Preventing Shadow AI: Technical and Organizational Measures for Companies

ChatGPT, Claude, Gemini, Copilot, and numerous smaller AI services are just a few clicks away for employees. That is exactly where the problem lies for companies: while the official IT department is still debating privacy, contracts, and approvals, employees may already have started using AI for emails, documents, source code, spreadsheet analysis, or customer data.

This unauthorized use of AI systems is often called shadow AI.

For companies, shadow AI is not just a privacy issue. Uncontrolled AI use can carry trade secrets, personal data, credentials, source code, and internal documents outside the organization.

A blanket ban on ChatGPT and other AI services rarely solves the problem.

Anyone who wants to effectively prevent, or at least significantly reduce, shadow AI needs a combination of technical controls, clear rules, secure alternatives, and trained employees.

What is shadow AI?

Shadow AI refers to the use of AI applications within a company that have not been officially reviewed, approved, or controlled by the relevant departments.

Shadow AI is therefore a modern variant of the already familiar shadow IT. We covered the underlying risks and typical triggers in detail in Shadow AI at Work: The Risks of Unapproved ChatGPT Use (Read article). This article picks up exactly where that one left off and shows the concrete technical and organizational implementation.

A typical example:

An employee needs to draft a lengthy customer email. The company does not officially provide an AI system. So the employee opens ChatGPT or another AI service privately and pastes the original customer request into the input field.

Within seconds, an entire customer document may have been transferred to an external service.

Similar situations happen every day:

• Developers have proprietary source code analyzed.
• Employees upload spreadsheets containing customer data.
• HR departments have applications summarized.
• Administrators paste log files or configurations into a chatbot.
• Sales staff have proposals and contracts optimized.
• Support staff submit customer tickets to public AI services.
• Employees install browser extensions with built-in AI.
• Staff use private AI accounts for work tasks.

The problem is not necessarily the AI itself.

The problem is the lack of control over which data goes where, and how it is subsequently processed.

Why shadow AI is dangerous for companies

With classic SaaS applications, IT can usually check which service is being used, who is allowed to access it, and which data is being processed.

Shadow AI bypasses this process. As a result, several risks arise at once.

1. Loss of confidential company data

Employees often don't recognize which information is actually confidential.

A seemingly harmless prompt can contain:

• internal server names
• IP addresses
• customer names
• contract information
• revenue figures
• source code
• passwords or API keys
• internal URLs
• ticketing system data
• security configurations

Full file uploads are especially problematic. A spreadsheet, for example, can contain numerous additional sheets and metadata beyond the data actually needed.

2. Privacy problems from personal data

Once personal data is transferred to an external AI system, it must be checked whether this processing is permissible under data protection law.

Article 5 of the GDPR requires, among other things, purpose limitation, data minimization, and appropriate security and confidentiality of personal data. Companies must also be able to demonstrate compliance with these principles.

An employee normally cannot carry out this assessment themselves when spontaneously using an online AI service. That is precisely why not just any AI service should be allowed to handle company data.

3. Trade secrets can be exposed

Even before personal data is affected, shadow AI can become economically dangerous.

Source code, product strategies, price lists, contract details, technical documentation, or planned projects may be among a company's most valuable information.

Germany's BSI (Federal Office for Information Security) explicitly recommends that companies identify and classify their especially important data and protect it accordingly. This data should therefore not be copied into public AI services without control.

4. Unknown AI providers become part of the supply chain

An AI service is ultimately an external service provider. Companies should therefore check, among other things:

• Where is the provider based?
• Where is data processed?
• Which subprocessors are used?
• How long are inputs stored?
• Are inputs used for training or product improvement?
• Is there a data processing agreement in place?
• Which security certifications exist?
• Can administrators centrally manage users?
• Are there audit and compliance features?
• Does the service support SSO?
• Can usage be logged?
• What options exist for data deletion?

Many of these controls are missing with a private user account.

5. Companies don't know which AI is actually being used

Without monitoring, another problem arises: IT cannot protect what it doesn't know about.

Beyond the well-known platforms, thousands of specialized AI services now exist for translation, presentations, programming, image generation, transcription, research, marketing, or document analysis.

That's why it isn't enough to simply block a few well-known domains.

Just banning shadow AI? Why that usually doesn't work

An obvious reaction is: "Let's just block ChatGPT."

Technically, that's certainly possible. A company could, for example, block known AI domains via firewall, DNS filtering, secure web gateway, proxy, CASB, or endpoint security.

But that doesn't make shadow AI disappear. Employees can still:

• use other AI services,
• use personal smartphones,
• use mobile data connections,
• install browser extensions,
• use personal devices,
• process data at home, or
• use as-yet-unknown AI platforms.

A complete technical ban is therefore difficult to enforce in many environments.

A much more sensible approach is: don't ban AI — make unsafe AI use unnecessary through secure, easy-to-use alternatives.

Technical measures against shadow AI

1. Define approved AI services

The most important step is an official AI allowlist. The company decides which AI services may be used. For example:

Category	Rule
Approved company AI	Use permitted
Approved AI with restrictions	Only for non-confidential data
Unreviewed AI services	Use not permitted
Private AI accounts	Not permitted for company data
Local AI	Permitted depending on internal security policy

This creates a clear foundation for all further technical controls.

2. Use company accounts instead of private accounts

Approved AI systems should ideally be provided through centrally managed company accounts. Ideally, the platform supports:

• single sign-on
• central user management
• roles and permissions
• MFA
• audit logs
• central billing
• deactivation of departed employees
• defined retention periods
• administrative privacy options

This gives IT control over the systems in use again.

3. Use SSO and an identity provider

Where possible, AI use should be tied to the company identity. Typical platforms for this include:

• Microsoft Entra ID
• Okta
• Keycloak
• Google Workspace Identity

This makes it possible, for example, to define: employee group A may use AI, group B may not.

Conditional access rules can also be applied. An AI platform could, for example, be reachable only from managed devices and only after successful MFA login.

4. Use a secure web gateway, proxy, or CASB

Central web control is one of the most effective technical measures against shadow AI. Depending on the infrastructure, solutions such as secure web gateway, CASB, SSE, SASE, DNS security, or a classic corporate proxy come into question.

Filtering shouldn't only be based on individual domains. Modern systems can recognize SaaS and GenAI services by category and log or restrict their use.

The goal: identify unknown AI services before sensitive data ends up there.

5. Use data loss prevention

One of the most important technologies against shadow AI is DLP – data loss prevention. DLP systems can prevent certain information from being transmitted externally via web forms, file uploads, or other channels.

What can be detected, for example:

• credit card numbers
• personal data
• internal document classifications
• customer numbers
• personal identifiers
• confidential file types
• source code
• specific keywords

A good DLP strategy doesn't necessarily prevent AI use itself. Rather, it prevents critical company data from ending up in unapproved AI services.

6. Classify data consistently

DLP works significantly better when companies already know how sensitive their data is. A simple classification can already consist of four levels:

Public – information that may be published anyway.

Internal – information for employees whose disclosure would nonetheless cause no serious harm.

Confidential – customer data, internal calculations, contracts, or technical information.

Strictly confidential – trade secrets, credentials, security-critical information, or especially sensitive personal data.

This classification can then be directly linked to the AI policy.

7. Control browsers and browser extensions

Shadow AI is no longer limited to websites. More and more extensions integrate AI directly into the browser. They can, for example:

• analyze web pages,
• read emails,
• rephrase text,
• automatically fill in forms,
• summarize documents.

Company browsers should therefore be centrally managed. For browser extensions, an allowlist is generally preferable to unrestricted installation of any extension.

8. Centrally manage endpoint devices

Additional controls can be set up on managed company devices, including:

• MDM
• endpoint management
• application control
• endpoint DLP
• EDR/XDR
• software inventory
• browser policies

This makes it possible to detect when new AI applications or AI clients are installed on company devices.

9. Make AI use visible in the SIEM

Companies with central security monitoring should also account for AI services. Proxy, firewall, DNS, CASB, and endpoint data can, for example, be forwarded to a SIEM. Interesting events would include:

• first-time access to an unknown AI service
• unusually large uploads
• use of private AI accounts
• mass access to GenAI services
• data transfers from especially protected areas
• installation of new AI browser extensions

It's important that this monitoring serve information security and not be used secretly to monitor employee performance.

In Germany, introducing technical systems designed to monitor employee behavior or performance can trigger co-determination rights under Section 87 of the Works Constitution Act (BetrVG). Section 90 BetrVG now also explicitly names the use of artificial intelligence in work procedures and processes.

Data protection officers and, where one exists, the works council or staff representative body should therefore be involved early on.

10. Deploy a central AI gateway

For larger companies, an AI gateway can be an interesting solution. Employees don't access different AI APIs directly. Instead, communication runs through a controlled company service:

Employee → AI gateway → approved AI model

The gateway can, for example:

• perform authentication,
• approve models,
• centrally manage API keys,
• log inputs,
• detect sensitive data,
• remove personal information,
• cap costs,
• set rate limits,
• provide different models.

This approach is especially interesting for developers. Instead of using ten different AI APIs with ten different API keys, there is a single, centrally controlled access point.

The same least-privilege principle also applies to AI agents that are allowed to access such a gateway or company systems directly. We explain how to properly limit permissions for AI agents in AI Agent Permissions: Security Rules & Least Privilege (Read article).

11. Offer local AI for especially sensitive data

Another option is running local AI models. Technologies like Ollama or other self-hosted inference platforms make it possible to run certain language models within your own infrastructure.

That doesn't automatically mean local AI is secure. There too, access control, patch management, model provenance, logging, network segmentation, permissions, and privacy must be considered.

For especially sensitive use cases, an internal AI platform can still be a useful addition. Germany's BSI explicitly points out that, alongside AI-specific measures, classic security mechanisms such as access controls and the isolation of system environments remain important for generative AI. We cover exactly how to harden an Ollama and Open WebUI setup like this for privacy and GDPR compliance in Local AI for Companies: Running Ollama and Open WebUI Safely (Read article).

Organizational measures against shadow AI

Technical measures alone aren't enough. Employees need to know what's allowed and why certain things are prohibited.

1. Create a clear AI policy

Every company that uses AI, or whose employees have internet access, should by now be thinking about an AI policy. We show in detail how such a policy can be structured in AI Usage Policy for Companies (Read article). At minimum, it should address:

• Which AI services may be used?
• Which accounts must be used?
• Which data may be entered?
• Which data is fundamentally prohibited?
• May files be uploaded?
• May source code be analyzed?
• May AI be used for customer communication?
• Do results need to be reviewed?
• How are new AI services requested?
• Who answers questions?

An 80-page compliance policy that nobody reads doesn't help much. A clear, understandable set of rules that employees can grasp within a few minutes is far more effective.

2. Introduce a simple traffic-light model

For many companies, a traffic-light model works particularly well.

Green – allowed. For example: publicly available information, general text correction, brainstorming, idea generation, generic wording, content with no company reference.

Yellow – only with approved company AI. For example: internal documents, internal processes, anonymized data, program code, technical documentation, internal meeting notes.

Red – never enter into external AI systems. For example: passwords, API keys, private keys, highly sensitive personal data, trade secrets, non-anonymized customer data, confidential contract information, security-critical credentials.

This lets employees assess a situation much more quickly.

3. Create a simple approval process

A common cause of shadow IT and shadow AI is bureaucracy. If reviewing a new tool takes six weeks while an employee needs to get a task done today, there's a strong incentive to just use the tool anyway.

That's why there should be a simple process:

Employee discovers AI tool → submits a request → IT/privacy reviews it → approval or rejection.

For small companies, a form or a ticket in the existing ticketing system may already be enough.

4. Maintain an AI inventory

Approved AI applications should be documented. The inventory could, for example, contain the following information:

Information	Example
AI application	Company chatbot
Responsible department	Marketing
Purpose	Text creation
Data category	Internal
Provider	Provider X
Contract reviewed	Yes
Privacy reviewed	Yes
Approved until	2027-12-31

Such a registry prevents a situation where nobody knows which AI systems are actually in use across the company.

5. Train employees regularly

Employees need to understand why certain rules exist. An effective AI security training shouldn't just say: "Don't enter customer data into ChatGPT." It should show how quickly a mistake can happen.

For example: "Please analyze this log file and find the error." Sounds harmless. But the log file may contain:

• usernames
• internal hostnames
• IP addresses
• session IDs
• internal URLs
• email addresses
• tokens

Examples like this build real understanding.

The EU AI Act requires providers and deployers of AI systems to take measures to promote sufficient AI literacy among staff and other people who use AI systems on their behalf. Since August 2, 2026, further key provisions of the AI Act apply or are being enforced. We explain exactly what this AI literacy obligation means in AI Literacy 2026: EU AI Act Training and Certificates (Read article).

AI training is therefore no longer just a security topic — it is increasingly part of structured AI governance.

6. Management must follow the rules itself

An AI policy doesn't work if executives and managers keep uploading confidential documents to private AI accounts themselves. Information security must therefore be supported by company leadership.

Germany's BSI also explicitly treats cyber security as a strategic issue and a leadership responsibility.

7. Create a point of contact for AI

Employees should know who to turn to. This could, for example, be:

• the IT department
• an information security officer
• a data protection officer
• an AI officer
• an internal AI governance team

The specific setup depends on company size. What matters is not the title, but the fact that a defined process exists at all.

Shadow AI and NIS2

For companies subject to Germany's NIS2 regulation, this topic carries additional weight.

Germany's NIS2 Implementation Act took effect on December 6, 2025. Since then, the BSI has pointed out the corresponding registration and reporting obligations.

The underlying NIS2 Directive requires a risk-based approach to cybersecurity and names, among other things:

• risk analysis
• incident handling
• supply chain security
• secure procurement and maintenance of IT systems
• cyber hygiene
• training
• access control
• asset management
• multi-factor authentication, where applicable

Unapproved AI services should therefore also be factored into cyber risk management at affected organizations.

Which departments should tackle shadow AI together?

Shadow AI is not a purely IT problem. At least the following areas can be affected:

• IT: technical approvals, identity management, device management, and network access.
• Information security: risk analysis, monitoring, and security requirements.
• Privacy: review of personal data and external providers.
• Legal: contracts, liability, copyright, and trade secrets.
• Procurement: sourcing and contract management.
• HR: training and internal policies.
• Works council: co-determination on technical systems and AI use, where applicable.
• Executive management: risk ownership and strategic direction.

This kind of collaboration is exactly what distinguishes real AI governance from simply blocking ChatGPT.

A 30-day plan against shadow AI

Companies don't need to launch a multi-million-dollar AI security project overnight. A pragmatic start can happen within a few weeks.

Week 1: Take stock

Determine:

• Which AI services are already in use?
• Which departments use AI?
• Which data is being processed?
• Are there private accounts in use?
• Which browser extensions are installed?

Week 2: Define rules and approved systems

Establish:

• approved AI services
• prohibited data
• data classification
• responsibilities
• an approval process for new tools

Week 3: Activate technical controls

Depending on infrastructure:

• SSO
• MFA
• DNS filtering
• proxy
• CASB/SSE
• DLP
• browser policies
• application control
• SIEM rules

Week 4: Inform and train employees

Then:

• publish the AI policy
• run a short training session
• name a point of contact
• collect feedback
• regularly review new AI applications

This already creates a solid foundation.

The best shadow AI strategy: enable instead of just banning

Companies need to accept a reality: employees use AI because it lets them get tasks done faster.

Banning all AI systems doesn't automatically eliminate the demand. In the worst case, usage simply shifts from the controlled company network to personal devices and accounts.

The more sustainable strategy is therefore: make secure AI easily available. When an employee has an officially approved company account, the incentive to use an unknown free AI service is much lower.

This changes IT's actual task. Not: "How do we prevent all use of artificial intelligence?" But rather: "How do we enable our employees to use AI without losing control over company data?"

That should be the goal of modern AI governance.

Checklist: preventing shadow AI in your company

Companies should check at least the following points:

• Define approved AI services
• Prohibit private AI accounts for company data
• Create an AI policy
• Classify data
• Configure DLP rules
• Use SSO and MFA
• Control browser extensions
• Review a secure web gateway or proxy
• Deploy CASB/SSE where needed
• Account for AI use in security monitoring
• Review an AI gateway for API use
• Maintain an AI inventory
• Vet providers before approval
• Train employees regularly
• Involve privacy teams early
• Involve the works council where applicable
• Create an approval process for new AI tools
• Consider local or private AI for sensitive use cases
• Regularly update the rules

Frequently asked questions about shadow AI

What is shadow AI?

Shadow AI refers to the use of AI systems within a company that have not been officially reviewed, approved, or controlled by the relevant departments. Employees may use private ChatGPT accounts, AI websites, browser extensions, or other AI services for work tasks without IT, privacy, or information security ever finding out.

Can a company block ChatGPT entirely?

Known AI services can be blocked via DNS, firewall, proxy, or secure web gateways. Fully preventing all AI use is much harder, however, since employees can turn to other services, personal devices, or mobile data connections.

Which data should never be entered into public AI systems?

Passwords, API keys, private keys, trade secrets, confidential contract information, and sensitive personal or non-anonymized customer data are especially critical.

Is an AI policy enough on its own?

No. A policy is important, but it should be complemented by technical measures such as identity management, DLP, browser controls, logging, and, where appropriate, web filtering.

Is DLP enough to prevent shadow AI?

Also no. DLP can prevent or detect the leakage of certain data, but it does not control every form of AI use. DLP should therefore be part of a broader AI governance and security strategy.

Do employees need to be trained on AI use?

Training is strongly recommended for security reasons. In addition, Article 4 of the EU AI Act requires measures to promote sufficient AI literacy among staff and other people who use AI systems on behalf of providers or deployers.

Conclusion: preventing shadow AI means regaining control

Shadow AI will become a challenge for companies over the coming years similar to shadow IT – probably an even bigger one. That's because using an AI service often doesn't even require installing software. A browser is enough.

That's why a strategy of "we'll block ChatGPT and the problem is solved" doesn't hold up long-term. Companies instead need multiple layers of protection:

Technical controls protect data and make unknown services visible. Organizational rules define what's allowed. Approved AI systems create secure alternatives. Training helps employees recognize risks themselves. And AI governance ensures that new AI applications don't become part of the company's IT in an uncontrolled way.

The goal shouldn't be to keep artificial intelligence out of the company. The goal is to use it in a controlled, traceable, and secure way.

Companies that provide employees with suitable AI tools while technically protecting sensitive data address shadow AI at its root cause — instead of constantly chasing after the next AI service to block.

Sources and further reading

This article reflects the legal and factual status as of September 10, 2026. Primary sources used include the European Commission on the current enforcement status of the EU AI Act, the consolidated version of the AI Act on EUR-Lex, the GDPR, information from Germany's BSI on cyber security and generative AI, the German Works Constitution Act, and Germany's NIS2 Implementation Act.

Note: This article does not constitute individual legal or data protection advice.

Related topics and sources

Shadow AI at Work: The Risks of Unapproved ChatGPT Use (Read article)

AI Usage Policy for Companies (Read article)

AI Agent Permissions: Security Rules & Least Privilege (Read article)

AI Literacy 2026: EU AI Act Training and Certificates (Read article)

EU AI Act for Small Businesses: What SMEs Need to Know (Read article)

Local AI for Companies: Running Ollama and Open WebUI Safely (Read article)