Artificial intelligence (AI) for SMEs is used first and foremost for repetitive tasks: invoices, cash management, customer service, writing. In Switzerland, the specialists quoted by the press in October 2026 recommend the same method for getting started: a specific use case, data kept under control and human validation of sensitive decisions.
These tools are changing in nature. Originally designed as assistants, they are becoming agents capable of acting in several software applications at once. This guide complements 2clics’ feature on AI in business.
What can AI actually do in an SME?
An AI assistant responds to a one-off request. In everyday use, it is asked to write a text, summarise information or prepare an analysis, notes the specialist website Inside IT on 26 October 2026.
An agent goes further. It analyses a situation, looks for the relevant information, proposes a solution, then carries out a series of actions. SAP promotes this idea under the name “Autonomous Enterprise”. “Humans steer the process,” sums up Alexander Finger, chief technology officer of SAP Switzerland, in PME magazine. AI can then “execute and orchestrate”.
| Task | What AI does | What remains human |
|---|---|---|
| Supplier invoices | Extracts the data, spots duplicates, unusual amounts and changes to bank details | Check the flagged cases and approve the payment |
| Cash flow forecasting | Analyses past flows, open invoices and payment habits, then proposes scenarios | Choose between preserving liquidity, adjusting spending or seeking financing |
| Customer service | Checks whether an item is available and could, eventually, take payment, place orders or recommend other products | Set the rules, particularly on prices and discounts |
| Sales proposal or report | Writes the document | Decide under what conditions the agent may change a price or commit to an expense |
The invoice and cash management examples come from Alexandre Buga, in an analysis published by PME on 23 October 2026 on the chief financial officer as seen by Deloitte.
Which tasks should be entrusted to AI, and which should be kept?
The potential depends on the process more than on the sector, according to Cyril Louis, managing director of Mavericx, an IT consulting firm based in Geneva and Zurich. Automation pays off most where many interactions, large volumes of data and repeated decisions come together.
The goal is not necessarily to cut jobs. According to Cyril Louis, companies mainly want to automate repetitive tasks and use the time saved elsewhere. Customer service could thus shift from a mere cost centre to an additional sales channel.
Some decisions must remain beyond the machine’s reach. Alexander Finger cites the case of a chatbot that had granted discounts on cars, which were then sold at a reduced price. SAP says it has ruled out facial recognition and systems that make decisions about people.
Human resources illustrate this boundary. AI can be used there to search for and synthesise information. According to Alexander Finger, it must not be used to judge employees. For an SME, the dividing line remains simple: the agent prepares, the human decides anything that commits the company or a person.
Why does data matter more than the AI model?
An agent is only as good as the information it can access. For Marc Beierschoder, partner in charge of AI and data at Deloitte, competition is shifting from the model to the company’s context. Customer data, contracts, documents and business rules are often spread across separate systems. Making them technically accessible is not enough.
Alexander Finger puts it his own way: “You can change LLMs, but not the connections to SAP’s systems.” The large language model (LLM) is therefore only one part of the set-up. Controlling the accessible data, the permitted actions and the rules matters at least as much as the choice of model.
This framework must be clear, especially as agents multiply. Marc Beierschoder calls for defining which data an agent can see, which business rules it follows and which actions it carries out on its own. To connect an agent to other sources, Cyril Louis cites the Model Context Protocol (MCP), which allows applications and AI models to exchange context.
Data quality then becomes central. If the information is incomplete, contradictory or poorly contextualised, an agent’s autonomy only speeds up errors, PME magazine points out.
Where does a Swiss SME’s data go?
Data residency, meaning the place where data is stored and processed, becomes more complicated with agents. According to Cyril Louis, these systems can depend on several language models and external services. Not all of them are operated in Switzerland. The issue is sensitive for finance, healthcare and major Swiss brands.
Alexandre Buga points out that an SME must comply with Swiss data protection legislation, including when choosing a tool. Before passing on data about customers, suppliers or employees, it must know:
- where this data is processed;
- for what purposes it may be reused;
- who can access it.
He recommends suitable professional tools and simple rules on which data is allowed. Swiss providers are meeting this demand. The Elca group offers a sovereign AI platform, where data and models remain under the customer’s control. In October 2026, it took a majority stake in Beyond Civic, a Lucerne-based specialist in sovereign data exchange.
What human oversight should be planned?
The more leeway an agent has, the more costly an error can be. The advice given to SMEs in PME magazine is therefore to proceed step by step. First, the agent is given a simple task, then it is allowed to chain several operations together. Human validation remains the rule for sensitive decisions.
An internal directive is not enough. According to the swissVR Monitor cited by Alexandre Buga, many companies report having a governance framework for generative AI. But fewer than one in five have all content validated by a person.
Monitoring matters too. Cyril Louis speaks of “observability”: knowing what agents are doing, for which tasks, and what resources they consume.
Human review itself has a cost. A Harvard-affiliated study measured 20 to 30% more activity among developers who use AI, with no significant increase in completed software. Human review becomes the bottleneck: it rises on average from 7.03 to 10.48 days.
How much does AI cost for an SME?
Investing in AI does not necessarily mean savings, says Alexandre Buga. The financial return is not immediate, and the tool must first create value for the company.
Usage costs matter too. A solution can work while needlessly consuming a lot of resources, warns Cyril Louis. The architecture must therefore use the available resources wisely.
OpenAI, in a report on Europe, highlights more affordable pricing, better accessibility of tools and trained teams. In it, the provider claims that European SMEs are speeding up on AI agents. Its figures concern its own products and say nothing about Switzerland.
AI for SMEs: how to take the first steps?
- Choose a use case with a clearly identifiable benefit, rather than aiming straight away for an autonomous enterprise.
- Appoint a person responsible for the project.
- Set the costs, expected results and measurement criteria: time saved, errors avoided, better decisions.
- Decide which tasks are automated, which exceptions are checked and who is accountable for them.
- Know where the data goes and choose a suitable professional tool.
- Evaluate the tool after a defined period, then adapt it or stop it.
What Swiss rules apply to AI in business?
The legal framework is under construction. According to the SME portal of the State Secretariat for Economic Affairs (SECO), Switzerland has signed the Council of Europe Framework Convention on AI. It has committed to adapting Swiss law accordingly. A preliminary draft regulation is to be drawn up by early 2027.
The portal sets out its aims: strengthening Switzerland as a location for innovation, safeguarding fundamental rights and increasing public trust in AI. The concrete effects for SMEs are not yet known.
Key figures
- Nearly nine out of ten chief financial officers believe that human work will remain predominant in their function in three years’ time (Deloitte survey).
- Fewer than one company in five has all generative AI content validated by a person (swissVR Monitor).
- Agents produce 55% of the tokens generated by OpenAI’s Enterprise and Business customers in the European Union.
- Weekly users of Codex and ChatGPT Work in European SMEs have grown more than eightfold in less than six months, according to OpenAI.
- With AI, code review rises on average from 7.03 to 10.48 days (Harvard-affiliated study).
What we don’t know yet
- The share of Swiss SMEs that use AI, and the share that already entrust tasks to agents.
- The real cost of a first project and the measured gain in a Swiss SME.
- The content of the preliminary draft law and any obligations it may impose on companies.
- Switzerland’s position in the adoption of AI agents in Europe.
Sources
- PME magazine, autonomous enterprise: “Entreprise autonome: jusqu’où laisser l’IA décider?”
- Inside IT, enterprise AI: “Bei Enterprise-KI entscheidet nicht Modell, sondern Kontext”
- SECO SME portal: home page
