and deployers since 2 February 2025. It was rewritten by the Digital Omnibus on AI, Regulation (EU) 2026/1744 of 8 July 2026, published on 24 July 2026 and in force since 27 July 2026 (EUR-Lex).

The new article 4 asks providers and deployers to take measures to support the AI literacy of their staff and of others who use AI systems on their behalf, taking into account what those people already know and how the tools are used. It does not require you to guarantee any specific level for any individual. The Commission and the Member States must help, SMEs in particular.

The European Commission's Q&A on AI literacy adds that no certificate is needed and that internal records of training are enough. It gives the example of staff who use ChatGPT for advertising copy or translation: they should be told about specific risks such as hallucination. The supervision and enforcement rules apply from 2 August 2026 (European Commission).

Rule 1: Choose the tools and the type of account

Write down which tools may be used for work, and with which account. The vendors describe the difference themselves.

• OpenAI says that for its services for individuals (Free, Plus and Pro) it may use your content to train its models, unless you switch off Improve the model for everyone under Settings, Data controls. Temporary chats are not used for training (OpenAI).
• For ChatGPT Business, ChatGPT Enterprise and its API, OpenAI says it does not train on business data by default, and it offers a Data Processing Addendum for the GDPR (OpenAI).
• Microsoft says that Copilot and Copilot Chat used by an organisation, signed in with a work or school account, fall under its Data Protection Addendum and Product Terms, with Microsoft as data processor, and that prompts and responses are not used to train foundation models (Microsoft).
• With a personal Microsoft account, the Microsoft Services Agreement and the Copilot Terms of Use apply instead (Microsoft).

Rule 2: No personal or client data in a free consumer account

Personal data is any information about a person who can be identified: a name, an email address, a phone number, a CV. Health data and a few other categories get extra protection under article 9 of the GDPR. When a supplier processes personal data for you, the GDPR requires a contract with that supplier (article 28), and you may only use the data you need (article 5) (GDPR). With a free consumer account you accept terms as an individual, without such a contract.

The Belgian Data Protection Authority (GBA) advises caution before typing sensitive information, such as medical or financial data, into a chatbot. It suggests pseudonyms or removing names, addresses and ages, and it advises checking default settings for chat history and AI training. Its test question: would you share this with strangers at a dinner table (GBA)?

In practice: write "client A" instead of a name, and keep passwords, contracts and unpublished prices out of every chat tool. One more detail from OpenAI: on its plans for individuals, a thumbs up or down can make the whole conversation available for training, even after you opted out.

Rule 3: Check every answer before it goes out

A chat tool can write a sentence that sounds right and is wrong. Treat each answer as a draft and check names, numbers, dates, prices, legal or tax statements, links and quotes against your own source before it goes out. The person who sends the text is responsible for it. Do not let a tool answer clients directly without someone reading the answer first.

Rule 4: Label AI content where the law asks for it

Article 50 of the AI Act has applied since 2 August 2026 and splits the work (AI Act, article 50).

• The providers, such as OpenAI and Microsoft, must mark what their generative systems produce in a machine-readable way. For systems already on the market before 2 August 2026, the Digital Omnibus gives them until 2 December 2026.
• You, as deployer, must disclose that an image, audio or video is artificially generated or manipulated when it is a deep fake: content that resembles existing persons, objects, places or events and would falsely appear to be real. For an evidently artistic, creative, satirical or fictional work, a disclosure that does not spoil the work is enough.
• AI text published to inform the public on matters of public interest must also be disclosed, unless a person reviewed or edited it and someone holds editorial responsibility for it.
• The disclosure must be clear and come at the latest when people first see or hear the content.

A realistic AI photo of a shop interior, a customer or a team that does not exist needs a label such as "Image generated with AI".

Rule 5: Teach it in an hour and write down that you did

One hour is often enough for a small team: go through the house rule below, show two answers that sounded right and were wrong, and agree what never goes into a chat tool. Note the date and who attended; the Commission says internal records are enough. Repeat it for new staff and new tools.

Rule 6: One person decides, and mistakes are reported

Name one person who approves new AI tools and any connection between a tool and the mailbox, the agenda or shared files. Ask staff to tell that person the same day when wrong data was pasted or a wrong answer was sent.

A one-page house rule to copy

House rule: AI chat tools at [business name]. Read and signed by every staff member.

1. Allowed for work: [for example ChatGPT Business, or Copilot Chat with your work account]. Only with the account the business gives you.
2. Never paste: names, contact details or files of clients, staff or patients, health data, ID or bank numbers, passwords, contracts or prices that are not public. Write "client A" instead of a name.
3. Every answer is a draft. Check names, numbers, dates, prices, legal and tax points and links against your own source before anything goes out.
4. A realistic AI image, sound or video that could pass as real gets the label "Made with AI".
5. No AI tool is connected to the mailbox, agenda or shared files without approval from [name].
6. Pasted the wrong data or sent a wrong answer? Tell [name] the same day.
7. Questions: ask [name]. Last checked on [date].

Name, date, signature.

Checklist

1. Allowed tools and account types written down.
2. Work accounts handed out, free accounts no longer used for work.
3. Chat history and training settings checked in every allowed tool.
4. House rule signed, one-hour session held and noted.

Want help choosing the tools and setting them up around how your business works? See our AI automation page. Our security awareness workshop ends with a few agreements that fit your business, and the house rule above can be one of them.