AI glossary for Swiss SMEs

AI comes with a lot of jargon, and most of it is simpler than it sounds. Here we explain the terms you will hear most often, in plain language and with an example from everyday Swiss business.

AI (artificial intelligence): AI is software that can do tasks which used to need human judgement, such as understanding text, recognising images or answering questions. It does not think like a person, it recognises patterns learned from large amounts of data. Today most business use of AI is about language: reading, writing, sorting and summarising. In practice: a fiduciary uses AI to read incoming receipts and prepare the bookings for review.
AI agent: An AI agent is an AI system that does not just answer, it carries out several steps on its own to reach a goal. It can look things up, use tools such as your calendar or CRM, and decide what to do next. Good agents work within clear limits and ask a person before important actions. In practice: an agent reads a customer request, checks stock in the ERP, drafts a quote and puts it in your inbox for approval.
API: An API (application programming interface) is an agreed way for two software systems to talk to each other and exchange data automatically. Instead of a person copying data from one screen to another, one program asks the other through its API. Most modern business software, and every major AI model, offers an API. In practice: your webshop sends each new order through an API to your accounting software, without anyone typing it in.
Chatbot / AI assistant: A chatbot is a program that answers questions in a chat window, on your website, in an app or in internal tools. Older chatbots followed fixed scripts, while modern AI assistants understand free text and reply in natural language, often in several languages. A good assistant only answers from your approved information and hands over to a person when it is unsure. In practice: a garage's website assistant answers questions about opening hours and tyre changes at night and books the appointment straight into the calendar.
Context window: The context window is the amount of text a language model can take into account at once, including your question, any attached documents and its own answer. It is measured in tokens. Anything outside the window is simply not seen by the model, so very long documents may need to be split or searched first. In practice: a 200-page tender dossier may not fit in one request, so the system first picks out the relevant sections and only passes those to the model.
Embedding: An embedding turns a piece of text into a list of numbers that captures its meaning. Texts with similar meaning get similar numbers, even if they use different words or languages. This lets software search by meaning instead of exact keywords. In practice: a search for 'cancel my contract' also finds the document titled 'termination of subscription', and the French version too.
EU AI Act: The EU AI Act is the European Union's law on artificial intelligence. It entered into force in August 2024 and applies in stages, with rules that depend on the risk of each use: some practices are banned, high-risk uses such as certain hiring or credit decisions face strict requirements, and chatbots must make clear that people are talking to AI. It can also apply to Swiss companies that offer AI systems in the EU or whose AI output is used there. In practice: a Swiss company whose website assistant serves customers in Germany should state clearly that it is an AI assistant.
Fine-tuning: Fine-tuning means training an existing AI model further on your own examples so that it adopts a certain style, format or specialist task. It changes the model itself, which takes effort, good training data and repeating whenever things change. For most SMEs, giving the model the right documents at the moment of the question (RAG) is cheaper and easier to keep up to date. In practice: a company fine-tunes a small model on thousands of past classified support tickets so it sorts new ones exactly the way the team does.
GDPR: The GDPR (General Data Protection Regulation) is the European Union's data protection law, in force since May 2018. It can also apply to Swiss companies, for example when they offer goods or services to people in the EU or track their behaviour online. Its principles are close to the Swiss nFADP, but some duties differ and fines can reach 20 million euros or 4 percent of worldwide annual turnover. In practice: a Swiss online shop delivering to customers in Austria needs to check its AI tools against the GDPR as well as the nFADP.
Generative AI: Generative AI is AI that creates new content, such as text, images, audio or code, instead of only sorting or predicting. Tools like ChatGPT are the best-known examples. The results are often very good, but they are produced by probability, so facts and figures should be checked before use. In practice: a carpentry firm uses generative AI to draft quote cover letters in German and French, which the owner then reviews and sends.
Hallucination: A hallucination is when an AI model states something false with full confidence, for example an invented figure, a non-existent law article or a wrong opening time. It happens because language models produce the most likely-sounding text, not verified facts. The risk drops sharply when the model must answer only from your own documents and show its sources. In practice: a website assistant that may only answer from the company's price list and FAQ will say 'I do not know' instead of inventing a price.
Human in the loop: Human in the loop means a person checks and approves the AI's work at key points before anything important happens. The AI does the preparation, the human keeps the decision and the responsibility. This is the safest way to start with AI and is often expected for sensitive tasks in public administration. In practice: the AI drafts replies to citizen enquiries at the municipal office, and a staff member reads, adjusts and sends each one.
Integration: An integration connects two or more software systems so that data flows between them automatically. It removes double entry and copy-paste between tools such as your CRM, ERP, webshop, email and spreadsheets. Integrations are usually built on APIs, and they are often where AI saves the most time. In practice: when an enquiry arrives by email, the integration creates the contact in the CRM, attaches the documents and notifies the right person.
Intelligent document processing (IDP): Intelligent document processing combines OCR with AI to read documents, understand what they are and pull out the data you need. Unlike simple OCR, it knows that a number is an invoice total, a due date or a VAT amount, even when every supplier uses a different layout. Uncertain fields are flagged for a person to check. In practice: supplier invoices arriving as PDF or paper scans are read, checked and prepared for posting in the accounting software.
Large language model (LLM): A large language model is an AI model trained on huge amounts of text so that it can understand and write language. It is the engine behind tools like ChatGPT, Claude, Gemini or Mistral. An LLM is very good at reading, summarising, translating and drafting, but on its own it does not know your company's data. In practice: an LLM summarises a 30-page consultation document into one page for the municipal council, in German and French.
Machine learning: Machine learning is the part of AI where software learns rules from examples instead of being programmed rule by rule. You show it many past cases, and it learns to recognise patterns and make predictions for new ones. It is behind spam filters, demand forecasts and fraud detection, and it is also how language models are built. In practice: a bakery uses machine learning on past sales and weather data to estimate how many loaves to bake each morning.
nFADP (Swiss Federal Act on Data Protection): The nFADP is the revised Swiss Federal Act on Data Protection, in force since 1 September 2023. It applies to private companies and federal bodies that process personal data, while cantonal and communal offices fall under their cantonal data protection laws. It requires, among other things, transparency towards the people concerned, appropriate data security and care when personal data is disclosed abroad, and serious breaches must be reported to the Federal Data Protection and Information Commissioner (FDPIC). In practice: before sending customer emails to an AI service, an SME checks where the data is processed and states this in its privacy policy.
OCR: OCR (optical character recognition) turns images of text, such as scans, photos or image-only PDFs, into text a computer can search and edit. Classic OCR recognises the letters but does not understand what they mean. Modern OCR also handles handwriting and poor scans much better than a few years ago. In practice: decades of scanned building permits in a municipal archive become searchable by address or parcel number.
On-premise vs cloud hosting: On-premise means the software and AI model run on servers you own or control, in your office or your own data centre. Cloud hosting means they run in a provider's data centre and you use them over the internet. The cloud is quicker to start and needs no hardware, while on-premise gives the most control over where data goes but costs more to set up and maintain. A middle path is a cloud hosted in Switzerland under a clear contract. In practice: a cantonal office keeps sensitive case files on its own servers with an open-weight model, while a small shop uses a Swiss cloud service for its website assistant.
Open-weight model: An open-weight model is an AI model whose trained parameters, the so-called weights, are published so anyone can download and run it on their own hardware. Examples include models from Mistral, Meta (Llama) and Qwen. They are not always fully open source, since the training data and licence terms vary. Their main advantage is that data never has to leave your own infrastructure. In practice: a law firm runs an open-weight model on a server in its office so that client files are never sent to an outside provider.
Prompt: A prompt is the instruction or question you give an AI model. The clearer the prompt, with context, the desired format and examples, the better and more consistent the answer. In business tools, most prompts are written once by the developer and run in the background, so staff do not need to write them each time. In practice: the prompt behind an email assistant says which tone to use, which languages to reply in and when to forward a message to a person.
RAG (retrieval-augmented generation): RAG is a method where the AI first searches your own documents for the relevant passages and then writes its answer based on them. This way the model can answer questions about your company without being retrained, and it can show which sources it used. When a document changes, the answers change too as soon as the new version is indexed. In practice: staff ask 'How many days of holiday do apprentices get?' and the assistant answers from the current staff regulations, with a link to the exact page.
RPA (robotic process automation): RPA is software that imitates a person clicking through screens: it opens programs, copies values, fills in forms and presses buttons, following fixed rules. It is useful for old systems that have no API. Because it depends on the exact screen layout, it can break when a program is updated, and it cannot handle cases it was not told about. Combined with AI, it can also deal with less structured input. In practice: an RPA bot transfers approved holiday requests every evening from an email form into an old HR system that has no interface.
Token: A token is the small unit of text a language model reads and writes, usually a word or part of a word. In English, 100 tokens are roughly 75 words, and German, French and Italian often need somewhat more. Tokens matter because AI providers charge per token and the context window is measured in tokens. In practice: summarising a 10-page contract costs only a few cents in tokens, while processing thousands of documents a month adds up and is worth estimating in advance.
Vector database: A vector database stores embeddings, the number lists that capture the meaning of texts, and finds the ones closest to a question very quickly. It is the search engine inside most RAG systems. Your documents stay in their original place, the vector database holds the searchable index of their content. In practice: all 2,000 pages of a company's manuals and procedures are indexed in a vector database, so the assistant finds the right paragraph in under a second.
Workflow automation: Workflow automation means letting software carry out a recurring sequence of steps that people used to do by hand, such as receiving, checking, forwarding and filing. Classic automation follows fixed rules, while AI steps can handle the parts that need reading or judgement. Each step is logged, so you can always see what happened and why. In practice: every incoming job application is acknowledged, summarised, stored in the right folder and sent to the hiring manager, without anyone touching it.