Quick answer
A chat assistant works in 4 steps: it is trained on a very large amount of text beforehand, it breaks your request into small pieces of text (tokens), it predicts what comes next, and it writes that out. This is this site's plain-words summary of Google's and Anthropic's documentation; Google's course calls large language models "essentially autocomplete mechanisms".
Key takeaways
- Follow one request, a reply to a customer email, through 4 steps: training, tokens, prediction, output. Each step below is tied to a sentence from Google's Machine Learning Crash Course or Anthropic's documentation.
- Training happens before you type. Google says the primary ingredient is a phenomenal amount of text; the sentence does not mention your prices, calendar or customers, so give them to the model in the request.
- Google's course says prompt wording matters: 'LLMs like context.' A request with the facts in it gives the model more to continue from.
- Google's course says LLMs hallucinate, meaning their predictions often contain mistakes. This site's habit: check every name, date, number and promise the draft adds.
How does AI work, in simple terms?
A chat assistant works in 4 steps: it is trained on a very large amount of text beforehand, it breaks your request into small pieces of text (tokens), it predicts what comes next, and it writes that out. That is this site's plain-words summary of the sources below. Google's Machine Learning Crash Course says "LLMs are essentially autocomplete mechanisms that can automatically predict (complete) thousands of tokens."
This page explains language models, the kind of AI behind chat assistants; other kinds of AI exist. NIST's glossary, in a definition it credits to NIST SP 800-218A, describes artificial intelligence broadly as "A machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations, or decisions influencing real or virtual environments." The page follows one request through the four steps: "Draft a reply to this customer email asking if we can start the job next Tuesday." Each step has a sentence from Google's course or Anthropic's documentation (read October 7, 2026), then what it means for you.
Disclosure: this page was drafted with Claude, a model made by Anthropic, and Anthropic's documentation is one of the sources quoted. The method note at the bottom says more.
What happens to the request, step by step?
Four steps, and only the last three happen when you press enter. The first happened long before.
| Step | What a source says | What it means for you (this site's view) |
|---|---|---|
| 1. Training (before you arrive) | Google: "The primary ingredient in building an LLM is a phenomenal amount of training data (text), typically somewhat filtered." | Training starts from a very large body of text. Google's sentence does not mention your prices, calendar or customers, so give them to the model in your request. |
| 2. Your request becomes tokens | Google: "A token could be a word, a subword (a subset of a word), or even a single character." | The model works on pieces of text, called tokens. Splitting your request into tokens is a step you never see. |
| 3. The model predicts what comes next | Google: a language model "estimates the probability of a token or sequence of tokens occurring within a longer sequence of tokens." | The reply is built by estimating what is likely to come next. This site's view: a likely sentence is not the same as a confirmed fact, so check the specifics. |
| 4. Output | Google: "LLMs hallucinate, meaning their predictions often contain mistakes." | A draft can read well and still contain a wrong detail. You are the check. |
Step 1: What happens before you type?
The model is trained first, on a very large amount of text. The large language model (LLM) is the product of that training. Google's course describes the task in one sentence: "The model trains by trying to predict those missing tokens," where tokens are hidden from the training text on purpose. Anthropic's glossary names the first phase: "Pretraining is the initial process of training language models on a large unlabeled corpus of text."
Training is done by the company that builds the model, not by you; Google's course says "You probably will never train an LLM from scratch." Anthropic's glossary adds that pretrained models are "not inherently good at answering questions or following instructions", and that "Claude is not a bare language model; it has already been fine-tuned to be a helpful assistant." Only Anthropic's page was read on that point, so this page does not say how other products handle it.
Step 2: What does the model do with your wording?
It splits your request into tokens, the units it reads and writes. In the example email, "Tuesday" might be one token or several; that detail is the model's, not yours. Google's page on prompt engineering says "end users clarify how the LLM should respond to their prompt" and, on examples without context, "LLMs like context." Its illustration: a bare "apple:" could return information about the technology company rather than the fruit.
For the email, "Reply to this" gives the model little. "Reply to this email from a customer; we can start Tuesday at 8 a.m.; we need a deposit confirmed first; keep it to four sentences" gives it the facts and the shape. Anthropic's glossary describes the "context window" as "the amount of text a language model can look back on and reference when generating new text", and says this "is different from the large corpus of data the language model was trained on". In plain words: what you paste is the model's working notes for this one job.
Step 3: How does the model decide what to write?
It estimates which token is likely to come next. Google's course defines a language model as something that "estimates the probability of a token or sequence of tokens occurring within a longer sequence of tokens." On its fill-in-the-blank example, the prediction "might be the highest probability" or "a random selection from tokens having a probability greater than a certain threshold".
Think of the word suggestions above your phone keyboard. That is this site's analogy, and its only tie to the sources is Google's "autocomplete mechanisms" sentence. It stops there: it says nothing about how either is built or how reliable either is.
Two practical points follow. Where a product allows a random selection, two runs of the same request can differ, so a second try can give a different email; the course does not say how any product is set. And whether the model "understands" your customer is not settled by the sources read here, so this page does not say. For an owner the useful point is narrower: a likely continuation and a correct statement are different things.
Step 4: Why can the finished draft be wrong?
Google's course states it directly: "LLMs hallucinate, meaning their predictions often contain mistakes." It lists this among the problems with using LLMs to make predictions. For the email, the risky parts are the specifics: a start time you did not give, a deposit rule you do not have, a promise you did not make. The course says "often", not always, and it does not say which details fail.
If you want the model to answer from your documents, that is a separate technique. Anthropic's glossary describes it: "In RAG, a language model is augmented with an external knowledge base or a set of documents that is passed into the context window." Whether a given product does this is stated on that product's pages.
What should you do with this?
Give the facts, ask for the shape, then check what it added. This is this site's habit, built on the four steps above, not a vendor procedure.
- Put your facts in the request. Dates, prices, terms, tone, length. Neither source read here says your business is in the training text, so put the facts in the request.
- Say what the reply must not do. For example: "do not promise a time we have not confirmed." The course shows examples and context shaping output, not that any instruction is always obeyed.
- Mark everything the draft added. Underline each name, date, number and promise that was not in your request.
- Check the underlined parts. Compare dates, prices and names with your own records, or ask a person who knows, then send.
For tools that write to customers on their own, the stakes rise. Read our guide to AI agents for small businesses and the practical guide to AI automation. A tool built for exactly this job is covered in AI email responders.
How was this page made?
The steps come from Google's Machine Learning Crash Course (three pages), Anthropic's documentation glossary (pretraining, the context window, retrieval) and NIST's glossary entry for artificial intelligence. Google and Anthropic also sell AI products, so read their pages as their descriptions, not independent tests. This page was drafted with Claude, made by Anthropic, one of the sources read; it is not one of several vendors compared. The email, its details and the habit are this site's illustration, not a case study.
Next step: take the next customer email you would answer, paste in your facts, ask the assistant for a draft, and underline everything it added that you did not give it.
FAQ
Does AI look things up on the internet when I ask it something?
Ask the seller whether the product searches the web or your files; the sources read here do not say. Google's and Anthropic's pages describe training on text and then predicting the next tokens, not a lookup. Anthropic's glossary describes retrieving from documents as a technique: 'a language model is augmented with an external knowledge base or a set of documents that is passed into the context window'.
Does what I type teach the AI about my business?
Read the product's own terms and settings before you paste in anything private. Google's course says 'Prompt engineering doesn't alter the model's parameters.' That sentence is about writing and refining prompts, not about what a given product stores or reuses afterward.
Is one AI tool the same as another under the hood?
Ask the seller which model runs the product and what happens when it is wrong. The sources here describe language models in general; tools can differ in the model they use and in what they add around it.
Is a chat assistant the same thing as AI?
Treat a chat assistant as one kind of AI, not the whole field. NIST's glossary, in a definition it credits to NIST SP 800-218A, describes artificial intelligence as 'A machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations, or decisions influencing real or virtual environments.' This page covers only language models.