Enterprise AI in Switzerland: A Practical Guide for Companies

26/08/2026

Swiss companies are using AI to check invoices, forecast demand, search internal documents, answer customer questions and analyse production data. The technology becomes harder to manage once companies connect it to internal systems, customer records, financial data or factory equipment. This guide explains which business tasks suit AI, what companies should prepare before they introduce it, how to control access and data, what Swiss law requires today and how to judge whether an AI project is worth the cost.

What you will learn

which company tasks suit AI;
how to choose a first project;
what companies should fix in their data before using AI;
how AI can search internal documents and business systems;
how companies should control AI that can update records or trigger software;
how finance teams and manufacturers already use AI;
how to choose between different AI models;
what companies should check before sending confidential information to an AI provider;
how Swiss law and the EU AI Act can affect Swiss businesses;
what companies pay for when AI becomes part of daily work;
how to measure whether AI saves time or money.

What is enterprise AI?
Which company tasks suit AI?
Choosing the first AI project
Preparing company data
Using AI to search internal documents
Giving AI access to company software
AI in finance
AI in manufacturing
AI in HR
Choosing an AI model
Where company data is processed
What enterprise AI costs
Measuring whether AI improves the work
AI security and system permissions
AI regulation in Switzerland
When the EU AI Act applies to Swiss companies
Deciding where employees still need to make the decision
Testing AI before everyday use
Enterprise AI implementation checklist
What Swiss companies should prepare for next
Further reading on Supralogic

What is enterprise AI?

Enterprise AI means using artificial intelligence for company work. The term covers very different jobs. A finance team can compare invoices with purchase orders. A manufacturer can analyse machine readings and warn maintenance staff about unusual changes. An internal assistant can search company policies or technical manuals. Customer-service software can read a request, check account information and prepare a reply.

Companies have used machine learning for forecasting and pattern recognition for years. Generative AI made it easier to work with text, images and documents because employees can give instructions in ordinary language.

Companies are now connecting these models to ERP systems, CRM platforms, document libraries and production databases. A model that only drafts text has limited access. A model connected to internal software can read company information and, depending on its permissions, update records or start another task.

The company therefore needs to decide exactly what the system can read, what it can change and when an employee must approve the next step.

Which company tasks suit AI?

Customer-service teams often spend several minutes on every case searching customer records, product information and company policies. AI can collect some of that information before the employee answers.

Finance employees compare invoices with purchase orders, investigate unusual payments and prepare forecasts from several data sources.

Procurement teams can search supplier contracts and extract information from documents. Legal teams can find clauses in large collections of agreements. HR teams can use AI for administrative work such as preparing job advertisements, organising applications or answering common employee questions.

Manufacturers can analyse machine readings, inspect products with cameras or use production and inventory data when planning output.

Companies should define the job before they choose the technology. “Use AI in finance” gives a project team very little to test. “Reduce the time employees spend checking invoices against purchase orders” gives the team a specific task and a result it can measure.

AI is not necessary for every repetitive job. If a company can describe exactly what software should do in every case, ordinary automation may cost less and be easier to control.

A payment above CHF 50,000 can trigger a second approval through a fixed rule. AI becomes more appropriate when software needs to read different document formats, understand written language or compare current activity with historical data.

Choosing the first AI project

A company should choose work it already understands and can measure. Consider a business that receives 5,000 supplier invoices each month. Employees enter information, compare each invoice with a purchase order and investigate anything that does not match.

The finance team can record how many hours the work takes, how many invoices require manual review and how often employees find errors. After introducing AI for part of the job, the company can compare those figures again.

Customer-service managers can compare resolution times and repeat contacts. Maintenance teams can record downtime and repair costs. Planning teams can compare forecasts with actual sales. Legal departments can record how many hours lawyers spend searching existing contracts. These numbers tell management more than the number of employees who have opened an AI application.

Preparing company data

Many businesses already hold the information required for an AI project, but their records may be inconsistent.

The same supplier may appear under several names. Old customer records may still be active. Two factories may record the same machine reading differently. Employees may have saved several versions of the same policy without clearly identifying the current one.

Experienced employees often correct these inconsistencies themselves. Someone in accounts payable recognises that two slightly different supplier names belong to the same business. An engineer knows that older maintenance records use a previous machine code.

A model does not automatically know any of this. If supplier records contain duplicates, the model may treat normal payments as activity involving new suppliers. If an internal assistant searches several versions of a policy, it may quote one that employees should no longer follow.

The company should identify which records employees and software should trust, archive outdated material and check who can access each source.

Manufacturers face an additional problem because older production equipment often records data in different formats. A factory may therefore have years of machine data without having consistent data that engineers can use directly for AI.

Using AI to search internal documents

Generative AI can help employees search large collections of company material.

An employee might ask:

What expenses can I claim during a business trip to Germany?

Which warranty applies to this machine?

Which contracts with this supplier expire this year?

What payment terms did we agree with this customer?

The application searches approved documents and passes the most closely matching sections to the language model. The model prepares its answer from those documents. Developers commonly call this retrieval-augmented generation, or RAG.

The search needs to find the correct source. If the system retrieves an old travel policy, the model may give the employee outdated information. If document permissions are too broad, the AI may show someone confidential material that the employee could not normally open.

Companies should show the source when employees use AI for contracts, policies, warranties or other work where the precise wording can affect a decision. The employee can then open the original document and check it.

Giving AI access to company software

Companies need stronger controls when AI can change a record or trigger another application.

A customer-service assistant may prepare a reply without changing anything. Once the company connects it to the service platform, the system could also update the case, change account information or issue a refund.

A procurement application may prepare a purchase order. Wider permissions could allow it to create the order directly in the ERP system.

A finance application may flag an unusual payment. The company must decide separately whether the software can stop the transaction or only ask an employee to investigate it.

Companies already restrict employee and software access according to job responsibilities. AI systems need the same limits.

A customer-service system does not need payroll records. An invoice-processing system does not need unrestricted access to every financial database. Software that prepares a purchase order does not automatically need permission to approve payment.

Companies should also record important actions. If an AI system updates a customer record, blocks a transaction or creates an order, administrators should be able to see what it did and which instruction led to the action.

AI in finance

Finance gives companies several AI tasks that they can measure clearly.

Forecasting is one example as finance teams already use historical sales, orders, costs and assumptions from different parts of the business. Machine-learning models can analyse more variables and update forecasts as new figures arrive.

Finance employees still need to change their assumptions when the company changes. Historical sales may become less reliable after an acquisition, a major price increase or the loss of a large customer.

AI can also help teams review transactions. Normal financial rules identify clear breaches. A system can detect a duplicate invoice or a payment above an employee’s approval limit.

Machine learning can compare a transaction with previous activity involving the same supplier. Suppose a supplier normally sends invoices between CHF 5,000 and CHF 10,000 and has used the same bank account for several years. A new invoice for CHF 45,000 with different bank details deserves a manual check even if the payment breaks no formal rule.

The software can tell the employee which details changed. The employee can then verify the new bank account with the supplier before releasing the payment.

AI in manufacturing

Manufacturers have used machine learning in production for many years. Predictive maintenance uses machine data to identify signs that equipment may be developing a fault. Machines can record vibration, temperature, pressure, power consumption and other measurements. A model can compare current readings with data collected before earlier failures.

Maintenance staff can inspect the equipment when those readings resemble a previous fault.

Manufacturers also use computer vision to inspect products. Production planners can combine order, inventory and historical demand data when deciding what to manufacture. Engineers can use generative AI to search technical manuals and maintenance records.

Factories also have technical requirements that ordinary office applications rarely face. A production line may need a response in milliseconds. Some sites cannot depend on a permanent cloud connection. A manufacturer may also decide that certain production data should remain inside the factory.

Companies can therefore run some AI models on servers or devices located close to the machinery. Older equipment can complicate the work. A plant may contain machines installed over several decades by different manufacturers. Engineers may need to standardise the data before they can use it reliably for AI.

AI in HR

AI can reduce administrative work in recruitment. HR teams can use it to prepare job advertisements, organise applications, schedule interviews and summarise interview notes.

Large employers receiving thousands of applications can save employees many hours on these jobs. Companies should apply more scrutiny when AI influences who receives an interview or job offer.

A model trained on previous hiring decisions can reproduce patterns contained in those decisions. Removing a person’s name or gender does not guarantee that the model will ignore related information because other variables can correlate with them.

Employers also need to comply with data-protection and employment rules when they process candidate or employee information.

A company can therefore treat interview scheduling differently from candidate assessment. Software can find a meeting time without deciding whether the applicant should continue in the recruitment process.

Choosing an AI model

Companies do not need the largest available model for every job. Large general-purpose models can work with text, images, documents and code. They can follow relatively complex instructions, which makes them convenient during early testing. A smaller model may be enough once the company knows exactly what the system needs to do.

A business extracting invoice numbers from millions of documents has different requirements from a legal team analysing long contracts. Classification and extraction can often run on smaller models at a lower cost.

Companies can also use several models. One may handle document analysis, another may process images and a smaller model may classify routine customer requests.

Before sending confidential information to a provider, the company should check where the provider processes the data, whether it keeps prompts, whether it uses customer data to train models and which contractual protections apply.

The company should choose the model according to the job, required accuracy, operating cost and the way the provider handles its information.

Where is company data processed?

AI services can receive contracts, customer records, financial information, source code and other confidential material. Companies should know where providers process that information and what happens after the request.

They should check:

where the provider processes the data;
whether it stores prompts and outputs;
whether it uses customer data for training;
which staff or subcontractors can access the information;
how it encrypts the data;
whether the customer can export its information when it changes provider.

Keeping data in Switzerland can help with some requirements, but the location of the server does not describe every dependency. A service hosted in Switzerland may still use an international cloud provider, an external AI model, identity software or other services operated abroad. Companies should identify the providers involved and decide which dependencies they can accept.

What does enterprise AI cost?

Companies can test AI without committing a large initial budget. They may begin with several software licences or a small application connected to a model through an API.

Regular use adds other costs because developers may need to connect the AI to existing software. Teams may need to clean data. Security specialists must configure access. Employees need training. Someone must monitor errors and update the system when data, business conditions or model versions change.

Generative AI also creates ongoing computing costs. Providers commonly charge according to how much information the model processes. A few employees summarising documents occasionally may cost very little. An application processing hundreds of thousands of customer messages can generate a much larger bill.

The model itself affects the cost. The reason is that larger models generally cost more per request, while smaller models may perform routine work well enough.

Companies should calculate what the system costs for the job it performs. If AI processes invoices, calculate cost per invoice. If it handles customer requests, calculate cost per case. If it analyses contracts, calculate cost per document. Management can then compare the AI system with the current way employees perform the work.

Measuring whether AI improves the work

Adoption figures tell management very little about financial return. Hundreds of employees may use an AI assistant without reducing costs or serving customers faster.

Companies should measure the work they changed. A customer-service team can compare resolution times, repeat contacts and errors.

A finance team can compare forecast accuracy. Accounts payable can record processing time, mistakes and the number of invoices that still need manual review.

A manufacturer can compare equipment downtime and maintenance costs. Moreover, managers also need to know what happens to the time employees save.

Suppose an employee previously spent ten hours each week checking documents and now spends five. The company has reduced the time required for that job, but it has not automatically removed five hours of payroll cost.

The financial benefit depends on whether the company can process more work without hiring another employee, reduce spending on outside support, prevent losses or move staff to work that requires their experience.

AI security and system permissions

An AI application should only receive the information and software access required for its job. A tool that answers questions from technical manuals does not need employee salary records. A customer-service application does not need unrestricted access to the company’s financial database.

Development teams sometimes give prototypes broad access because it makes testing easier. Security teams should review those permissions before employees use the system for normal work.

Companies should also set separate permissions for individual actions. An application may prepare a purchase order without receiving permission to release a payment. A customer-service system may update routine account information while refunds above a fixed amount still require an employee.

Employees also need clear rules for public AI services. A person can easily paste a customer contract, spreadsheet or confidential email into a public chatbot because the tool saves time. The company should tell employees which services they may use and what information they may submit.

A specific policy gives employees clearer instructions than telling them simply to “use AI responsibly”.

AI regulation in Switzerland

Switzerland does not currently have one overarching law that regulates AI. Existing Swiss law still applies when companies use it. The Federal Act on Data Protection covers relevant processing of personal data, while employment law, intellectual-property law and sector-specific rules can also apply according to the task and industry.

The Federal Council decided in February 2025 to prepare targeted AI legislation rather than introduce a direct Swiss equivalent of the EU AI Act. Federal authorities are preparing a consultation draft for the end of 2026.

Companies do not need to wait for new legislation before keeping proper records of their AI systems. They should know which application they use, what data it receives, which provider processes the information and who inside the company is responsible for it.

An HR system that processes candidate information needs a different legal and security review from a marketing tool that rewrites public website copy.

When does the EU AI Act apply to Swiss companies?

A company headquartered in Switzerland can still fall within parts of the EU AI Act. The answer depends on the company’s role, where it provides the system and how the AI or its output is used in the European Union.

Swiss companies should therefore assess individual systems connected with EU activities rather than assuming that a Swiss headquarters automatically places them outside the legislation.

Several parts of the Act already apply. Requirements for general-purpose AI models began applying in August 2025, while further provisions, including transparency rules under Article 50, became applicable on 2 August 2026.

Those transparency rules cover specified AI systems that interact directly with people and certain forms of AI-generated or manipulated content. A Swiss manufacturer using an internal forecasting model only in Switzerland therefore faces a different legal assessment from a Swiss technology company selling an AI service to customers in the EU.

Deciding where employees still need to make the decision

Companies should also decide how much human review each AI system needs according to the consequences of an error.

An employee using AI to rewrite an internal email can read the draft before sending it. A payment flagged as possible fraud requires a proper investigation. The employee should see which transaction details caused the alert and should have access to the supplier records needed to check them.

Recruitment, credit, safety, regulated reporting and high-value financial decisions may also require an employee to examine the evidence before the company acts. A human approval step only works when the reviewer has enough time and information to make an independent decision. If an employee must approve hundreds of AI recommendations every day, the review can become routine clicking.

The leadership should therefore name the person or role responsible for the final decision and specify what the software can do without approval.

Testing AI before everyday use

A pilot can show whether a model performs a task under test conditions. Everyday work introduces more variation.

Employees upload unexpected documents. Customers ask questions the test team did not include. Company data changes. Another software provider updates its system. The AI provider releases a new model version.

The company should decide who investigates incorrect results, who can disable the application, how developers test updates and which logs the company keeps.

For work with larger financial, legal or operational consequences, companies can initially run AI alongside the current system.

A payment-monitoring model can flag transactions without blocking them. A forecasting model can produce a forecast while finance continues using its existing method. Maintenance software can warn engineers without automatically stopping equipment.

Employees can compare the AI results with actual decisions and record the types of errors that occur.

The company can automate more of the work after the system performs consistently and employees understand where it still makes mistakes.

Enterprise AI implementation checklist

1. Choose one clearly defined job

Select work that takes a measurable amount of time, produces frequent mistakes or requires employees to search large amounts of information.

2. Record current performance

Measure processing time, cost, errors or another relevant figure before introducing AI.

3. Check the data

Confirm that the records are accurate enough for the job and that the company can use them legally and securely.

4. Test real cases

Include incomplete documents, unusual requests and examples where experienced employees need to use judgement.

5. Restrict access

Give the application only the information and software permissions required for the job.

6. Ask experienced employees to review the results

Employees who already perform the work can identify mistakes that may look reasonable to developers or managers outside the department.

7. Compare the new process with the old one

Measure accuracy, employee time, operating cost and the amount of manual work that remains.

8. Automate selected steps

Allow the software to complete work independently only where the company accepts the consequences of an error.

9. Keep measuring after launch

Company data, business conditions and model versions change. Teams should continue checking performance after employees begin using the system.

What Swiss companies should prepare for next

Businesses are connecting AI to more internal data and software. Employees will continue using tools that write, summarise and search documents, while more applications will also update records, check transactions and complete parts of routine work.

Companies will probably use several models rather than one model for every job. Large models can handle demanding document analysis or complicated instructions, while smaller models can process routine classification and extraction at a lower cost.

Manufacturers will continue running some models close to production equipment when they need faster responses or want to keep production data inside the plant.

AI systems that can change records or trigger other software require stricter permissions than tools that only produce text. Companies still need accurate records, named responsibilities, controlled access and measurements that show whether the new system improved the job. A company that understands an existing process can test whether AI can remove specific manual steps. If its records are inconsistent, responsibilities are unclear or too many systems have unrestricted access to company data, management should correct those problems before giving AI more authority.

Swiss companies do not need to begin with a company-wide AI programme. They can choose one expensive, slow or error-prone job, measure how employees perform it today and test whether AI produces a better result.