How to write a corporate AI usage policy (template outline)
Nanda Teknoloji Engineering Team · · 8 min read
In brief
A good AI usage policy should be short, actionable, and backed by technical controls. It defines scope, data classes, approved tools, roles, prohibited uses, and the exception process; instead of blocking employees, it makes the safe path easy.
A corporate AI usage policy is a short, actionable document that defines for what purpose, with what data, with which tools, and with whose approval employees and applications may use artificial intelligence tools. An effective policy contains seven components: scope, data classes, approved tools, prohibited uses, roles and responsibilities, an exception process, and violation management. The value of the policy lies not in its length but in an employee being able to find the answer to the question “can I use this tool with this data?” in thirty seconds. Below is an outline you can adapt to your organization, with the points to watch in each section.
Why is a written policy needed?
In the Shadow AI picture, employees usually turn to tools outside the organization not with bad intent but to speed up their work. If there is no policy, the employee decides by their own judgment; if the policy is overly restrictive, they work around the rule. A written policy does three jobs: it clarifies expectations, it forms the basis for technical controls (access, filtering, logging), and it documents responsibilities in the event of an incident. It also provides a basis for accountability in uses involving personal data; we cover this aspect in our AI and KVKK article.
Policy outline: section by section
1. Purpose and scope
Write whom it covers (employees, contractors, interns), which tools (chat services, code assistants, image generators, browser extensions, models embedded in products), and which devices (corporate and personal). Keep the scope general, such as “generative AI and LLM-based tools”; list tool names in an appendix, because names change.
2. Definitions
Define terms such as generative AI, prompt, approved tool, personal data, and confidential data in a few sentences. Use everyday language instead of legal language.
3. Data classes and permission matrix
This is the backbone of the policy. Align with the organization’s existing data classification; do not invent a new class system. An example matrix:
| Data class | Example | Approved corporate tool | Unapproved external tool |
|---|---|---|---|
| Public | Published announcement, product page | Permitted | Permitted |
| Internal use | Meeting notes, internal procedure | Permitted | Prohibited |
| Confidential | Contract, financial statement, source code | Conditional (approval and logging) | Prohibited |
| Personal / special category | Customer information, health data | Only in a tool that has undergone legal assessment | Prohibited |
Explain the “conditional” cells in the matrix: what approval, what logging, what masking is required?
4. Approved tools and the acquisition process
The approved tool list should be published in an accessible place inside the organization and have a known owner. Define a short process for new tool requests: who applies, what security and legal look at (data processing terms, retention, use for model training, location), and within how many business days a decision is given. If the process is slow, the employee does not wait and finds their own way.
5. Acceptable and prohibited uses
Write with concrete examples:
- Permitted: Preparing draft text with general information, summarizing public content, getting code suggestions in an approved tool.
- Prohibited: Uploading a customer list to an unapproved tool, writing authentication credentials into a prompt, putting AI output into official correspondence without verification.
- Requires approval: Using AI in decisions that affect people, such as hiring, performance evaluation, or credit.
6. Output responsibility
Model output may contain errors, bias, or copyrighted content. The policy should state clearly the verification responsibility of the person who uses the output, and define how AI use in content going to customers will be handled.
7. Roles and responsibilities
Give the duties of the policy owner, approvers, data owners, information security, the KVKK officer, and managers in a short table. “Everyone is responsible” means no one is responsible.
8. Exception and violation process
With what form and to whom does a person requesting an exception apply? How long is the exception valid? How is a policy violation reported? Keeping reporting within an educational rather than punitive framework prevents incidents from being hidden.
9. Review
Add a version number, effective date, and review calendar. New tool types, legislative changes, and post-incident findings are update triggers.
Linking the policy to technical controls
A written rule remains symbolic if it cannot be enforced. Every critical rule in the policy should have a technical counterpart:
- The approved tool list is reflected in access rules at the network and browser level.
- Data class rules are enforced in the AI DLP layer that inspects prompt content.
- Usage logs and quotas are collected through a central AI Gateway.
- Exceptions are defined with an expiry and close automatically when the time is up.
To set up a risk framework, the NIST AI Risk Management Framework can be a reference source for classification and governance structure.
Common mistakes
- The policy is written only by legal or only by IT. If business units are not part of the process, the rules do not fit real workflows.
- Embedding tool names in the body. Tools change; the policy has to be rewritten from scratch every time.
- Skipping training. Announcing the policy by email is not enough; short, example-based training is needed.
- Making the safe path harder. If the approved tool is unusable or access is slow, the employee returns to unapproved tools.
- Tying violation reporting to punishment. Employees hide the mistake, and the organization learns of the incident late.
Example: what the policy’s front page might look like
The first page the employee reads should be short and action-oriented. The structure below can be used as an “at a glance” box that summarizes the long text:
- You may use: Approved corporate tools, with content the data class matrix permits.
- You may not use: Passwords, keys, customer personal data, and confidential documents in unapproved tools.
- Ask: If you are unsure, contact your manager and the information security unit; in case of doubt, the default is “do not send.”
- Verify: AI output is your responsibility; check it before using it.
- Report: If you made an erroneous submission, report it immediately; the aim is not to blame but to limit the impact.
Using these five items as a first page that links to the full policy text ensures the document is actually read.
Role-based rules: Is one policy enough for everyone?
Different roles carry different risks; the policy should take this into account.
| Role | Typical use | Additional rule example |
|---|---|---|
| Software developer | Code suggestions, debugging | Confidential information and private repository code only in approved tools |
| Human resources | Résumé summary, job posting text | Personal data is masked, the decision stays with a person |
| Legal | Contract drafting, legislation review | Party information and confidential clauses in the in-house tool |
| Marketing | Content drafts, image generation | Copyright and brand consistency check, disclosure of AI use |
| Customer service | Response drafts, summaries | Customer data is masked; responses are verified |
| Management | Meeting summaries, reporting | Approved closed environment for strategic and financial information |
Keep role-based additions as appendix pages without bloating the core policy; that way a change affects only the relevant team.
Rollout timeline for the policy
Adoption matters as much as publishing the policy. A staged rollout can follow this order:
- Preparation: Inventory of current use, identifying stakeholders, writing the draft.
- Pilot: Trying it with one or two units; drawing an FAQ and examples from real questions.
- Publishing and training: Short, example-based sessions; separate briefing for managers.
- Turning on technical controls: Starting from monitoring mode and gradually enabling rules.
- Feedback and revision: The first update with the findings of the first three months.
In training, use realistic cases instead of abstract principles: “While writing a reply to a customer, you pasted the email into a chat tool; what should you have done?”
Measuring the policy’s success
Compliance is not measured only by “the rate of those who signed.” More meaningful indicators: increased use of approved tools, decreased traffic to uncontrolled tools, the duration and outcomes of exception requests, and the number of reported incidents and time to closure. A rise in the number of reports can initially be a positive sign: it means employees are speaking up instead of hiding things.
How should the exception process work?
The exception request is a mechanism that keeps the policy flexible; but if left uncontrolled, it renders the rule meaningless. The request form should include the purpose, data class, tool to be used, duration, and a compensating measure (masking, additional logging, manager approval). The decision time should be known in advance, approval should be granted for a limited period, and it should close automatically when the period ends. The same exception being requested again and again shows that the policy should be reviewed at that point.
Decision summary
Build the policy as a two-page core plus appendices. Put the data class matrix at the center, keep the approved tool path fast, link every rule to a technical control, and review at least once a year. Rather than waiting for a perfect policy, starting with a publishable first version and improving it with monitoring data is more effective.
Moving forward with Nanda Teknoloji
Nanda Teknoloji carries out the discovery, design, migration, and operation steps in projects like these: an inventory of current AI use, mapping the policy to technical controls, and monitoring in operation. For a holistic approach, you can look at the AI security solution.