What is a large language model?

In one sentence

A large language model, or LLM, is the part of an AI app that reads text and writes new text, having learned the patterns of language from a huge amount of writing. It is very good at understanding what someone means and writing a natural reply, but it only knows your business if you tell it.

In more detail

Imagine someone who has read an enormous library: books, websites, letters, manuals and conversations in many languages. They have not memorised every page, but they have soaked up how people write, how questions are usually answered and how one idea tends to follow another. Ask them almost anything and they can write a sensible-sounding answer in the right tone. That is roughly what a large language model is.

When the model writes, it builds the reply a small piece at a time, each time choosing what is most likely to come next given everything it has read so far in the conversation. It does this so well that the result reads like a person wrote it. It is not looking anything up while it writes unless it has been given a tool to do so.

The model is the engine, and the app is the car. Claude is made by Anthropic, ChatGPT runs on models from OpenAI, and Grok comes from xAI. Each company trains its own models and puts them inside an app you can talk to. The app adds things around the model, such as a chat window, the ability to read files, and connectors that let it use outside tools.

Because it learned from general writing, a model knows a great deal about the world in general and nothing at all about your business in particular. It does not know your opening hours, your prices, which van is booked on Thursday or that you stopped offering a service last month. If it has not been told, it may guess, and a guess written fluently can sound exactly like a fact.

A model also has a cut-off. It learned from writing gathered up to a certain date, so it may not know about recent events or changes unless the app gives it a way to read newer information. For a business, the newest information that matters is usually your own: today's availability, this week's offer, the message a customer sent five minutes ago. That is why a model works best when it is connected to your messages and given a clear set of notes. The example below shows the same question answered with and without those notes. Names and details are examples.

An example

What a language model is good at, and where it needs help

The left column is why language models feel so useful. The right column is where most mistakes come from, and every item on it is fixed in the same way: give the model the facts, the tools to check the rest, and clear rules about when to hand over.

Model, app and agent: three words that get mixed up

• The model is the language engine itself. It reads and writes text. On its own it cannot send anything anywhere.
• The app is what you open and type into, such as Claude, ChatGPT or Grok. It wraps the model in a chat window and adds features.
• The agent is the app, or a program built on a model, once it has tools it can use to take actions for you, such as reading your WhatsApp messages and sending a reply.

So when people say "the AI answered my customer", what usually happened is this: an agent, built on a model, read the message through a tool, wrote a reply using the model and sent it through another tool. Each part can be checked and limited separately, which is good news for you.

Why the notes you give it matter so much

A language model fills gaps. That is part of what makes it good at conversation, and it is also why it can be wrong. The fewer gaps you leave, the fewer guesses it makes. A short set of notes covering your facts and your boundaries does more for reply quality than any choice of model. Here is a simple starting point. Names and details are examples.

You answer customer messages for Kingsway Keys & Locks.

Facts:
- Open Monday to Saturday, 08:30 to 17:30. Closed Sundays.
- Standard house keys: £6.50 each, cut while you wait.
- Security and car keys: ordered in, about a week, price on request.
- Emergency lock-outs: call 020 0000 0000. Do not quote a price.

Rules:
- If a price or service is not listed above, say you will check and a
  person will reply. Never guess a price.
- Hand the chat to a person for complaints, insurance jobs or anyone locked
  out right now.
- Keep replies short and friendly. Use British spelling.

The line "never guess a price" is doing a lot of work here. Models will usually follow a clear rule like that, and it turns a confident wrong answer into an honest "let me check".

Common misunderstandings

• Thinking it searches the internet for every answer. Unless the app has a search tool switched on, the model writes from what it learned and what is in the conversation.
• Thinking it learns from each chat with your customers. In most set-ups the model itself does not change from one conversation to the next. What changes is the notes and history you give it.
• Thinking a bigger or newer model removes the need for notes. Better models make fewer mistakes, and every model still needs to be told your prices.
• Thinking it knows when it is wrong. A model can write a wrong answer with exactly the same confidence as a right one.
• Thinking the model sends the message. The model writes. The tools you connect decide whether and how anything is sent.

Why it matters for your business

• Knowing that the model only knows what you tell it explains why good notes are the first job, not an afterthought.
• Knowing it can sound sure while being wrong tells you which replies to hold for a person, such as prices, refunds and promises.
• Knowing the model writes and the tools send helps you put limits in the right place: on the tools and the accounts.
• Knowing the model, the app and the agent are different parts makes it easier to switch apps later without starting again.
• Understanding the strengths lets you hand over the work a model does best, like understanding messy messages and answering in any language.

How Spaces handles it

Spaces is not a language model and does not choose one for you. You bring the AI app you already use, such as Claude, ChatGPT, Grok, Cursor, Codex, OpenClaw or Hermes, and connect it to Spaces through the MCP server. The model inside that app does the reading and writing. Spaces gives it the tools to see your WhatsApp, Instagram, LinkedIn, Messenger, X and email conversations and to reply in them.

To fill the gaps a model would otherwise guess at, Spaces keeps a business brief: your hours, what you sell, prices and links, common questions, tone, languages, what never to promise, when to hand a chat to a person and an urgent contact. The agent is told to read it before replying, and because it lives in Spaces rather than in one app, every agent you connect reads the same facts.

Whatever the model writes, sending goes through Spaces, so every message is paced and counted against the account limits. On WhatsApp, for example, sends are spaced at least 10 seconds apart and capped at 400 a day per number. If you want to read replies before they go, ask your AI agent to hold replies as drafts, and nothing goes out until you approve it.

In short

A large language model is a very good writer that has read a great deal and knows nothing about your business until you tell it. Give it your facts, tell it where the edges are, and let the tools around it handle checking, sending and limits.

If you remember one thing, make it this: the model fills gaps, so leave as few as you can.

Is a large language model the same as an AI agent?

No. The model is the part that reads and writes text. An agent is an app or program built on a model that also has tools, so it can take actions such as reading your messages and sending replies. Claude, ChatGPT and Grok become agents once you connect tools to them.

Which language model is best for answering customers?

The leading models from Anthropic, OpenAI and xAI all write good customer replies. The difference your customers notice comes far more from your notes and rules than from the model. Start with the app you already use and are comfortable with.

Does the model learn from my customer conversations?

In most set-ups the model itself does not change from chat to chat. It reads the conversation and your notes each time it replies. To make it better, improve the notes rather than hoping it learns on its own.

Why did it make up a price?

Because it was asked a question its notes did not answer, and models fill gaps with what sounds likely. Add the real price to your notes, and add a rule that it should say it will check rather than guess any price that is not listed.

Does a language model know what day it is or what is in stock?

Not by itself. The app may tell it the date, and it can only know stock, bookings or availability if you give it a tool or notes that hold that information. Spaces does not connect to shops or booking systems, so your agent needs its own tools for those.

Can a language model write in other languages?

Yes, and this is one of its real strengths. Most models can read and reply in many languages. Tell it in your notes which languages you are happy for it to answer in and which facts must stay exactly as written, such as prices.

Do I need to understand how the model works to use it?

No. You need to know three things: it only knows what you tell it, it can be confidently wrong, and the tools you connect decide what it can actually do. With those in mind you can set it up sensibly.

Does Spaces include a language model?

No. Spaces connects your messaging accounts to the AI agent you already use and keeps sending inside safe limits. The model belongs to the app you bring, such as Claude, ChatGPT or Grok.