01What it is
A large language model, or LLM, is a program trained on very large amounts of text. From that training it has learned how language works and a great deal of general knowledge. It can read, summarise, sort and write text, answer questions about documents, translate, and, in newer models, describe images.
Well-known models are Claude by Anthropic, GPT by OpenAI and Gemini by Google. Companies use them through a chat window such as ChatGPT, or, inside their own software, through an API: the application sends a text to the model and receives the answer back.
02How it works
A language model works with tokens, small pieces of text, often a word or part of a word. Given a text, it calculates which token is most likely to come next, adds it, and repeats. The answer is written token by token.
The model only knows what it learned in training and what it is given in the request. It has no access to your files or systems unless an application passes that information to it. This is why a business application usually does two steps: it first searches its own data for the relevant passages, then hands those passages to the model together with the question. The model answers from what it was given and can name where each statement came from.
The model does not remember earlier requests by itself. Each request is processed on its own, and the application decides what context to include.
03What they are good for
Language models do well at work that used to need a person reading through text. Typical uses in a company:
- Sorting incoming email or support requests by topic and urgency, with a draft reply.
- Pulling figures and statements out of reports, contracts or presentations.
- Finding a document by its content when nobody remembers the file name.
- Summarising long documents or meeting notes.
- Describing images on slides or in photos, so they become searchable.
- Answering questions about a company's own records, with links to the source.
04Their limits
A language model can state something false with full confidence. This is called hallucination. It happens most when the model is asked about something it was not given and does not know.
An application can reduce this a lot: by giving the model the relevant documents, asking it to answer only from them, and showing the source next to every answer so a person can check it in one click. For decisions with consequences, a person should check the result before it is used.
Models are also not calculators. For sums and totals, the application calculates and the model only phrases the result.
05Data protection and training
The main concern for companies is what happens to the data sent to a model. The answer depends on the provider and the contract.
Free and private versions of chat tools may use conversations to train future models. Business offers and API access from the large providers exclude training on customer data by contract. Providers also differ in where the data is processed and how long requests are kept, for example to detect abuse.
Through cloud providers, the large models can be used with processing in the EU: Claude through Amazon Bedrock, OpenAI's models through Microsoft Azure, Gemini through Google Cloud. If personal data is processed, a data processing agreement under Art. 28 GDPR is needed with the provider.
06Costs
Through an API, models are billed by usage: per token sent to the model and per token it writes back. Larger, more capable models cost more per token than smaller, faster ones.
Costs therefore depend on how much text an application processes. Sending only the relevant passages instead of whole documents, using a smaller model where it is good enough, and processing a document once instead of on every question keep the bills low. We estimate the running costs for an application before it is built and set spending limits with the provider.
07How we use them
We call models only through providers that process the data in the EU and rule out training on it by contract. Kadvern One, our own operating system, runs its assistant on Claude through Amazon Bedrock in Frankfurt. For a document search we are building for a client, we use OpenAI's models through Microsoft Azure with processing limited to the EU.
AI answers in our applications name their source, so the person reading can check them. The code of an application stays independent of a single provider, so the model can be exchanged later if a better or cheaper one appears.
We also write our code with AI agents. How they check each other's work is described in the method section on our home page.
Frequently asked questions
Will our data be used to train the model?
Not with the providers we use. Their business and API terms exclude training on customer data, and we only use those terms.
Does the data leave the EU?
We set up processing in the EU, for example Claude on Amazon Bedrock in Frankfurt or OpenAI's models on Microsoft Azure with an EU data zone. The providers are US companies; if that matters for you, we discuss it before the project starts.
What if the AI gives a wrong answer?
It can. That is why our applications show the source of each answer and why decisions with consequences stay with a person. Answers based on your own documents are far more reliable than answers from the model's general knowledge.
What does it cost to run?
It depends on the amount of text processed. For most business applications the model costs are a small part of the running costs. We estimate them before building and set a monthly limit with the provider.
Can we use our existing ChatGPT or Claude accounts instead?
For individual work, yes, as long as everyone uses the logged-in business account. An application that works with your own data and shows sources needs API access, which is set up separately.