AI Is a Prediction Engine, Not a Brain

AI does not think. It predicts. Every time you type a prompt and hit enter, you're asking a very sophisticated autocomplete system to guess what words should come next. That single insight will make you better at using AI than 90% of people who use it daily.

Once people stopped treating it like a genius colleague, their prompts improved immediately.

How does a large language model actually work?

Here's the 60-second version. A large language model (LLM) is trained on enormous amounts of text, roughly the equivalent of millions of books. During training, it learns patterns: which words tend to follow which other words, in what contexts, with what tone. When you give it a prompt, it uses those patterns to predict what should come next, one token at a time.

Think of it like this: if someone says "The capital of France is...", you'd predict "Paris" without needing to reason about geography. You've seen that pattern thousands of times. LLMs work the same way, just across billions of patterns instead of thousands.

Here's what actually happens each time you hit enter:

1. You write a prompt -- your question or instruction in plain language.
2. The model breaks it into tokens -- small chunks of text (roughly 3/4 of a word each).
3. It processes the tokens -- running them through layers of pattern-matching (neural network layers, technically).
4. It predicts the next token -- picks the most probable next piece of text.
5. It repeats step 4 -- generating one token at a time until it decides it's done.

[IMAGE: Five-step LLM flow diagram showing prompt to tokens to processing to prediction to output]

• Type: diagram
• Filename: llm-five-step-flow.png
• Alt text: Diagram showing the five steps of how an LLM processes a prompt and generates a response
• Caption: Every AI response follows this same five-step loop

What are tokens and why should you care?

Tokens are the atomic unit of AI. Models don't read words the way you do. They chop text into tokens, which are roughly three-quarters of a word on average. The word "unbelievable" becomes multiple tokens. The word "cat" is one.

Why does this matter for business use? Three reasons:

• Cost. You pay per token on most platforms. GPT-4o costs about $2.50 per million input tokens and $10 per million output tokens as of early 2026 (OpenAI). A 1,000-word document is roughly 1,300 tokens.
• Context window. Every model has a maximum number of tokens it can handle in one conversation. GPT-4o handles 128,000 tokens. Claude 3.5 Sonnet handles 200,000. Go over the limit and the model starts forgetting earlier parts of your conversation.
• Precision. Knowing how tokenisation works helps you understand why AI sometimes stumbles on unusual words, code, or non-English text.

You can see exactly how text gets tokenised using OpenAI's free tokenizer tool. Paste in a paragraph from your website and watch it split into coloured chunks. It's surprisingly satisfying.

What can AI do well, and where does it fall flat?

This is the part most AI guides skip, and it's the part that saves you the most time. AI is genuinely brilliant at some tasks and genuinely terrible at others. Knowing which is which stops you from trying to force a square peg into a round hole.

AI is strong at:

• Drafting and editing text (emails, reports, social posts)
• Summarising long documents
• Brainstorming and ideation
• Translating between languages
• Reformatting data (CSV to JSON, messy notes to clean tables)
• Explaining complex topics simply

AI is weak at:

• Maths (it guesses rather than calculates, though tool-use is closing this gap)
• Knowing what happened after its training cutoff
• Citing real sources accurately
• Making subjective business decisions
• Anything requiring real-world sensory experience
• Consistently following very long, complex instructions

When I ran this live with a group of 20 founders, roughly 60% were trying to use AI for tasks in the "weak" column. One bloke was asking ChatGPT to calculate his quarterly tax obligations. Please don't do that. Use a spreadsheet.

Why does AI make things up?

Hallucination is a fancy word for "the model confidently generated text that is factually wrong." It happens because the model is predicting plausible-sounding text, not looking things up in a database.

A 2024 study by Vectara found that even top-tier models hallucinate between 3% and 27% of the time depending on the task (Vectara Hallucination Index). That's improved since, but it hasn't hit zero and probably won't for years.

Three rules for managing hallucinations:

1. Never trust a specific claim without checking. If AI gives you a statistic, a date, or a name, verify it.
2. Ask for sources. Models will sometimes cite real papers and sometimes invent fake ones. Check the URLs.
3. Use AI for structure, you for facts. Let AI draft the framework of a competitor analysis or report, then fill in verified data yourself.

Which model should you use for which job?

Not all models are equal, and the most expensive one isn't always the best choice. Here's a practical comparison as of early 2026:

Model	Best for	Context window	Relative cost	Speed
GPT-4o (OpenAI)	General business tasks, writing, analysis	128K tokens	Medium	Fast
Claude 3.5 Sonnet (Anthropic)	Long documents, nuanced writing, code	200K tokens	Medium	Fast
Gemini 1.5 Pro (Google DeepMind)	Multimodal (text + images), large context	1M tokens	Medium	Medium
GPT-4o Mini	Quick drafts, simple Q&A, high volume	128K tokens	Low	Very fast
Claude 3 Haiku	Fast classification, simple summaries	200K tokens	Low	Very fast

The rule of thumb: Start with a cheaper, faster model. Move up only when the output quality isn't good enough. In our workshops, 70% of common business tasks worked perfectly well with the smaller models, saving roughly 80% on costs.

[IMAGE: Decision tree for choosing an AI model based on task type]

• Type: diagram
• Filename: model-decision-tree.png
• Alt text: A decision tree diagram helping users choose between GPT-4o, Claude Sonnet, Gemini Pro, and smaller models based on task requirements
• Caption: Start cheap and fast, move up only when you need to

FAQ

---

Next up: The Difference Between 'Meh' and 'Wow' -- learn the five-part prompting framework that turns vague requests into sharp results.

This is Post 1 of 7 in the AI for Business free course.