LLM stands for Large Language Model - a system trained on an enormous quantity of text that predicts which word logically follows the previous one. From that single capability come the 4 things such models do: writing and reshaping text, summarising long documents, translating, and answering questions asked in plain language.
The model does not search a database of answers; it predicts the continuation of text. Based on everything it saw during training, it calculates which word most likely follows, then repeats that word by word until an answer forms.
Because of that the model has no concept of truth, only of probability. A statement that sounds convincing and one that is actually correct look similar to it - and that is where hallucination arises, the invention of a fact that does not exist.
The same mechanism explains why the model is strong at language and weak at arithmetic. Predicting the next word captures the shape of a sentence beautifully but does not perform an exact calculation.
Reshaping text that already exists - summarising a long document, translating, changing tone, extracting data from an untidy message.
Understanding intent. The model reliably recognises that "how long do I wait for the site" and "what is your build timeline" mean the same thing, which classic rule-based automation found difficult.
Work where the input is unstructured - an email, a document, a photo of an invoice. That is why it is used in business process automation.
The model knows nothing about your company unless it is given the information. Prices, timelines and terms are not part of its training, which is what RAG is for - an approach where the model answers from your documents.
Its knowledge has a cut-off in time. It knows what existed up to the moment of training and does not know later events unless they are supplied.
Consistency is not guaranteed. The same question asked twice can produce a differently worded answer, which is a problem anywhere repeatability is required.
And most important for business use: the model does not know what it does not know. Without an explicit instruction to admit ignorance, it will fill the gap with something plausible.
For business use you almost never need your own model. Existing models connected to your data cover practically every case at a fraction of the cost.
The choice of model matters less than how it is configured. A well-configured cheaper model with access to the right documents outperforms a more expensive model without context.
Prices are measured by volume of text processed and have mostly fallen in recent years. That is why systems are built so the model can be swapped without rewriting the rest - what is the best choice today usually is not in a year.
Human review remains necessary anywhere output goes straight to a client. The model speeds up preparation, but responsibility for a sent quote stays with the company.
For a company considering adoption, the first question is not which model but which task. The model is a tool, and a tool is chosen after the job, not before.
Cost is charged by volume of text processed and for most business applications is surprisingly low - often less than a single monthly business software subscription. The larger cost is almost always the build and integration, not the model itself.
Cost falls further with sensible configuration. Short, precise prompts and sending only the relevant parts of a document rather than the whole thing can halve the bill with no loss of quality.
When deciding, separate two things. The first is intent - how well the model recognises what the user is actually asking, which is already very good even on cheaper models. The second is access to your data, which is a matter of how the system is set up. The second decides answer quality far more than the first.
Because models change quickly, the system is built so one can be replaced without rewriting the rest. What offers the best price-to-quality ratio today usually does not a year from now.
Last updated: 17 August 2026
Send the address and we will tell you where you stand - no obligation.
Send an enquiry
