Module 1Academy

AI basics for teams

What a language model does, where it goes wrong, and what may go into it. For the whole team, not only the technologists.

AI basics for teams is a half-day course for anyone who will work with AI. The team learns how a language model arrives at an answer, why it can be convincingly wrong, and which data does not belong in it. After the module, each participant writes a prompt for their own work and checks the result.

After this module the team can
  • 01The team describes in its own words how a language model composes an answer, and why that answer is sometimes wrong
  • 02Each participant writes a prompt for their own work and checks the answer with a fixed control step
  • 03The team names the data that does not belong in an AI tool, by category
  • 04The team makes a list of its own tasks AI can help with, and of tasks it cannot
Agenda
  1. 01

    What a language model is

    Prediction, not knowledge. We show why a model completes a sentence the way it probably sounds, and what that means for facts, amounts and names.

  2. 02

    Where it goes wrong

    Invented sources, wrong figures and an overconfident tone. Each participant tries to provoke an error and learns the control step that catches it.

  3. 03

    Which data may go in

    Personal data, client files, contracts and trade secrets: per category, with the question of what the business version of the tool does with it.

  4. 04

    Practice on your own work

    Each participant brings one task. We write the prompt together, read the answer critically and note what is missing.

Why this is the first module

Most problems with AI in a team start not with the technology but with expectations. Someone asks a language model for a figure, gets a figure that reads like an answer, and the figure goes out of the door. That mistake is not dramatic when someone catches it, but nobody catches it when the team does not know where the model is often wrong.

That is why the Academy starts here, before the conversation about tools or automation. A team that knows how a language model works chooses better prompts, checks faster and asks the vendor the right questions. The team needs that basis before it runs a system that answers without a person in the loop.

How a language model arrives at an answer

A language model estimates, word by word, which word most likely follows what is already there. That is why it reads smoothly and sounds certain. It has no database of facts to look things up in; it learned patterns from an enormous amount of text. That makes it strong at phrasing, summarising and structuring, and weak at precise facts, figures and sources it does not have in front of it.

For a team that means a simple rule: let the model work with material you supply yourself, and check everything that later goes into a decision, an invoice or a letter to a client. That rule is the core of the module, and we practise it until it becomes automatic.

Where it goes wrong, and how to catch it

Three mistakes are common. The model invents a source that looks right but does not exist. It calculates an amount wrongly while the reasoning reads neatly. And it answers in a confident tone even when the question was unclear. The technical literature calls these hallucinations.

The control step is the same in every exercise: take the claims out of the answer, look them up in a source you trust, and stop when a claim cannot be found. For figures the extra rule is that the calculation must be visible. A team that has done this twenty times gets fast at it.

Which data does not belong in it

The largest risk is not the error in the answer but the information that disappears into the prompt. Personal data of clients or staff, files with sensitive content, negotiation documents and trade secrets do not belong in a tool by default. Whether they may depends on the account you use and what the vendor promises by contract, which differs per tool and per subscription.

We go through what you can do per category: anonymise, use a business account with a data processing agreement, or do not give the task to a language model at all. The details of the AI regulation and the GDPR come back in module five. Here the team learns the question it must always ask: what happens to this text, and who has put that in writing?

What the team does afterwards

The module ends with two lists the team takes away. The first holds tasks AI can support well, with the control step next to each. The second holds tasks AI should not do, such as a decision about a person or a legal judgement without a human check. Those two lists are the starting point for module three, where we decide which work is worth giving to a system.

Frequently asked questions

  • Is this training about one specific tool?

    No. We use the tool your team already has or is choosing, but the module is about working with language models in general. That knowledge stays when the tool changes.

  • Do participants need to prepare anything?

    Yes, one task from their own work that comes back every week or every day. Without that task the exercise is too abstract.

  • Is the module also for managers?

    Yes, and they do the same exercises. A decision about AI is only sound when the person deciding has seen the errors themselves.

Which team needs this module?

For half day €2,450 excl. VAT. In the conversation we say whether the module fits, or whether a build day or the adoption programme is better.