Company Policy for Employee ChatGPT Use: What to Allow and Ban

Company Policy for Employee ChatGPT Use: What to Allow and Ban

A good company policy for employee ChatGPT use says, in plain language, which tools are approved, which types of data can be pasted into them and which never can. It also says what to do when someone makes a mistake and who decides the doubtful cases. Banning everything rarely works: people keep using the tools, only without rules and without anyone seeing it.

Key takeaways

• Sort data into three simple bands (free, restricted and prohibited) and make it clear what belongs in each.
• Approve tools by name and by account type. A personal account and a company account do not have the same contractual treatment.
• Treat informal use as information, not as a crime: offer an approved path before you punish.
• Personal data, trade secrets, credentials and client documents need explicit rules, grounded in data protection law (in Brazil, the LGPD).
• The policy needs an owner, a periodic review and a channel for questions that answers in minutes, not weeks.

Why banning everything does not work

When a company blocks generative AI without offering an alternative, the work does not stop. Someone still asks ChatGPT to summarize a meeting, rewrite an email or make sense of a spreadsheet, only now on a phone, through a personal account, with nobody aware of it. The risk does not shrink. It becomes invisible.

That is why the policy should start with a practical question: which tasks is your team already trying to solve with AI, and which of them are safe enough to allow under a rule? A policy that acknowledges real usage is more likely to be followed than one that pretends it does not exist.

There is also an important distinction between three kinds of statements, and the policy should respect it. What a vendor says in its documentation and terms is a fact you check at the source. What your team does day to day is implementation experience, which you learn by asking and observing. What you think might happen to your data is a hypothesis, and it should be treated as one until it is verified. Mixing the three is the origin of most bad rules.

What data can and cannot be pasted into AI?

The most useful rule is the one that fits in the head of someone in a hurry. Three bands are usually enough.

The free band holds content that is already public or not sensitive. The restricted band holds internal content that can only go into approved tools, on company accounts. The prohibited band holds what must not be pasted into any generative AI tool, unless there is a formal, documented decision.

Examples of content that usually fits the free band:

• Website copy, already published marketing material and public announcements.
• Generic knowledge questions that include no company or client data.
• Drafts on topics that do not reveal strategy, clients or internal numbers.

Examples of restricted content, which goes only into an approved tool on a company account:

• Internal documents without personal data, such as procedures, meeting notes without sensitive names and draft policies.
• Source code that contains no secrets and is not the competitive core of the business.
• Aggregated business data, anonymized so that no one can be identified.

Examples of prohibited content:

• Personal data of clients, employees or job candidates (a name with a tax ID, address, phone number, support history, resumes).
• Sensitive personal data as defined by the LGPD, such as health, biometrics, racial origin, religion or union membership.
• Passwords, tokens, API keys, certificates and any other credential.
• Contracts under confidentiality, proposals under negotiation and third-party information the company has committed to protect.
• Trade secrets, formulas, unreleased pricing strategy and acquisition plans.
• Full WhatsApp conversations with clients, because they contain personal data and context the client did not agree to share.

These bands do not replace legal analysis. To understand the legal basis, the purpose and the duties of the data controller, see the guide on LGPD and artificial intelligence. The ANPD, Brazil's national data protection authority, publishes official guidance that should be your reference whenever there is doubt about processing personal data.

How to handle informal use by your team

Informal use, often called shadow AI, is where most policies fail. If the first reaction to discovering it is punishment, people start hiding it. If the first reaction is curiosity, the company learns what the team needs.

An approach that tends to work has four moves:

• Amnesty for a defined period: anyone who declares what they already use is not punished for past use, as long as they stop whatever is prohibited.
• A simple survey: a short form asking which tool, for which task and with what type of data.
• An approved alternative: for each legitimate use found, offer a tool and a rule instead of just a no.
• Proportionate consequences: from the policy's start date, use in the prohibited band is handled under the disciplinary rules you already have.

The survey usually reveals very similar tasks across departments: summarizing text, drafting replies, reviewing spreadsheets, translating. That helps you decide where to invest in an approved solution, and the logic of which processes to automate with AI first helps you prioritize.

Step by step: how to write the policy

The policy does not need to be long. It needs to be clear, short enough to be read and specific enough to be enforced. One possible path:

• Define the scope: who is covered (employees, interns, contractors) and which tools (chatbots, assistants built into editors, browser extensions, mobile apps).
• List the approved tools, the required account type and who administers access. Check the vendor's terms for how data is handled on each plan type instead of assuming.
• Publish the three data bands with everyday examples from your own company, not generic ones.
• Write absolute prohibitions in short sentences, with no implied exceptions.
• Set the human review rule: nothing generated by AI goes to a client, a contract, payroll or a decision about people until a responsible person has read it and taken ownership.
• Name a policy owner and a question channel with a short response time.
• Describe the incident procedure: what to do on noticing that prohibited data was pasted, whom to tell and how fast.
• Schedule a periodic review, because tools, plans and terms change.

If the company plans to go beyond individual use and connect AI to systems, the policy has to work together with action approval rules. The AI agent approval policy template helps separate what the AI may suggest from what it may execute.

Illustrative scenario: the analyst who wants to summarize complaints

Illustrative scenario: a customer support analyst receives a file with hundreds of client messages and wants a summary of the main reasons for complaints. She opens a personal account on an AI assistant and pastes the whole file, including names, phone numbers and order numbers.

Under the proposed policy, this use falls in the prohibited band on two counts: the account is not the approved one, and the content contains client personal data. What the policy should offer as the correct path:

• Remove or mask names, phone numbers and identifiers before any analysis.
• Use the approved tool, on a company account, with only the complaint text, already stripped of identifying details.
• Ask the model for categories and generic examples, and check a sample manually against the original messages.
• Log the task in the question channel, so the case becomes an example in the next version of the policy.

Note that the analyst did nothing malicious. She had a real problem and a tool within reach. The policy exists to let her solve the same problem safely.

Decision checklist before pasting any text

A short list, taped near the keyboard or pinned in the team channel, is worth more than a long document. Before pasting anything into generative AI, the person answers in order:

• Is the tool one of the approved ones, and am I on the company account? If not, stop.
• Does the text contain anyone's personal data, even a little? If so, remove or mask it, or do not use it.
• Does it contain a password, key, token or credential? If so, never paste it.
• Is it under confidentiality with a client, supplier or partner? If so, do not paste it without formal authorization.
• Could I explain to a client or to the legal team why I sent this? If not, ask first.
• Who will review what the AI returns before it leaves the company? If nobody, the task is not ready for AI use.

The last item matters as much as the data items. A good share of real damage comes from wrong answers that were accepted without being read, not only from leaks. The catalog of risks for applications built on language models maintained by OWASP is a good reference for understanding, for example, overreliance on model output and exposure of sensitive information.

How the map, limit, pilot and transfer method applies

A usage policy is a good first governance project because it follows the same path as any AI initiative. Map: find out which tools and tasks already exist in the company, and with what data. Limit: decide the data bands, approved tools and permitted actions. Pilot: test the policy with one department or one real workflow, logging questions and incidents against an agreed metric, such as the number of questions to the help channel that turned into a rule. Transfer: hand over the policy with an internal owner and decide whether to expand, adjust or stop.

For workflows where AI starts handling personal data on a recurring basis, such as customer support or HR, use stops being purely individual and may require a formal assessment. See the guide on data protection impact assessments for AI projects to know when an impact report enters the conversation. The NIST AI risk management framework also helps structure the thinking of mapping, measuring and managing risk without depending on fashion.

Frequently asked questions

Can a company simply ban ChatGPT? It can, but that usually pushes use onto personal accounts and phones, where the company loses all visibility. A ban is only effective when it comes with technical blocking, an approved alternative and a clear explanation of the reason.

Can I paste client data if I only delete the name? Usually that is not enough. A phone number, email, order number, address and even the context of the complaint can identify the person. If identification is still possible, the data remains personal data under the LGPD, and the safest rule is to mask it properly or not use it.

Does a company account on an AI tool solve the privacy problem? It usually comes with different terms and controls than personal accounts, but that depends on the vendor and the plan, and it should be verified in the official terms and documentation. Even with a company account, the company remains responsible for deciding which data goes in, on what legal basis and for what purpose.

How often should the policy be reviewed? Review it at intervals set by the policy owner, and whenever a tool changes plan, terms or features, or when an incident exposes a gap. A short, frequent review is better than a large rewrite that gets forgotten.

Sources

• ANPD: Autoridade Nacional de Proteção de Dados
• NIST AI Risk Management Framework
• OWASP Top 10 for LLM Applications