Most people describe AI as a machine that thinks. It isn’t. A large language model is a prediction engine, and once you see it that way, every decision about where to use it in your business gets easier.
What AI actually is
Ask a model what the capital of France is and it does not look up an answer. It predicts the next most likely word, then the next, until the sentence is finished. “The capital of France is” reliably produces “Paris” because that sequence appears everywhere in the text it was trained on.
The unit it predicts in is a token, roughly three-quarters of a word. Tokens are the alphabet of a language model, and everything it reads or writes is measured in them.
That single mechanic explains both the power and the failure modes. It is why models are fluent in any domain you name, and why they will state something false with exactly the same confidence as something true.
What actually changed since 2023
GPT-3.5 shipped as a text-only model with a hard knowledge cutoff. Ask about anything after September 2022 and it either guessed or refused. Three things have changed since, and each one moved AI from a novelty to something you can staff a workflow with.
- Multi-modal input. The same model that reads text now reads images, audio and video. A photo of a roof is a valid input to a workflow.
- Tools instead of memory. Models now recognise what they don’t know and go and search for it. A model trained three months ago can still answer a question about yesterday.
- Looping instead of answering. An agent keeps checking and reasoning with itself until a task is finished, rather than returning one response and stopping.
Chatbot or AI employee?
That third change is the whole game. Tell an agent to fetch your emails and draft replies and it works through every one, going back and forth with itself. That back-and-forth is what we mean by reasoning. It then executes the tools that write the drafts.
To a language model, searching the internet and sending an email are the same kind of action. Once that is true, the difference between a chatbot and an employee is just how long you let it run.
What this means for your business
Part of an organisation is run by people. A growing part is run by agents. The question is no longer whether AI can do a job. It is which jobs you hand over, and who picks the work back up on the day the agent gets one wrong.
Start with the vocabulary. Tokens and context windows next, then agents and tool calls. You cannot scope work you cannot name.
