What AI actually is

Artificial intelligence is the broad field: software that handles tasks we normally associate with human judgment, such as recognising speech, sorting messages or writing text.

Machine learning is one way to build AI. Instead of writing every rule by hand, you show the system many examples and it learns the patterns. A spam filter or a lead-scoring model is machine learning. It outputs a label or a number, not new content.

Generative AI is machine learning that produces new content: text, images, audio or code. A large language model (LLM) is the generative kind you will use most. It is trained on a very large amount of text and produces text in response to text.

This matters when you evaluate tools. When a vendor says a feature is AI-powered, ask which kind. A scoring model and an LLM fail in different ways and need different checks.

An LLM starts life through pretraining: it reads a huge body of text and learns to predict the next word given everything before it. Anthropic's glossary notes that a model at this stage is not good at answering questions or following instructions. Further training turns it into an assistant that follows instructions.

When you send a prompt, the model generates its reply one small piece at a time, each piece chosen because it fits what came before. It is not looking anything up unless the app around it gives it a search or retrieval tool. It is producing the most fitting continuation based on patterns it learned.

Many current models are also multimodal, meaning they can take images or documents as input as well as text. The core behaviour is the same: input in, generated output out.

LLMs are strong at work that involves messy language: summarising a long email thread, drafting a reply, pulling a name and phone number out of a free-text form, classifying a support ticket, rewriting copy for a different audience, or translating.

A realistic business example: a lead fills in a contact form with a rambling paragraph. A model can turn that paragraph into a one-line summary, a service category and an urgency flag that your CRM can store. A rule-based filter would struggle because no two leads phrase things the same way.

A model has no knowledge of events after its training data ends unless you supply it. It cannot take actions such as sending an email or updating a contact unless it is connected to tools. It can be confidently wrong, and the same prompt can produce different wording each time.

It also does not remember you between separate conversations unless the app stores memory and feeds it back in. And it is the wrong tool for exact, repeatable work like adding up invoice totals or checking whether a date has passed. A formula or a normal automation step does that reliably and for free.

A useful rule from day one: deterministic tasks get deterministic automation. Use a model where the input is unstructured language and some judgment is needed, and check its output where mistakes are costly.