AI data leaks can start with an ordinary task, like pasting a contract into a chatbot. Clear rules on confidential data, approved tools, and output checks make the safe choice the easy one.
Write a generative AI policy that names approved tools and tasks, defines what confidential data employees can enter, requires output checks, and gives everyone a clear route for approval or help. Pair those rules with practice: a policy cannot replace access controls or employees knowing what to do.
TL;DR
- A generative AI policy needs explicit confidential data rules, approved uses, named owners, and an incident reporting route.
- Employees should practice an approved task, check the result, and know when to ask for help.
- EU AI Act Article 4 literacy measures support informed use; training alone does not guarantee compliance.
Why this matters
A policy gives employees a common boundary. Yet, this doesn’t automatically ensure that everyone is aware of this boundary when making everyday decisions. “Can I paste a document into an AI tool, connect a shared folder, or send an AI-written answer to a customer?”
For example, an employee drafting a supplier email needs to know whether contract terms can enter the approved tool, whether the tool can access other files, and who checks the draft. Another employee needs a safe first task using public information before handling sensitive work. Both need the same foundation, with different levels of support.
Five steps to a confidential data policy
Start with the decisions employees actually face. Give each rule a named owner, an action employees can take, and an exception route that does not depend on guessing whom to contact.
1. Define who and what the policy covers
State whether the policy covers employees, contractors, and other people working on your behalf. Include browser tools, workplace assistants, embedded AI features, file uploads, connectors, and agents when they fall within your organization’s use of generative AI.
Assign responsibilities explicitly:
- Business owners approve the purpose and decide who reviews the work.
- IT and security assess tools, access permissions, and technical settings.
- Legal and privacy teams assess relevant contractual, privacy, and legal requirements.
- HR and learning teams organize training and maintain learning records.
- Employees follow the approved workflow and report mistakes or uncertainty.
A named role is more useful than a general instruction to obtain management approval. Tell employees where to find the current person or team responsible.
2. Classify the information before choosing the tool
Use your existing information classification scheme rather than creating a competing AI-only scheme. If employees cannot tell which category applies, require them to ask the information owner before submitting the material.
The following is an example decision table, not permission to upload any particular category. Your security and legal owners must set the actual boundaries.
| Information category | Example | Policy decision to document | Limit to address |
|---|---|---|---|
| Public information | Published product documentation | Which approved tools and tasks permit use | Public access does not remove copyright or accuracy checks |
| Internal information | An unpublished process guide | Which tools and account settings permit use | Internal material can still reveal business operations |
| Confidential information | A supplier contract | Whether a specifically approved workflow permits use | Contract terms and access restrictions still apply |
| Restricted information | Credentials or highly sensitive records | Prohibited uses and any controlled exceptions | General tool approval is insufficient |
Treat personal data and confidentiality as overlapping questions. A public document can contain personal data, while confidential commercial information need not identify anyone. The policy must account for both.
3. Approve the task, account, and data together
Tool approval should identify the account or environment, permitted tasks, allowed information categories, and required settings. Approval for drafting public marketing copy does not automatically approve uploading employee records.
Require the reviewing team to check contractual terms, use of submitted content, retention, deletion, access controls, subprocessors, and relevant data transfers. Record the decision and its conditions. An enterprise label alone does not answer those questions.
Include connected data sources in the review. An assistant that reads a shared folder needs scrutiny of folder permissions and retrieval boundaries, not just the text an employee types into a prompt.
4. Give employees a repeatable working sequence
Teach this sequence through an approved workplace task. It gives employees who are less confident a concrete starting point and gives experienced users the same boundaries.
- Choose approved task: Confirm that the intended use falls within the approved purpose.
- Check data: Identify sensitive information and remove anything unnecessary before submission.
- Use approved tool: Use the authorized account, configuration, and access permissions.
- Review result: Check facts, calculations, confidential content, and suitability before use.
- Ask for help: Stop and contact the named owner when the task or information falls outside the rules.
Checking the result is a separate responsibility from checking the input. A permitted prompt can still produce an inaccurate statement, an inappropriate recommendation, or an answer that should not reach an external recipient.
5. Write usable rules and an incident route
Use short instructions that explain the action and the reason. The following are example policy clauses to adapt to your approved tools, contracts, and procedures:
- Data entry: Employees must confirm that the information category and task are permitted for the specific AI environment before entering text, uploading files, or enabling a connection.
- Minimum necessary information: Employees must include only the information needed for the approved task and use public or synthetic examples where appropriate.
- Human review: The person responsible for the work must check AI-generated material before relying on it or sharing it.
- Uncertainty: Employees must pause and ask the designated owner when they cannot establish permission.
- Incidents: Employees must promptly report suspected disclosure, unauthorized access, or unexpected exposure through the organization’s incident process.
Explain what employees should preserve for the response team, without encouraging them to copy sensitive content into another unsuitable system. Security and privacy owners should determine containment, investigation, and any notification duties.
Why confidential data rules vary
A useful policy has a shared foundation and documented conditions for specific workflows. These factors change the decision:
- Information sensitivity: A public press release and an unpublished acquisition document require different handling.
- Contractual duties: Customer or supplier agreements can restrict disclosure or processing.
- Tool configuration: Retention, content use, sharing settings, and administrative controls affect approval.
- Access scope: A connector can expose information beyond the document an employee intended to use.
- Purpose and consequences: Drafting an internal agenda differs from supporting an employment decision.
- Applicable law: Privacy, sector-specific requirements, and AI regulation depend on the organization, activity, and jurisdiction.
Document exceptions as bounded approvals: the task, information, environment, owner, and review conditions. A blanket exception for a senior employee removes the very boundaries the policy is meant to establish.
What should the policy say about the EU AI Act?
For global companies, assess EU scope rather than assuming headquarters determine applicability. The EU AI Act covers defined providers and deployers, including certain organizations outside the EU; legal teams should assess the company’s role and activities, and whether any use case counts as high-risk.
Article 4: Teach people to apply the rules
Article 4 requires providers and deployers to take measures to support the development of AI literacy among their staff and other people dealing with the operation and use of AI systems on their behalf. It does not require them to guarantee a specific level of AI literacy for any individual. The measures take into account people’s technical knowledge, experience, education, and training, the context of use, and the people affected.
Translate that into relevant practice. Someone summarizing an approved public document needs different instruction from someone configuring an assistant with access to confidential records. A shared baseline supports both, with additional instruction for higher-responsibility roles.
Article 5: Identify prohibited uses
Article 5 prohibits specified AI practices, subject to the provision’s definitions and exceptions. For example, its prohibition on using AI to infer emotions in workplaces has an exception for systems intended for medical or safety reasons.
Have legal owners translate relevant prohibitions into explicit restrictions and escalation examples. A confidential data policy cannot make an otherwise prohibited use acceptable merely by restricting the input.
Article 50: Assign transparency responsibilities
Article 50 has generally applied since Aug. 2, 2026, and sets transparency obligations for specified providers and deployers. These include requirements concerning certain interactions with AI and certain artificially generated or manipulated content, with qualifications and exceptions in the provision.
Assign responsibility for identifying relevant outputs and deciding what disclosure is required. Do not teach a blanket rule that every AI-assisted document requires the same label.
These references describe Regulation (EU) 2024/1689, the EU AI Act, as amended by Regulation (EU) 2026/1744, at a general level. Training supports employees in carrying out rules; it does not establish legal compliance or replace legal advice.
How do you prove employees understand the policy?
Measure decisions, not just attendance. A course completion record shows that someone completed the course; a scenario assessment can show whether that person recognizes an unauthorized disclosure or knows when to seek approval.
Use clearly labeled hypothetical examples:
- Customer contract: Can the employee identify the confidentiality restriction and find the approval route?
- Public report: Can the employee complete an approved summary task and check the answer against the source?
- Connected assistant: Can the employee recognize that folder access needs review before enabling a connection?
- Accidental upload: Can the employee locate the incident process and explain the immediate reporting action?
Define the expected answers before assigning training. Keep learning records separate from claims about operational behavior: passing an assessment does not prove that every later action follows policy.
Should you build training internally or buy it?
Choose the delivery approach after defining what employees must demonstrate. A learning management system distributes and records learning; it is not, by itself, the content or the employee capability you need.
| Approach | Best for | Advantage | Limitation to plan for |
|---|---|---|---|
| Internal training | Company-specific rules and workflows | Direct connection to internal approvals and examples | Your team owns creation, updates, and assessment |
| Large learning platforms | Selecting material for different learning needs | A catalog-based buying approach | You must assess relevance and connect content to policy |
| Specialist companywide AI training | A shared employee foundation | A focused route to organization-wide instruction | Verify assessment quality and fit with local rules |
| Learning management systems | Assigning learning and maintaining records | A delivery and administration layer | Content and meaningful assessments still need an owner |
Ask suppliers to demonstrate a realistic employee decision, the expected answer, and the evidence your team receives. Also ask who updates legal references and how company-specific instructions reach employees.
Can employees use confidential data in an approved AI tool?
Tool approval alone is insufficient. Employees also need permission for the information category, task, account configuration, and connected sources involved. The policy should make those conditions visible before they upload anything.
Does removing names make a prompt safe?
Removing names does not automatically make information anonymous or nonconfidential. Other details can identify a person, and commercial secrets remain sensitive without personal names. Teach employees to assess the remaining information rather than treating name removal as permission.
FAQ
What should a generative AI policy include for confidential data?
A generative AI policy should define permitted tools, tasks, and information categories, along with review duties, approval owners, and incident reporting. Employees need instructions they can apply before entering information or enabling access.
Can employees paste a customer contract into an AI assistant?
Employees should paste a customer contract only when the specific task, information, and AI environment are authorized. Tool approval alone does not resolve contractual restrictions or confidentiality duties.
Is an enterprise AI account enough to protect confidential information?
An enterprise AI account does not by itself establish that a workflow is appropriate. Your team must assess contractual terms, settings, access permissions, and the information involved.
Does AI literacy training guarantee EU AI Act compliance?
AI literacy training does not guarantee EU AI Act compliance. Article 4 requires measures to support AI literacy, while other obligations depend on the organization’s role, systems, and activities.
When do the EU AI Act transparency rules apply?
Article 50 has generally applied since Aug. 2, 2026. Its obligations concern specified providers, deployers, and uses, with qualifications and exceptions that require a scope assessment.
How do we help employees who are unsure about using AI?
Give employees an approved task using permitted information, demonstrate the checks, and name the person who can help. Assess whether employees can repeat the task and recognize when to stop.
What should an employee do after uploading confidential information by mistake?
The employee should promptly use the organization’s incident reporting process. Security and privacy owners should direct containment, investigation, and any required notifications.
Implement a policy test
The most useful policy test is an employee decision. Give someone a hypothetical document and ask which tool they can use, what information they must remove, who checks the result, and where they ask for approval. Any unanswered question identifies a rule or teaching step to fix before expanding use.
