AI and websites in plain words.

LLM stands for large language model. It predicts word by word what fits next and in this way produces answers, summaries, translations or drafts. It does not understand like a person does, but it is very good at recognising patterns in language. Chat assistants such as ChatGPT or Claude are built on these models.

The clearer the prompt, the more useful the result. A good prompt says what the result is for, who it is for, what tone to use and how long it should be. A short example of the result you want often helps more than a long explanation.

This includes your question, the conversation so far and any attached documents. What is not in the context, the model does not know about your business: not your prices and not your customers. The amount is limited. This limit is called the context window, and with very long documents details can get lost.

A token is a word or part of a word. German, French and Italian texts usually need more tokens than English ones, because longer words are split into several pieces. Providers bill use of their interface by tokens, separately for input and output, and the size of the context is also measured in tokens.

Language models write fluently even when they lack a basis for what they say. The result can be wrong figures, invented sources or legal articles that actually cover something else. The risk drops considerably when the AI works with your own documents and names its sources. Important statements are still checked by a person.

Multimodal models describe photos, read scanned invoices or understand spoken language. Some also produce images or speech themselves. This makes it possible to automate tasks where someone used to type things out by hand.

With closed models, for example from OpenAI, Anthropic or Google, every request goes through the provider's servers. With open models, the model files can be downloaded and run on your own hardware or with a hosting company of your choice. Closed models are often more capable and easier to use, open models give more control over where the data stays.

An agent can search, read files, fill in forms or prepare emails, deciding by itself which step comes next. That helps with tasks made of many small steps, but it is harder to control than a single answer. Every agent therefore needs clear limits: what it may do, what it may not do and where a person has to approve.

Classic automation always does the same thing, for example filing every incoming invoice in the right folder. It is cheap and reliable. AI is only needed where the input looks different every time, such as freely written emails or photos. Often a mix works best: fixed rules for the process and AI only for the step that needs understanding.

Before automating anything, it pays to write the process down: what comes in, who does what, which programs are involved and what comes out at the end. This usually shows quickly which step takes the most time. That step gets improved first, not everything at once.

API stands for application programming interface. Through an API, one program can ask another for something or hand over data, without anyone clicking or copying. Many accounting, shop and CRM programs offer such an interface. How good it is often decides how easily a program can be connected.

With an integration, the AI reads directly from the mailbox, the file storage or the ERP and writes results back there. Without it, work stays at copying between windows and the benefit is smaller. A good integration only accesses what it needs for its task.

For accounting, payroll or email there are proven standard programs. A custom solution pays off where your process is special and no product fits, often as a small app or web application connected to your existing software. Custom software also needs care, so maintenance and further development belong in the budget from the start.

RAG stands for retrieval-augmented generation. For each question, the relevant passages are found in manuals, contracts or internal pages and passed to the model. The AI then answers with your knowledge instead of general knowledge and can name the source. The quality depends heavily on how current and well organised the documents are.

The AI prepares, a person decides. This makes sense for anything that goes out of the company, costs money or has legal consequences, such as quotes, payments, terminations or replies to customers. Over time it becomes clear where the suggestions are reliable and where checking remains necessary.

Chat assistants are usually billed per person and month. Automations through an interface are billed per use instead, measured in tokens. The cost per request is often small but adds up with many requests, long documents or expensive models. A spending limit with the provider prevents surprises.

The revised Federal Act on Data Protection has applied since 1 September 2023. Among other things, it requires you to say openly which personal data you process and why, to use only as much data as necessary and to protect it appropriately. If you pass data to a service provider, such as an AI provider, you remain responsible for it. With customers in the EU, the GDPR may apply as well.

When an external provider processes personal data for you, such as a cloud, email or AI service, there needs to be a contractual basis. It states what the data is used for, how it is protected and which other companies are involved. Large providers offer such agreements as standard, often called a DPA.

This includes name, address, phone number and email, but also a customer number or a photo if someone can be recognised from it. Health data, religious views or information about criminal proceedings count as sensitive and require extra care. Before you process customer data with AI, it is worth asking which details are really needed.

Many AI and cloud services run in data centres in the United States, some offer locations in the EU or in Switzerland. Switzerland recognises EU countries as having adequate data protection. For the US, this has applied since September 2024 only to companies that have joined the Swiss-US Data Privacy Framework, otherwise additional contracts are needed. Besides location, access and retention period also matter.

With some free and personal versions of chat assistants, input can be used for training, depending on the settings. Business versions and use through the interface usually exclude this by contract. It is also worth checking how long input is stored, for example for abuse monitoring.

An AI solution should only access what it needs for its task and should not see more than the person using it. The same applies to staff: not everyone needs access to payroll data or all customer files. Granting rights sparingly and removing them as soon as someone leaves prevents many problems.

If a password is stolen, nobody gets into the account without the second factor. It is one of the most effective and cheapest protective measures there is. It belongs on every account that handles money, customer data or your website: email, accounting, hosting and domain.

A domain is rented yearly from a registrar. It matters that it is registered in your company's name and that you have access to the account yourself, not just the agency or a former employee. If the domain expires, the website and often the email addresses stop working.

The DNS settings say which server the website is on and which service receives your email. A wrong entry can take down the website or email. Changes therefore belong in experienced hands and should be documented.

This can be a classic web hosting package, a website builder with hosting included or a platform that delivers pages quickly worldwide. What matters are availability, security updates, backups and where the data is stored, especially when forms with personal data come in through the website.

With HTTPS, nobody along the way can read what visitors type into a form. Browsers warn clearly about pages without encryption, and search engines prefer encrypted pages. For this, the website needs a valid digital ID, which most hosting providers now issue for free and renew automatically.

CMS stands for content management system. With it, you update opening hours, offers or news yourself, without calling the agency every time. Well-known examples are WordPress, Webflow or website builders such as Wix. Every CMS needs updates, and the more extensions are installed, the more upkeep is required.

SEO stands for search engine optimisation. It includes texts that use your customers' words, a clean technical foundation, short loading times and links from other websites. For local businesses, what matters most is showing up for searches like “electrician Winterthur”. SEO works over months, and no particular position in search results can be guaranteed.

The profile shows address, opening hours, phone number, photos and reviews. For local businesses it is often the first contact, even before the website. It is worth managing the profile yourself, keeping it up to date and replying to reviews in a factual and friendly way.

This includes readable contrast, descriptions for images, keyboard navigation and clear labels in forms. Everyone benefits, including older people or someone reading on a phone in the sun. In Switzerland, accessibility is mandatory for public authorities. If you sell into the EU, for example through an online shop, check whether the EU accessibility rules apply to you.

Many visitors come via their phone and mobile data. If a page takes several seconds, some leave before they have seen anything. The most common causes are oversized images, lots of embedded add-ons and slow hosting. Google measures loading with the so-called Core Web Vitals and takes them into account in search.

ChatGPT comes from OpenAI, Claude from Anthropic, Gemini from Google, and Microsoft Copilot is built into Windows and Microsoft 365. All of them write, summarise, translate, analyse files and search the web. The differences lie in the details, in how they fit into existing software and in the data protection terms of each version.

Llama comes from Meta, Mistral from a French company, and with Apertus, ETH Zurich and EPFL have released an open Swiss model. Such models run on your own servers, with a hosting provider in Switzerland or, in small versions, even on a powerful laptop. They require technical know-how, and the licences differ, especially for commercial use.

Speech recognition turns voice messages or meeting recordings into text. Speech output, for example from ElevenLabs, reads texts aloud with realistic voices and can also power phone assistants. Many services now understand Swiss German, but not always reliably. Anyone recording conversations needs the consent of the people involved.

Midjourney and FLUX mainly create images, Runway and Higgsfield mainly videos. The results work for drafts, mood boards and social media posts. For real products, staff and references, real photos usually come across as more credible. Usage rights need to be clarified, and so does consent when people can be recognised.

Zapier and Make run as cloud services and are quick to set up. n8n can also be run by yourself, for example on a server in Switzerland. All three can include AI steps, for example summarising an email and forwarding it to the right person. As the number of processes grows, you need an overview: who built what, and what happens when a step fails?

Amazon Web Services, Microsoft Azure and Google Cloud are the largest providers. All three operate data centres in Switzerland and also offer AI models there, although not every model at every location. Billing is based on usage, which is flexible but needs cost control. There are also Swiss cloud providers whose data centres are all in Switzerland.