AI for business: what a language model actually does (and what it does not)

The situation

You have seen the demo. Someone types a question and the model answers in three seconds with text that sounds perfect. The next day a customer messages you on WhatsApp and you think: why isn't the AI answering this for me? Then someone on your team tries it, and the model invents a price you do not have and a warranty you do not offer.

Both experiences are real. To know when to rely on AI and when not to, you need to understand one thing: what a language model actually does.

What a language model does

A large language model (LLM) does one thing: it takes text and predicts the text most likely to follow. That is all. It learned this from enormous amounts of text, and because it has seen so much, its predictions look like thinking. Three consequences follow, and you need to keep all of them in mind:

• It has no memory unless you give it one. Every conversation starts from zero. If it knows your customer's name, it is because someone put it in the context, not because it "remembered".
• It knows nothing about your business. Your opening hours, prices, stock, return policy — none of it is in the model. It exists only if you hand it over with every question.
• It is just as confident when it is wrong. A correct answer and an invented one are produced by the same mechanism. Tone is not a signal of truth.

What it is good at, and what it is not

What it is good at in a business

The model is excellent at text tasks where you already have the material and want speed:

• Drafting: emails, product descriptions, review replies, the first version of any text.
• Classifying: "is this message a complaint, a quote request or spam?" At hundreds of messages a day, that is hours.
• Extracting: from the invoice that arrived by email, pull out the supplier, amount and due date into a table.
• Summarising: a 40-message thread becomes five lines with the decisions taken.
• Answering from your own documents: give it the right text (contract, product sheet, procedure) and it answers from that, not from imagination. That is called RAG and it has its own lesson.

What it is bad at

• Facts you did not give it. Asked about your opening hours without having them, it will produce plausible hours. Plausible, not real.
• Arithmetic at scale. It can add two numbers. Do not let it compute VAT on 3,000 spreadsheet rows — that is what formulas and code are for, and they never get it wrong.
• Guarantees. A model cannot promise it will always be right. A well-built system around it can reduce errors and say "I don't know", but zero does not exist.
• Decisions that carry responsibility. Who answers when the model promised a discount? You do. So the decision stays with a human.

Chat, API or system

These are three different things, and confusing them costs money:

Chat is a personal tool. A system is a product. Between them is not a step but a project: who supplies the data, who verifies, what happens when the model is wrong.

The three questions before any AI project

1. Where is the data? If the answer is "in Maria's head" or "in a PDF from 2022", the project starts with documentation, not with AI.
2. Who approves? What may the model do alone (propose an answer) and what needs a human (send a quote, issue a refund). Draw the line before, not after the first incident.
3. How do we measure? Hours saved per week, share of questions resolved without a human, errors reported. Without a number before, you cannot say afterwards whether it was worth it.

What it costs, realistically

The model itself (per call) has become steadily cheaper and for most SMEs it is the smallest line on the bill. The real costs are: preparing the data (clean, current documents), building the system, testing with real questions, and maintenance — because your prices change, models change, and someone has to check every month that the answers are still right. Budget AI like a junior employee: cheap per hour, expensive if nobody supervises.