What Is Shadow AI? Managing Employee ChatGPT Use Safely
Nanda Teknoloji Engineering Team · · 8 min read
In brief
Shadow AI is employees using AI tools such as ChatGPT, Claude, or Gemini with business data without the knowledge of IT and security teams. Banning usually does not work; managing it with visibility, an approved alternative, a clear policy, and DLP controls is more sustainable.
Shadow AI is employees using generative AI tools such as ChatGPT, Claude, and Gemini with business data without the knowledge or approval of the IT and security teams. The intent is usually good: drafting an email, summarizing meeting notes, debugging code. The problem is that in doing so, customer information, contract text, or source code is sent to a service outside the organization’s control. Once data leaves, where it is stored, who can see it, and how it will be deleted are no longer in the organization’s hands. The solution is not to ban, but to gain visibility, offer an approved alternative, write a clear policy, and back it up with technical controls. This article explains where the risk comes from and how you can manage it step by step.
Why is Shadow AI spreading?
Access to generative AI tools takes a browser and a few minutes of sign-up; no installation, purchase approval, or IT support is needed. The productivity gain is also immediately visible to the employee. Corporate processes, on the other hand, take weeks with the steps of evaluating a new tool, legal review, and procurement. This speed gap naturally creates a vacuum, and employees fill it with personal accounts. For the definition of the concept, you can also see the Shadow AI glossary entry.
The main risks of Shadow AI
| Risk | How it arises | Example |
|---|---|---|
| Data leakage | Sensitive data sent to an external service as a prompt | Pasting a customer list to have it summarized |
| Personal data and KVKK | Personal data transferred to a service abroad or one without a contract | Using HR records for analysis |
| Loss of intellectual property | Sharing source code or a product plan | Pasting an entire module for debugging |
| Unreliable output | The model producing a wrong but convincing response | Unverified information in a legal or financial text |
| Compliance and contract violation | A customer contract prohibiting data sharing with third parties | Uploading a customer document to an unapproved tool |
| Identity and access | Personal accounts and extensions gaining access to corporate data | An AI browser extension obtaining permission to read email content |
For the personal data dimension, it is wise to consult the resources of the KVKK Authority and your organization’s legal department.
Visibility: first find out what is being used
You cannot manage what you cannot measure. The first step is to work out which AI tools are used in which units, and for what purpose.
- Network and security gateway logs. Show the traffic volume and user distribution toward known AI service domains.
- SaaS and browser extension inventory. AI extensions installed in corporate-managed browsers are listed.
- Identity provider records. Accounts opened with corporate email and OAuth permissions are reviewed.
- Employee surveys and interviews. They reveal personal device use that technical monitoring cannot see, and the real needs. Use a non-accusatory tone; the goal is not to punish but to understand the need.
- Department-based use cases. Marketing, software, legal, and HR carry different risks; evaluate them separately.
Management strategy: steer instead of banning
In most organizations, a full ban leads to this: use shifts to personal phones and home computers, and visibility drops to zero. A more resilient approach has four pillars.
1. Offer an approved alternative
Provide a tool or in-house solution that meets employees’ needs and has been evaluated against corporate conditions (data retention, no use in training, access control). If the approved path is not easy and useful, no one will switch to it.
2. Write a clear, short policy
The policy should say, in table form, which data types may be sent to which class of tool. An example classification might look like this:
| Data class | Unapproved public tool | Approved corporate tool | In-house (private) model |
|---|---|---|---|
| Public information | Allowed | Allowed | Allowed |
| Internal use | Prohibited | Allowed | Allowed |
| Customer and personal data | Prohibited | Conditional, anonymized | Allowed |
| Confidential (source code, finance) | Prohibited | Prohibited or specially approved | Conditional |
For writing the policy, you can take a look at our article on the corporate AI usage policy.
3. Back it up with technical controls
A policy alone is not enough; you need controls that make mistakes harder. AI DLP solutions detect content in prompts such as personal data, ID numbers, and secret keys before they are sent, and mask or block them. An AI Gateway routes corporate AI traffic through a single point, centralizing access control, logging, and policy enforcement. The framework the two form together is covered under AI security.
4. Train, and build a feedback loop
Employees need to be shown “how to do it safely” as much as “what not to do”: removing and generalizing sensitive information, using sample data, not using a response without verifying it. A quick review channel that collects new tool requests from users reduces shadow use.
Frameworks and standards
To approach AI risk management holistically, the NIST AI Risk Management Framework offers guidance with its govern-map-measure-manage logic. For technical risks specific to AI applications, the OWASP Top 10 for LLM Applications list is the reference source.
Implementation plan: the first 30 days
- Week 1: Appoint an executive sponsor; form a small team from security, legal, HR, and IT.
- Week 2: Visibility work: logs, extension inventory, a short survey.
- Week 3: Write the data classification table and a first draft of the usage policy; evaluate approved tool candidates.
- Week 4: Open the approved tool to a pilot group, enable DLP and logging controls, and publish the training material.
- Ongoing: Monthly usage report, incident log, and policy review.
A scenario: a well-intentioned violation in the legal department
An employee in a company’s legal department pastes the text of a long supplier contract into a public AI tool to have it summarized. The contract contains commercial terms under a confidentiality clause and the counterparty’s personal data. The employee has no bad intent; the aim is to save time. However, the confidentiality obligation in the contract may have been breached, and personal data has been transferred to a third-party service. If what to do when the incident is noticed has not been defined in advance, the process gets muddled. In a prepared organization, by contrast, there is an approved tool or in-house model for contract types, the policy table explicitly addresses this data class, and the DLP control catches ID numbers and confidentiality language in documents like contracts and raises a warning.
Incident management: when data is sent to the wrong tool
Despite every control, a violation can occur. Employees being able to report without fear of punishment is the most effective way to limit the damage.
- Open a reporting channel. Set up a short form that can be anonymous and name a clear owner.
- Determine the scope. Which data was sent, to which tool, and with which account? Does the tool allow logging and deletion requests?
- Make a legal and contractual assessment. If personal data or customer data is involved, notification obligations are evaluated with the legal department.
- Take the necessary steps. A request to delete the account and history, closing the relevant sharing links, and informing the relevant parties if needed.
- Make a lasting improvement. Find which gap the incident came from: lack of training, an inadequate approved tool, or a technical control gap.
Common mistakes
- Only banning. Usage goes underground and visibility disappears.
- Not offering an approved alternative. The need persists, and the solution path is the employee’s own account.
- Writing the policy long and vague. No one reads it; tables and examples are more effective.
- Relying only on technical controls. DLP can miss contextual sensitive information that has no signature; it should be run together with training.
- Forgetting personal devices. Use outside the corporate network needs policy and managed device rules.
- Not verifying output. Text, code, or legal interpretation produced by the model must not be used without human review.
Decision summary
| Your situation | First step |
|---|---|
| You don’t know the usage | Visibility work and a survey |
| There is usage but no policy | A data classification table and a short policy |
| There is a policy but weak enforcement | Technical control with AI DLP and AI Gateway |
| You work with sensitive data | Evaluate an approved corporate or in-house model |
Moving forward with Nanda Teknoloji
On projects like this, Nanda Teknoloji carries out the discovery, design, migration, and operations steps: working out existing AI usage, designing the policy and data classification, and then enabling the DLP and gateway layers. You can find the general framework on the AI security page.