You don't need to understand how AI works.
You need to know when you can trust it.
This course turns a language model into a tool that saves you time in your daily work — without losing customer data along the way or putting a made-up figure into a quote.
You work at your own pace. Your progress is stored in this browser — you can stop at any time and continue later. No data is sent to us.
Four concrete skills
- Judge which of your tasks are a good fit for AI — and which are not.
- Write an instruction so that the answer is usable on the first try.
- Spot fabricated details before they reach a customer.
- Decide what you may type into an AI window — and what you never should.
A machine that keeps on writing
The most common misconception is also the most expensive one: the assumption that a language model looks things up.
A language model like ChatGPT or Claude has read an enormous amount of text and learned from it which words typically follow one another. When you ask something, it does not look up the answer — it composes the most likely continuation of your question.
That is why a wrong answer sounds exactly as convincing as a right one. The model cannot tell the difference.
Almost everything you need to keep in mind follows from this. A model is strong when the job is language: rephrasing, summarizing, structuring, translating, generating ideas. It is weak when the job is facts it doesn't have in front of it: prices, dates, names, legal articles, your figures from last week.
Strong
Processing text you provide. Turning ten lines of notes into a clean file memo. Making a draft more polite.
Weak
Supplying details it doesn't have. Ask it for your supplier's price, the latest court ruling, or Mr. Meier's appointment, and it will invent something plausible.
What is true about language models?
Choose the one statement that is correct.
Choosing the right tasks
The difference between enthusiasm and disappointment rarely comes down to the model. It comes down to the task you give it.
A simple rule of thumb: you supply the knowledge, you ask for the language. Anything the model is supposed to shape out of your own text works reliably. Anything that would require knowledge only you or your systems have is risky.
- Well suited: summarizing, rephrasing, adding structure, gathering ideas, first drafts, translating, explaining a technical text in plain words.
- Risky: calculations with real amounts, legal advice, current events, statements about people, anything that goes out unchecked.
Six tasks from everyday office work
Decide for each task: well suited or risky?
Four building blocks for every instruction
"Write me an email to a customer" produces an email that fits no one. Four extra sentences change the result completely.
A usable instruction states a role, a task, context, and a format. The role sets the tone. The task says what should come out. The context supplies the knowledge the model lacks. The format determines length and structure — and saves you the rework.
Build the instruction
Scenario: a customer complains about a late delivery. Toggle the building blocks on and watch the instruction grow.
Two habits help on top of that: provide examples ("this is how our emails usually sound: …") and follow up instead of starting over — "shorter", "less formal", "without the last paragraph" get you there faster than a second attempt.
Where the answer is made up
Fabricated details don't come with a warning color. They sit in the middle of the smoothest paragraph — and most of the time they are numbers, names, or references.
Don't check the whole answer — check four specific kinds of details: numbers (prices, deadlines, percentages), proper names (companies, people, products), references (legal articles, standards, sources), and dates. Everything else is language — and language is what the model does well.
What do you need to check before this goes out?
An AI produced this answer to a question about the warranty period. Mark the three passages you must not adopt unchecked.
The practical test: with every answer, ask yourself where the model could have gotten that detail. Was it in your question? Then it is probably reproduced correctly. Was it not? Then it is a guess — even if a source is cited.
The line between a useful tool and a data breach
Everything you type into an AI window usually leaves your company. With free services, it may additionally be used for training.
The rule is simpler than it sounds: treat the input field like a postcard to an outside service provider. Whatever you wouldn't write on that postcard doesn't belong in the field either. Personal data, health information, salary lists, passwords, entire contracts with names and terms — all sensitive.
Most of the time you don't even need those details: anonymize the text before pasting it in. "Customer A", "Company B", "Amount X" are perfectly sufficient for the model — the help with wording is exactly the same.
May this go into the input field?
Rate each case: harmless, only anonymized, or not at all.
If you regularly need to work with real customer data, that is not a prohibition — it is a question of tooling: business accounts that don't train on your data, or a model hosted in-house, solve exactly this problem. Talk to us about it.
Turning knowledge into habit
The most common reason AI delivers nothing in a company: after the training, nobody thinks of using it at the right moment.
Pick one single recurring task from your week — the meeting report, the standard reply to a frequent request, the translation into French. Do that one task with AI for two weeks. Only then add the second.
Your next steps
This list is stored in your browser. You can come back later and tick items off.
Four questions to finish
Your reference sheet
The essentials on one page — print it and pin it next to your screen.
Paste in the text the conversation should be about. Never ask for details only you could know.
Four sentences instead of one. Who should answer, what should come out, what is the situation, how long and in what structure.
Always verify these four kinds. The rest is language, and language is what the model does well.
Whatever you wouldn't write on a postcard to an outside service provider, don't type into the input field.
"Customer A", "Amount X" — the help with wording stays just as good, the risk disappears.
"Shorter", "less formal", "without the last paragraph". The conversation is the tool, not the first attempt.
You remain responsible for everything that carries your name.
When in doubt: don't type it in, don't send it out unchecked, just ask.