A similar pattern is currently emerging in many companies: Employees want to use generative AI productively but are limited by severely restricted internal solutions. Often, these solutions lack file uploads, powerful models, data integrations, or features that public offerings already provide.

The reasoning behind this is understandable. Companies want to comply with data protection, information security, and regulatory requirements. However, problems arise when the solution implemented is so technically limited that it offers little to no added value for day-to-day work.

This then creates a second problem: Employees turn to personal accounts or unsanctioned tools. As a result, the risk is simply shifted to an area that is less controllable.

The key question, therefore, is not whether companies should adopt AI. Rather, it is what a safe yet practical use of AI might look like.

Can companies use OpenAI, Microsoft, Google, or Anthropic?

Basically, yes.

Neither the Swiss Data Protection Act nor the GDPR prohibits the use of cloud-based AI systems. What matters is how a service is used, what data is processed, and what contractual and technical safeguards are in place.

If a service provider processes personal data on behalf of a company, this data processing must be governed by a contract. In practice, this is done through a Data Processing Agreement (DPA) or a data processing contract.

With regard to cloud services, the FDPIC points out that companies remain responsible for data processing even when they outsource it. In particular, they must vet the provider, define the purpose of the processing, ensure appropriate security measures, and take international data transfers into account.

A DPA is therefore an important prerequisite. However, it alone is not enough.

What Companies Should Check Before Release

Before introducing an AI service, companies should clarify at least the following points:

  • Business License: A Business, Team, or Enterprise version should be used for company data. These should be clearly separated from personal and free accounts.
  • DPA or Data Processing Agreement: The agreement must specify how the service provider processes personal data on behalf of the company.
  • Use of Data for Training: The company needs to know whether inputs, files, and outputs are used to improve or train general models.
  • Storage and Deletion: It is important to know how long data is stored and what options are available for deletion and retention.
  • Data Location and Cross-Border Transfers: What matters is not only where data is stored, but also where it is processed and which subprocessors are involved.
  • User Management: Depending on the level of risk and the size of the company, SSO, MFA, roles, permissions, logging, and, if applicable, DLP should be available.
  • Data Classification: Employees must know which public, internal, confidential, or particularly sensitive data may be processed in which system.
  • Use Case: A writing assistant should be evaluated differently than a system that supports personnel decisions, credit decisions, or other sensitive processes.
  • AI Literacy: Employees must understand what data they are permitted to use, how to verify results, and what the system's limitations are.

This creates a manageable framework without unnecessarily restricting usage.

What Companies Fear and What Is Actually Possible

Many concerns about cloud-based AI revolve around the same questions: Could confidential data end up with the provider? Will user input be used for training? Do foreign authorities have access to the data? Can employees upload information without supervision? And does the company lose control over its data?

These concerns are fundamentally valid. However, they do not automatically mean that cloud AI must be ruled out.

The specific details are what matter.

OpenAI

When it comes to OpenAI, a distinction must be made between consumer offerings and business offerings.

ChatGPT Business and Enterprise are designed for business use. According to OpenAI, company data is not used to train the general models by default. Contractual and administrative features are also available to business customers.

When using the system with corporate data, companies should pay particular attention to the DPA, data retention, user management, data residency, and the features used.

From a legal standpoint, what matters is not that OpenAI is a U.S. company, but whether data processing and international data transfers are carried out on a lawful basis.

Microsoft

Microsoft Copilot is a particularly obvious choice if a company already uses Microsoft 365.

A key advantage is that Copilot can access existing user permissions and company data. But this is also where the risk lies: If SharePoint, Teams, or file permissions have historically been set too liberally, AI can make these existing permissions visible.

Before implementation, therefore, it is important to evaluate more than just Copilot. The existing information architecture and rights management must also be in order.

From a data protection perspective, commercial use is governed by the Microsoft agreements and the corresponding DPA.

Google

For companies using Google Workspace, Gemini offers a similarly integrated solution.

Here, too, a clear distinction must be made between business Workspace offerings and personal Gemini accounts. Specific privacy policies and terms of service apply to the business versions.

In particular, companies should review which version of Workspace is being used, what administrative options are available, and how data retention, training, and data transfer are handled.

Anthropic and Claude

Claude can be particularly useful for document analysis, longer contexts, and development tasks.

At Anthropic, too, the distinction between consumer and commercial products is essential. For business data, companies should use Claude for Work or the corresponding API or enterprise offerings.

DPA, subcontractors, data retention, and training should be reviewed in the same way as with other providers.

Swiss providers

Swiss providers can offer additional benefits if data storage in Switzerland or a local contractual partner is important.

However, “Swiss-hosted” should not automatically be equated with “entirely Swiss AI.”

The following questions, among others, are relevant: Which models are used? Where does the inference take place? Are external APIs used? Are web search or other third-party providers involved? Which subcontractors are involved? And under what conditions is data stored?

You can find an overview of these offerings on swissAI under " Swiss AI Chats."

The question of which provider to choose therefore cannot be answered simply with “safe” or “unsafe.” Microsoft, OpenAI, Google, Anthropic, and Swiss providers can, in principle, be used within a data-protection-compliant enterprise architecture. The prerequisite is that the contract, product, configuration, data, and specific intended use are all compatible.

AI licenses do not cost the same for all users

It’s also worth taking a nuanced look at the costs. Not every employee needs the same features and usage limits.

  • Office users, approx. CHF/USD 20 to 30 per month: texts, summaries, simple file analysis, occasional research.
  • Heavy user, approximately CHF/USD 50 to 100 per month: frequent searches, larger documents, more complex analyses, agents.
  • Power users, developers, and analysts, approximately CHF/USD 100 to 200 or more per month: heavy usage, coding, large amounts of data, multiple models, and specialized features.

It usually makes more sense to tailor licensing to actual needs than to assign the same license to all employees.

At the same time, a power user should not be artificially restricted to a very small usage quota. If work is regularly interrupted by limits, the overall benefit of the implementation is diminished.

Why It Often Makes Sense to Use More Than One AI Tool

Even a strategy based solely on a single tool may not be the best fit for every company.

Microsoft Copilot excels within the Microsoft ecosystem. ChatGPT offers different capabilities in research, analysis, multimodal processing, and agents. Claude may be advantageous for certain document and development tasks. Additional specialized systems are available depending on the field.

In addition, the performance of the models is constantly changing.

A practical model could therefore involve offering a standard tool that is available to the majority of employees and providing certain user groups with additional tools or more powerful licenses.

The swissAI overview, “AI Models Compared,” shows which models are suitable for different tasks.

In-House Solution or Existing Platform: Benefits, Effort, and Costs

Having your own internal AI platform can be beneficial, especially when specific processes, special regulatory requirements, or strategically important internal data sources need to be integrated.

However, it should not automatically be considered a safer or cheaper option.

An internal AI platform is not a one-time development project. It must be operated and further developed on an ongoing basis.

Models change, APIs are updated, new features are introduced, security updates must be applied, and existing integrations can be affected by changes to external systems. At the same time, authentication, file uploads, data access, roles, monitoring, logging, and the user interface must be maintained.

This requires more than just technical development. Typically, several functions are involved:

  • AI Engineering and Technical Architecture
  • Front End and User Experience
  • Integration of Internal Systems
  • Security and Data Protection
  • Testing and Quality Assurance
  • Product Ownership
  • Business Requirements
  • Project Management
  • Support and Training

Even a leanly operated internal platform can therefore incur the following annual costs:

  • AI Engineer or Platform Owner, 80 to 100 percent: CHF 130,000 to 170,000
  • Front-end and Integrations, 10 to 20 percent: CHF 15,000 to 30,000
  • Product Owner, Business and Project Management, 10 to 20 percent: CHF 15,000 to 30,000
  • Security, Data Protection, and Testing: CHF 10,000 to 20,000
  • Hosting, APIs, Monitoring, and Tools: CHF 15,000 to 40,000
  • Total: approximately CHF 185,000 to 290,000 per year

A budget of CHF 150,000 to 200,000 per year is therefore more realistic for a very lean organization, where existing employees also take on additional responsibilities.

For a business-critical platform with multiple integrations, high security requirements, and ongoing development, the costs can be significantly higher.

In contrast, the cost of an external business platform for 100 employees is often in the low to mid five-figure range per year, even if a smaller group receives additional premium licenses.

Therefore, the economically sound question is not simply “in-house or cloud.”

A custom solution makes sense especially when it provides a specific business advantage or meets requirements that existing platforms do not cover.

To ensure broad access to powerful AI, it is often more efficient to use an established business platform and develop only the specific business functions in-house.

The internal IT department can then focus on data integration, access rights, security, processes, and the applications that truly set the company apart.

What Good AI Governance Actually Requires

Governance should establish a secure and clear framework for use. This includes, in particular:

  • Approved Business Tools: Employees know which systems are officially authorized for use.
  • DPA and Contract Review: Data protection and data processing are governed by contract.
  • Data Classification: It is clearly defined which data may be used in which tool.
  • Identity and Rights Management: Users, roles, and access are managed centrally.
  • Rules for Sensitive Applications: Additional requirements apply to HR, finance, healthcare, and other sensitive areas.
  • Different license tiers: Casual users and power users receive access tailored to their needs.
  • AI Literacy: Employees are trained to use AI safely and effectively.
  • Ongoing Review: Providers, models, and features change. Governance must therefore be updated regularly.
  • Controlled Alternatives to Shadow AI: If additional tools are needed, there should be a regulated review and approval process.

In this way, governance does not become a collection of prohibitions, but rather a tool that makes controlled use possible in the first place.

Conclusion

Companies don't have to choose between powerful AI and data protection.

OpenAI, Microsoft, Google, Anthropic, and Swiss providers can generally be used for business applications. However, this requires selecting the correct product version and reviewing the DPA, data processing, training, storage, international transfers, permissions, and the specific use case.

For many companies, a tiered model is likely to make sense: a standard business tool for widespread use, more robust licenses or additional systems for power users, and targeted in-house developments where they create genuine technical or strategic value.

A fully in-house platform can be useful, but it incurs significant ongoing costs and requires a sustained commitment of technical and organizational resources. Given that the annual cost ranges from approximately CHF 185,000 to 290,000, it is therefore important to carefully assess whether a fully in-house solution is truly necessary or whether existing platforms already meet a large portion of the requirements.

At the same time, the official solution should be robust enough that employees will actually use it. A highly restrictive internal platform does not automatically reduce the use of shadow AI. If key features are missing, the risk that employees will turn to private or unauthorized alternatives actually increases.

The central task of AI governance is therefore to balance security, data protection, and practical usability.

In a nutshell:

  • Cloud AI is not generally prohibited under the nDSG and the GDPR.
  • Corporate data requires appropriate business solutions and a certified DPA.
  • OpenAI, Microsoft, Google, Anthropic, and Swiss providers must be evaluated based on the same fundamental criteria.
  • Not all employees need the same license or incur the same costs.
  • Having several approved tools may be more effective than a rigid single-tool strategy.
  • Developing your own solution can be a good idea, but it’s significantly more work to maintain than it often seems when you first implement it.
  • Good governance establishes clear guidelines and enables effective utilization within those guidelines.

As of September 22, 2026. The specific assessment under data protection law always depends on the particular use case, the data being processed, and the contracts entered into.