Glossary · simply explained

LLM (large language model)

A large language model (LLM) is a neural network trained on vast amounts of text that models language statistically: it predicts the most likely continuation token by token — and can thus understand, summarise and translate text, write code and answer questions. LLMs are the foundation of services such as ChatGPT, Claude and the copilots in office tools.

For companies, LLMs are simultaneously a productivity lever and a governance topic: they measurably accelerate knowledge work but bring their own risks — from hallucinations via data leakage to new attack surfaces such as prompt injection.

How LLMs work — in brief

The technical basis is the transformer architecture: texts are split into tokens, represented as vectors and related via attention mechanisms. In pre-training the model learns language patterns from billions of documents; fine-tuning and alignment then direct it towards helpful, safe behaviour. The context window limits how much text the model can consider per request.

Important for expectation management: an LLM is not a reference work but a probability model — it can produce convincingly worded misinformation (hallucinations) and only knows the state of its training. It becomes reliable in enterprise use through connection to your own vetted knowledge sources (RAG) and human control of critical outputs.

Enterprise use: the switch points

  • Sourcing model: APIs of large providers, European hosting or self-operated open-weight models — depending on privacy and cost.
  • Privacy: clear rules which data may go into which models; check DPAs and processing locations.
  • Grounding: RAG instead of blind trust — answers from own sources with citations.
  • Governance: make usage visible (shadow AI) and map EU AI Act duties.

Frequently asked questions about LLM (large language model)

What does hallucination mean with LLMs?

The model produces fluent, plausible-sounding statements that are factually wrong — including invented sources. This is not an occasional bug but a consequence of the working principle. Antidotes: grounding via RAG, mandatory citations, review processes for critical outputs.

Do LLMs learn from our inputs?

That depends on the contract: business and API offerings of reputable providers exclude training on customer data by default; free consumer services partly do not. Exactly why business data belongs only in approved services with a DPA — and shadow AI needs reining in.

Run your own model or use an API?

APIs deliver top quality without infrastructure — at running costs and data flow to the provider. Open-weight models on own or EU infrastructure maximise control but demand GPU capacity and operational skill. Many go two-track: sensitive cases internally, breadth via API.

What are tokens and why do they determine costs?

Tokens are the model’s processing units — word fragments of ~4 characters on average. APIs bill per input and output token, and the context window limits the token amount per request. Long documents and chat histories therefore drive cost and latency.

Which security risks do LLMs bring?

Three classes: data leakage through uncontrolled use, manipulated model outputs through prompt injection — critical as soon as LLMs drive tools — and classic supply chain/API risks of the integrated services. Access control, output filters and AI gateways address this.

Wondering how this looks in your own network? Talk to KAEMI: we plan, build and manage the right solution with you.