The Suniworks journal
Local AI for Swiss SMEs: choosing where data is processed
Ollama offers local and cloud execution, while Apertus is a Swiss-developed model. Separate the model, the hosting arrangement and the operating responsibilities before choosing how to handle company documents.
Updated on .

For a Swiss SME considering local AI, start by specifying where company documents may be processed. Ollama offers a way to run models locally, while Apertus is a choice of model. Assess the model, the machine running it and the application's connections separately.
Choose a practical task before discussing equipment: drafting an email, summarising a document or locating a contract clause. Define which information the task may use and who will check the answer. The wider question of personal ChatGPT at work also deserves attention when setting rules for staff.
What matters before choosing a solution
- Ollama supports both local and cloud execution. Installing it does not, by itself, specify where processing happens.
- Apertus comes from the Swiss AI Initiative, through collaboration between EPFL, ETH Zurich and CSCS. Assess its hosting separately from its Swiss origin.
- Compare a defined task, an agreed set of documents and the work needed to operate each option.
Separate the model from the service running it
Ollama provides an execution environment
A model is the component that generates an answer from the information it receives. The Ollama website describes support for running models on Windows, Linux and macOS, either locally or through cloud infrastructure. Cloud execution means processing on remote infrastructure.
Ask your technical provider to name the chosen model, identify the machine that will run it and explain how employees will access it. Put those answers in the proposal. A product name alone is too broad for an acceptance criterion such as keeping document processing inside the business.
Apertus is a Swiss-developed model
The Swiss AI Initiative's Apertus page identifies EPFL, ETH Zurich and CSCS as collaborators in its development. It describes openness across training data, code, model weights and methods. Weights are the values learned during training that shape how a model responds.
Use that information to assess how the model is developed and documented. Then request a separate description of the proposed deployment. For a requirement such as processing within Switzerland, make the location of that deployment an explicit condition rather than deriving it from the model's origin.
Compare hosting arrangements against the same requirements
Use a common checklist when considering ChatGPT or Copilot, Infomaniak's Euria, or a local deployment using Ollama. The table below sets out questions for the proposed arrangement; it does not assign hosting guarantees or commercial terms to those named services.
| Arrangement under consideration | Data handling to establish | Setup to define | Costs to request |
| Remote service with processing abroad | Processing countries, retention and parties involved | User application, accounts and document access | Selected subscription, integration and administration |
| Service required to provide Swiss hosting | Exact scope of the Swiss hosting commitment, including supporting functions | Access method, accounts and business-system connections | Contracted service, support and integration |
| Local execution using Ollama | Target machine and permitted external connections | Ollama, selected model and employee-facing application | Equipment, deployment and ongoing operation |
For a hypothetical fiduciary firm, the trial might involve summarising an authorised document. Specify the input, the intended recipient and the information the summary must retain. A joinery business could instead assess a draft reply to a quotation request. These are ways to scope a trial, not claims about the quality either task will achieve.
Plan hardware around a defined workload
Request a configuration for the selected model and task, then require a trial on that configuration. Describe the documents it must handle, the requested output and expected simultaneous use. Give the technical provider concrete conditions against which to justify the equipment proposal.
The Ollama Linux documentation provides installation routes for AMD64 and ARM64 architectures. It also covers components and drivers for AMD and NVIDIA graphics processors. A graphics processor is a computing component; ask the provider to explain its role in the proposed setup rather than choosing equipment from its name alone.
Assign responsibility for running Ollama
The Linux documentation covers running Ollama as a system service, viewing logs and updating the installation. A service is a program managed by the operating system. Logs record information that can help the operator investigate its behaviour.
Include those operating tasks in the proposal before staff begin using the application. Name the person responsible for checking errors, approving updates and validating results after a change. For a limited trial, agree how problems will be recorded and who can decide to pause it.
Check the connections around the local model
Review the functions enabled around Ollama as well as the machine running the model. Its Claude Code integration documentation describes local and cloud model connections, alongside web search and page-fetching functions. This is a development integration, but it illustrates why connection choices belong in an architecture review.
For your business application, request a diagram tracing an authorised document from input to answer. Have the provider identify each component and the communications it is permitted to make. If processing must remain inside the business, write that requirement into the project's acceptance criteria and ask how it will be checked.
Where the project also involves generated software, establish who reviews and approves it. The question of responsibility for AI-generated code belongs alongside the deployment discussion, rather than being left until handover.
Decide whether local operation is worth the effort
Set the conditions for keeping the installation before starting the trial. For document summaries, prepare a checklist of essential information and ask the reviewer to record omissions. For clause retrieval, require the relevant passage and its location so the reviewer has a clear basis for checking the answer.
Compare budgets over the same scope. Request separate estimates for setup, any equipment and recurring operating work. Include the review effort in your assessment, and agree how it will be recorded. Treat a time saving as something to measure during the trial rather than a promised outcome.
Use the decision criteria to choose what to test. If local processing is a firm requirement, assess a proposed local architecture against it. If your requirement concerns processing within Switzerland, request an explicit commitment covering the proposed service and its functions. Keep those requirements distinct when comparing offers.
Start with a bounded trial
- Select a task and prepare fictional examples or documents explicitly authorised for the trial.
- Write down your processing requirements and request a written description of the proposed architecture.
- Define acceptable answers and appoint the person who will validate them.
- Run the trial on the proposed configuration, recording errors, review work and operating interventions.
Use those observations to decide whether to continue, change the configuration or reconsider the hosting arrangement. Before authorising business documents, have the responsible person approve the document flow, access permissions and answer-review procedure.