Named entity recognition is the NLP task of finding entities in free text and labelling them by type: people, organisations, places, dates, monetary amounts, products. The model settles two things at once, where the mention starts and ends and which category it belongs to.
Nothing works without context. "Washington" can be a person, a state or a city depending on the sentence, and "Ford" can be a brand or a surname. Rule-based systems gave way to models that read context for this reason. Languages with rich inflection add difficulty, since entity names take endings and capitalisation is not always a reliable clue.
The output is usually the input to the next step. Entities get linked to identifiers in a knowledge base, search queries are interpreted through those identifiers, and structured fields are extracted from documents.
A concrete case: a law firm needs party names, signature date, term and value pulled out of incoming contracts. Those fields sit in different places and different phrasings in every contract. After learning from a few thousand annotated contracts, the NER model finds the same fields in new documents and opening a file stops being a reading job.
The error modes are particular too. A model can draw the boundary wrongly, catching only part of a long organisation name. Partial matches like that get counted separately during evaluation.