Microsoft Copilot & Data Protection

How Companies Can Use AI Without Losing Sensitive Data

Microsoft Copilot is one of the most visible examples of how artificial intelligence is transforming the digital workplace: summarizing emails, analyzing documents, creating presentations, or automatically documenting meetings — all of this is suddenly possible within seconds.

The productivity gains are real. And they are significant.

However, this new level of efficiency raises an uncomfortable question that many companies are currently facing:

What happens to our sensitive data when we use Copilot?

This is not a marginal issue. It touches the core of modern IT strategies. Copilot only works effectively when it has access to data. And that is exactly where the problem lies. The risk of data leakage or access by US hyperscalers under the CLOUD Act remains.

This article explains how Copilot works technically, where the real risks lie, and how companies can use AI without losing control over their data.

Why Copilot Doesn’t Work Without Data

To understand why Copilot is critical from a data protection perspective, it is important to first understand how the system works.

Copilot is not an isolated tool. It is deeply integrated into Microsoft 365 and accesses multiple data sources, including:

• Emails in Outlook
• Documents in SharePoint and OneDrive
• Chats and files in Microsoft Teams
• Calendar data
• Organizational knowledge

This data is used to generate context-based responses. The more context Copilot has, the better the results.

But this also means:

Copilot requires access to content — in plain text.

And this is where the real challenge begins.

The Core Risk: Plain Text Access to Corporate Data

Many discussions around Copilot focus on privacy policies, contractual clauses, or data storage locations.

However, the real risk lies deeper — on a technical level.

Copilot processes data in a way that makes it interpretable for AI. This means:

• Content must be readable
• Content must be analyzable
• Content must be processable

In short:

AI requires access to plain text data.

This raises several critical questions:

• What data is being processed?
• Who can technically access it?
• Where does processing take place?
• What level of control does the company have?

Even if providers maintain high security standards, a structural issue remains:

As soon as data is processed in plain text, an access point exists.

Why Traditional Security Measures Are Not Enough

Many companies rely on existing security mechanisms within Microsoft 365:

• Access controls
• Role and permission concepts
• Audit logs
• Compliance features

These measures are important — but they do not solve the core problem.

They define:

Who within the organization has access.

But they do not prevent:

• Systems themselves from processing data
• Platforms from accessing content
• External components from analyzing data

In other words: the platform provider can technically access and process all plain text data — no matter how sensitive or business-critical it is.

A common misconception is:

“Our data is secure because we control permissions.”

In reality:

Permissions do not protect against system-level access.

And this is exactly what matters in AI systems.

The Core Conflict: Productivity vs. Data Sovereignty

This creates a classic dilemma in modern IT:

The more data Copilot can access, the better it performs.
The more data is protected, the less AI can use it.

This is not a configuration issue — it is a structural conflict.

A Pragmatic Approach: Separating Data

Instead of trying to resolve this dilemma, modern security architectures take a different approach:

They accept the conflict and separate the data.

This model has proven highly effective in practice:

Separation between plain text and encrypted data

The logic is simple:

• Non-sensitive data remains in plain text → Copilot can use it
• Sensitive data is encrypted → Copilot cannot access it

The result:

• AI remains usable
• Sensitive data remains protected

How Secure Copilot Usage Works in Practice

With a solution like eperi sEcure, this separation is implemented technically.

This means:

Copilot continues to work for:

• General research
• Document overviews
• Text suggestions
• Meeting summaries
• Non-confidential content

At the same time, the following are protected:

• Confidential documents
• Sensitive emails
• Internal communication
• Business-critical information

This data is stored in encrypted form and is not accessible to Copilot.

The Key Advantage: Security by Design

The fundamental difference compared to traditional approaches lies in the architecture.

Instead of trying to control access, this model ensures:

That certain data is technically inaccessible.

This is a fundamental shift.

Because:

• What is not accessible cannot be processed
• What is not processed cannot be exposed

This is data sovereignty at a technical level.

What About Using AI with Sensitive Data?

So far, we have a model that works well:

• Copilot for non-sensitive data
• Protection for sensitive data

But many companies are now asking the next question:

What if we want to use AI with sensitive data as well?

This is often where the greatest value lies:

• Contract analysis
• Internal knowledge bases
• Compliance evaluations
• Customer data
• Insurance data

And this is where the traditional Copilot model reaches its limits.

Why Public AI Is Not Designed for Sensitive Data

Systems like Copilot or ChatGPT belong to the category of Public AI.

These systems are designed to:

• Work with plain text
• Process data within their platform
• Deliver fast results

However, they are not built to process highly sensitive data under strict regulatory requirements.

This means:

• If data is protected, it cannot be used
• If it is to be used, it must be exposed

A classic trade-off.

The Next Evolution: Confidential AI

This is where a concept comes into play that is gaining significant importance:

Confidential AI

The idea is simple — but powerful:

Sensitive data remains protected and can still be used by AI.

How Confidential AI Works

Confidential AI is based on three core principles:

1. Encryption remains in place
   Data stays protected and is only exposed in a controlled way for AI processing.
2. Key control remains with the company
   The company decides when and how data is used.
3. Processing takes place in controlled environments
   AI operates in isolated, secure execution environments.

The result:

Data is processed without companies losing control.

Why Cloud Encryption Is the Foundation

Confidential AI only works if one key requirement is met:

Data must already be protected.

This is where cloud encryption comes into play.

A solution like eperi sEcure ensures that:

• Data is stored encrypted in the cloud
• Keys are not held by the cloud provider
• Access is technically controlled

This creates the foundation for:

• Secure Copilot usage
• Confidential AI

A New Model for AI in the Enterprise

Combining both approaches leads to a clear model:

Layer 1: Copilot (Public AI)

• Use for non-sensitive data
• Fast productivity gains
• No use for sensitive content

Layer 2: Confidential AI

• Use for sensitive data
• Controlled processing
• Full data sovereignty

This model is not only technically sound — it is becoming a best practice.

What Companies Should Do Now

Introducing Copilot is not just an IT project — it is a strategic decision.

Companies should ask themselves three key questions:

1. Which data can be processed by AI?
2. Which data must remain protected?
3. Which architecture supports both?

This is the real challenge.

Conclusion: Use AI Without Losing Control

Microsoft Copilot and other public AI models are powerful tools. They will fundamentally change how we work.

But with this new technology comes a new responsibility:

Data must be protected — not only organizationally, but technically.

The good news:

Companies do not have to choose between innovation and security.

With the right architecture, both are possible:

• Copilot for productivity
• Encryption for protection
• Confidential AI for sensitive data

Or in other words:

The future does not belong to Public AI — it belongs to Confidential AI.