An LLM, or large language model, is the technology that lets software read a sentence written in plain English and work out what it means, then produce a plausible reply in the same language. It learns to do that by reading huge amounts of existing text rather than following a fixed set of rules written for every situation in advance. On its own, an LLM only ever produces text: an answer, a summary, a draft. It has no memory of yesterday's conversation, no job it's responsible for, and no way to act on what it just told you unless something else is built around it to do that.
You've seen the term everywhere by now, on a login screen, in a product feature list, in a pitch that says "powered by" a model you've never heard of. Most of those pitches never explain what the words mean before asking you to trust the product with your customer emails or your invoices. That gap matters, because how much you should trust any AI tool, including one of our own agents, depends on understanding what the LLM underneath it is good at and where it tends to guess.
Once you know what the model is actually doing, buying decisions about AI stop being a guess of your own. You can ask a supplier a sharper question than "is this good AI." You can ask what the model was trained on, what happens when it doesn't know something, and who checks its work before it reaches a customer.
An LLM learns language patterns from text, not facts stored in a database
Ask a database a question and it looks up a stored answer, correct or not, exactly as somebody entered it. An LLM doesn't work that way. It's trained by reading enormous amounts of text and learning the patterns of which words tend to follow which other words, in which order, given the context so far. That training gives it a working grasp of grammar, tone, and the shape of an argument, without giving it a lookup table of verified facts. It can write a fluent, well-structured paragraph about your industry without having a single confirmed figure about your industry stored anywhere inside it. That's a strength for tasks like drafting and summarising, where fluent language is the whole job, and a weakness for anything that needs one exact number pulled from a real record.
An LLM answers with the most likely next words, which is why it can sound certain while being wrong
Every reply an LLM gives is built one likely word at a time, predicting what comes next given everything written so far. Most of the time that produces a correct, useful answer, because most questions have an answer well represented in what the model learned from. But the model has no separate step where it checks the words it just produced against the real world before sending them. It can produce a wrong answer in exactly the same confident, well-punctuated tone as a right one, because tone and accuracy come from two different places. We've written separately about what actually stops an AI agent from doing this in practice: the short version is that the fix sits in the agent built around the model, not in the model deciding to be more careful.
The LLM does the reasoning; an agent built around it does your business's job
An LLM answering one question in a chat window is not the same thing as an AI agent running part of your business. The model supplies the reasoning, reading a request and working out what it means, but it has no job description, no memory of what it did an hour ago, and no rule about when to stop and ask a person before acting. An agent is what you get when that reasoning sits inside something with a written job: what task it owns, what tools it's allowed to touch, and when it hands a decision to a person rather than deciding alone. The LLM reads the supplier email and works out what's being asked; the agent is the part that checks that against what was already agreed, drafts the reply, and waits for someone to approve it before it goes anywhere. Confusing the two is common, and it's also why "AI" and "AI agent" get used as if they mean the same thing when they don't.
Newer LLMs read more than text, but the core skill stays the same
The first widely used models only read and wrote plain text. Newer ones can also read an image, a scanned invoice, or a photographed form and describe what's in it, which matters to a business that deals in paperwork rather than clean data entry. That's a real expansion of what the model can take in, not a change to how it works underneath: it's still predicting the most likely correct response to what it's been shown, in language, rather than looking anything up in a verified record. A model that can read a delivery note photographed on a warehouse floor is more useful to more businesses than one that only reads typed text, but it carries the same basic habit of guessing fluently when it isn't sure.
An LLM doesn't know when it's guessing. Ask it something outside what it was trained on, or something that needs a number from today rather than a pattern learned months ago, and it can produce a wrong answer with the same tone as a correct one. That is a basic feature of how the technology works, and no future update removes it. It's why every agent built on an LLM at The Agentry has a human check on anything that leaves the business or spends money, rather than trusting the model's confidence as the only test of whether it was right.
You don't need to understand the maths behind an LLM to use one well. You need to judge any AI tool, including an agent, by what the model actually does: read, reason in language, and produce a draft, not a decision that's already final. Once you know that, the question stops being "is this AI good" and becomes "what job is this actually right for."