
AI Use Policy for Employees: What to Allow, Ban and Review
A practical employee AI use policy template: what staff can paste into AI tools, what is banned, and how to review the rules when models change.
Want to apply this in your company?
Book a call with Juan →An employee AI use policy is a short document that tells your team which information can be used in AI tools, which tasks are allowed, what is prohibited, and who decides on exceptions. It works best when written in plain operational language, tied to tools the company has approved, and given a review date. Without it, each person decides alone what to paste into a chat, and the problem tends to surface when it is already hard to fix.
Key takeaways
- Classify information into simple levels before you list tools.
- Allow low-risk uses by describing concrete tasks, not broad categories.
- Ban pasting customer personal data, passwords, access keys and confidential documents into unapproved accounts.
- Name an owner for the policy and a trigger for review, because tools and terms change.
- Train the team with everyday examples and offer a channel for questions.
Why write the rule before the incident
In most companies, AI is already being used before any formal decision is made. Someone summarizes a customer email, another person asks for help drafting a proposal, and a third transcribes a meeting in a free app. These habits are understandable because they usually save time, but they remove any control over where the data ends up.
The policy does not need to ban everything. It should separate what the company accepts, what it accepts under conditions, and what it does not accept under any circumstances. The more concrete the list, the less the team has to guess.
What staff can paste and what they cannot
We recommend a three-level classification, simple enough to fit on one page:
- Public: text already published on your website, marketing material and generic training content. It can be used in approved tools without special conditions.
- Internal: procedures, proposal templates without client prices and service scripts. Use only in approved business accounts and without identifying people.
- Restricted: personal data of customers, employees and patients, confidential contracts, banking data, credentials and information from legal cases. Nothing in this level goes into an AI tool without a formal review by the responsible person.
Personal data is the most sensitive point in Brazil, because the LGPD applies to its processing, including when that processing happens in a third party's tool. Before approving any tool, read the terms of the business account and confirm in the vendor's official documentation how data is stored, whether it is used for training, and in which region it is kept. These answers change over time, so the check should be part of the process, not just the memory of whoever set up the account.
What is prohibited: a checklist
The list of prohibitions needs to be short and specific. A first version can include the items below, adjusted to your sector:
- Pasting personal data of customers, employees or patients into personal accounts or unapproved tools.
- Sending passwords, tokens, API keys or system access details.
- Letting AI decide alone on credit, hiring, dismissal, sanctions or refusal of service, without recorded human review.
- Sending a client text generated by AI as a technical, legal or accounting opinion without review by a qualified professional.
- Connecting AI tools to folders, email inboxes or internal systems without approval from IT or the responsible person.
- Recording meetings with outside participants without notice and without an adequate legal basis.
- Evaluating people's performance based on data extracted by AI without clear criteria that the team knows.
Step by step: writing the first version
A useful policy starts from a survey of what already happens. The path below tends to work in companies of any size:
- Step one: ask each area which AI tools and tasks it already uses. The question is about mapping, not punishment, and the team needs to hear that clearly.
- Step two: classify the types of information that move through each area using the three levels above.
- Step three: list the approved tools, the account that must be used in each one, and the tools that stay prohibited for now.
- Step four: define who approves exceptions, within what deadline, and how the request is recorded.
- Step five: write the rule on one page, with examples of correct and incorrect use for each area.
- Step six: train the teams on real work tasks and keep a channel open for questions.
- Step seven: set the date of the first review and the triggers that bring it forward.
Illustrative scenario: a customer service agent at a services office has to answer a client who questions a charge. She pastes the full history into her personal chat, including name, tax ID, amounts and address, asks for a reply, and sends it. If the policy existed, the answer would already be defined: identification data stays out of the tool, the reply template is generic and the fields are filled in the official system, and complex cases go to the area lead. The need is still met, without exposing the client.
How to review the policy when models change
An AI policy ages quickly. Tools gain new features, such as reading files, browsing the web, remembering past conversations and connecting to email or CRM. Each new feature changes the risk, even when the tool keeps the same name. For that reason, the review should be triggered by events, not only by the calendar.
- A new tool or a new plan contracted by the company.
- A change in terms of use, privacy settings or storage location published by the vendor.
- A new feature that reads files, accesses the internet or takes actions in systems.
- An incident, a near miss, or a question that keeps coming up in the team.
- A change in data protection guidance or AI regulation that affects your sector. It is worth following the publications of the ANPD.
When reviewing, check the record of approved exceptions, which tools are actually in use, and whether the rule still fits the real work. When a tool reads external content, such as incoming emails or web pages, also consider prompt injection: a malicious text may try to push the model to act outside what was agreed. OWASP documents this risk and mitigation measures. To structure governance, the NIST AI risk management framework offers a useful vocabulary that does not depend on any specific tool.
If the policy involves monitoring use, care for the team matters as much as the rule itself. See how to handle that topic in how to measure AI productivity in your team without surveillance.
How the policy fits into the implementation method
In our implementation method, the policy is one of the first deliverables. The process follows four steps: map the current situation, limit data and actions, pilot a real workflow with a record and an agreed metric, and hand control to the team, with the decision to expand, adjust or stop. Applied to the policy, mapping is the usage survey, limiting is the data classification, the pilot is the adoption of one approved tool in one area, and the handover is training with a named owner. There is no ready-made text that fits every company, because what changes is the type of data each area handles.
For an example focused on a conversational tool, see company policy for ChatGPT use. If your team records meetings to summarize decisions, also read recording meetings with AI and consent under the LGPD.
Frequently asked questions
Who should approve a new AI tool in the company? In small companies, the owner or an operations manager can approve, as long as they consult IT or the provider responsible for the data. The approval should consider what kind of information the tool receives, where the data is stored and whether the account is a business account. The decision should be recorded somewhere accessible, with the date and the person responsible.
Does the policy need to be a legal document? It does not need contract language, but it has to be clear enough to be followed. A short version works for the team, with a technical annex that lists the tools and the data classification. If your company handles personal data at scale or operates in a regulated sector, it is worth having a lawyer who specializes in data protection review the text.
Can employees use AI on their phones with a personal account? The policy should address this case explicitly, because use on personal devices often escapes control. A reasonable rule is to allow only public content or generic questions in those accounts and to prohibit customer data and internal documents. When a team needs mobile access, the company should provide a business account with the same restrictions.
How often should the policy be reviewed? Beyond the periodic review on the calendar, the policy should be reviewed whenever one of the triggers above appears. For most companies, the most useful practice is a short review that comes with the log of exceptions and incidents. The person responsible for the policy should be known to the whole team.