LLM (large language model)

What matters for a company is what an LLM does well: understanding freely written text. A customer writes an email with no structure, with typos and three questions at once - and the model extracts what it is about, how urgent it is, and what needs to be answered. In the same way it reads an invoice in PDF, summarizes a long document, or writes a draft post in the brand’s tone. This used to require a person, because a classic program could only work with data in a fixed format.

Equally important is what an LLM cannot do: vouch for facts. A model generates probable text, not verified truth, and when it lacks information, it can convincingly make it up. That is why we do not let it answer from memory in automations. It receives specific source materials (price list, terms, order status), rules check the output (the price matches, the link exists, no promise was made that the company does not give), and a person decides at important steps.

Which model is used depends on the task. Sorting emails can be handled by a smaller, cheaper model; more demanding text or work with long documents needs a stronger one. We go through the choice of model, and where the data flows, during the process analysis, before anything is connected.