With AI in every other headline these days, it helps to take a moment and look at what it actually is - beyond the marketing and beyond the doomsaying. Here, the essence in plain language, with no jargon and no hype.
AI today is almost always the same thing: a language model trained on enormous amounts of text, which predicts which word is likely to come next. Technically, it's no more than that. In its effect, it's no less.
This page isn't a tech lecture. It's a short tour through the terms you'll be hearing everywhere in the coming time - so you can keep up without feeling small, and so you notice when someone is using those terms to impress rather than to explain.
Training
The model reads billions of texts and learns which words typically follow which.
Prediction
When you ask, it works out word by word what would most likely come next - that's how the answer takes shape.
Linguistic plausibility
What comes out sounds right. Whether it's true is another question.
When a language model answers your question, it doesn't think. It calculates. From what you typed in, and from everything it read during training, it works out: "Which word would be the most likely next one now?" Then the next one. Then the next. That's how an answer takes shape - word by word, without the model knowing what it's talking about.
That's not a put-down. It's an important distinction. The model has no idea what an apple is, what a contract means, what a knee operation actually involves. It has seen a great many sentences about apples, contracts and knee operations, and can plausibly carry them on. That's enough for many things. It's not enough for everything - and the difference is precisely what matters.
The technical foundation is called a Large Language Model, or LLM for short. Well-known examples include ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google) and Mistral (French, EU-based). They all work on the same principle: a neural network with billions of parameters, trained on huge amounts of text.
The breakthrough came with the Transformer architecture (Vaswani et al., "Attention is All You Need", 2017) - a mathematical approach that lets models attend to many places in a text simultaneously. Since ChatGPT (late 2022), LLMs have entered mainstream awareness. What has changed since then is the quality - not the underlying principle.
A language model has a cutoff date. Up until a certain point it has read texts from the internet, from books, from Wikipedia, from forums, from scientific publications. After that, nothing. Whatever happened after the cutoff, it doesn't know - even when it acts as if it did. And whatever was never publicly written down, it doesn't know either.
There are two consequences. First: for anything current, a language model is unreliable - yesterday's exchange rate, the new law, the rail strike announced this morning. Second: about your specific data - your customers, your prices, your processes - the AI knows nothing, because that data wasn't in the training set. If AI is to be useful in your business, you have to connect it to your data. That's exactly the step where most AI attempts in small and mid-sized businesses get stuck.
The technical term is knowledge cutoff. For current models, it typically sits between a few months and a year before today's date. Some models additionally come with a browsing mode - they can query the internet when the question calls for it. That makes them more current, but not necessarily more correct.
The technique for combining an AI with your own data is called Retrieval Augmented Generation, or RAG for short: before each answer, the AI is given a relevant excerpt from your own knowledge base to read, and answers on that basis. This is the standard pattern for AI in business today - and exactly what Wendwerk builds in projects where AI is supposed to work with your content.
When you use ChatGPT, Claude or a similar tool in your browser, the model isn't running on your computer. It runs in a data centre - OpenAI's in the US, Anthropic's in the US, Google's worldwide. Your input is sent there over the internet, the calculation happens there, the answer comes back. This matters, because your input leaves the building when you press enter.
For uncritical content - drafting an email template, shortening a piece of text, asking a research question - that's fine. For sensitive content - customer names, contracts, prices, internal figures - it becomes a question. What happens to the data? Is it used for further training? Does it stay in the EU? How we handle that is on the page How we handle data protection.
The free consumer tools (chat.openai.com, claude.ai) generally use your inputs for further training unless you explicitly opt out. The paid business contracts (OpenAI ChatGPT Team/Enterprise, Anthropic Claude for Work, API access under a contract) don't - here, the inputs are contractually excluded from training.
For client projects, Wendwerk uses exclusively the API access with a business contract - usually Anthropic Claude, and depending on the use case other providers too. Which providers are used in any given project is stated openly in the contract, and sensitive processing runs on routes that stay within the EU wherever the use case requires it.
AI is strong at anything to do with language: summarising, rephrasing, translating, drafting, tailoring a text to a particular audience. It's strong at sorting - placing a stack of texts into categories, reading sentiment out of reviews, picking the important ones out of a flood of emails. And it's strong at suggesting - replies, phrasings, templates that you only need to adjust.
What these have in common: AI is an accelerator for the first 80 percent. The draft, the rough shape, the first version. The last 20 percent - the part that makes the difference between "okay" and "really good" - stays with the human. Take that to heart and you gain time without losing quality. Ignore it and let AI handle the whole 100 percent, and you get mediocre output in large quantities.
A widely cited Harvard study from 2023 (Dell'Acqua et al., "Navigating the Jagged Technological Frontier") had consultants work on real tasks with and without AI. The result: on tasks that fell within the AI's strengths, the AI users were on average 25 percent faster and rated 40 percent better. On tasks outside its strengths, the AI was not only unhelpful but led to measurably worse results - because users followed its wrong suggestions.
From this follows a simple rule of thumb: use AI where it is strong - but know where that is. Language, sorting, suggesting, summarising are the strong fields. Arithmetic, precise fact-checking, legal judgement, medical diagnosis are the weak ones - there, AI can be a tool for the professional, but not the source of truth.
AI is not a consciousness, not a person, not intelligence in the proper sense of the word. It has no intentions, no opinions, no self-interest. What sounds like an opinion is a likely turn of phrase. What sounds like sympathy is a language pattern from the training data. That doesn't make AI any less useful - but it makes it a tool, not a counterpart.
And AI is also not an Artificial General Intelligence - that would be an AI that can do everything a human can, plus more. Today's language models are far from that, even if some reports make it sound otherwise. What models will go on to learn in the next few years is an open question. What they are today is clearly definable - and exactly that clarity helps you put them to good use.
The technical term for the "AI that can do everything" is Artificial General Intelligence, or AGI for short. Whether it's coming, when, and in what form, is disputed within the research community. What exists today is narrow AI - narrow systems specialised in particular tasks. Even a vast language model like GPT-4 is, in this sense, narrow AI: specialised in language processing, not in everything.
A recurring confusion is the ELIZA effect: back in the 1960s, Joseph Weizenbaum built a simple chatbot called ELIZA, which asked standard therapeutic questions. The test subjects developed emotional attachment to the machine - even though they knew it was just a program. With today's language models, the temptation is incomparably greater. Being aware of the temptation is the first step toward dealing with AI wisely.
We put AI to work where it honestly saves time - and not where it merely looks good. Three typical fields: First, in diagnosis - we use AI to analyse your answers from the questionnaire, summarise observations, suggest priorities. Second, in the software itself, when the use case carries it - a search that understands natural language, a classification of incoming emails, a suggestion system for replies.
Third, in the daily building itself - when developing your software, we work with AI tools that make us more productive. But in none of these three cases is AI on its own. Where it makes content decisions, we double-check. Where it's sensitive, it runs in contractually secured environments. Where it isn't needed, we leave it out - because a simple form is often better than a clever bot.
In our projects, we work mainly with Anthropic Claude via the API under a business contract. Depending on the use case, other models come in too - European providers, for instance, when a project explicitly calls for EU processing, or specialised models for classification, translation, or image processing. Which model is used for which purpose, we set down transparently in the project contract.
For everything AI decides in your software, there's a log: which input, which response, which human approved it. That doesn't only give legal traceability - it also lets you learn where the AI is good and where it regularly gets things wrong. If in doubt, we can remove the AI part later, or replace it with a simpler approach - the software stays functional, the AI portion is always swappable.
AI is a powerful tool that's easy to misunderstand. Those who understand it use it better - and don't become dependent on it.
This page has been deliberately sober. No doomsaying, no promises of salvation, no hype. Whoever knows what AI is and what it isn't can decide with a clear head where it helps in the business - and where the old, proven approach remains the right one. That's the stance with which we build AI into every one of our projects.
What risks AI brings and how we deal with them is on What AI can't do, even when it looks like it can. How we handle data protection concretely in our projects, you'll read on How we handle data protection. And why we never build AI first, you'll find on Tools first, AI second.