How AI really works

Self-check

Your score stays on your own device.

1. A colleague says: 'A newer, bigger model will know our contracts.' What is true?

   • No. Only a model trained on your contracts can
   • Yes. Bigger models have read more, your contracts too
   • No. It only knows them once it gets them as a source
   • Yes, as long as you tell it not to make things up

2. Your team starts using an AI assistant for customer questions. Which measure helps most to catch a made-up answer?

   • Let the assistant give short answers only
   • Each answer shows its source document
   • Ask the assistant how sure it is each time
   • Ask the same question twice and compare

3. A supplier says: 'With RAG, our assistant gives the right answer every time.' What is true?

   • True, because RAG retrains the model on your files
   • Not true. RAG makes answers less reliable than before
   • True, because it reads your own documents first
   • Not quite. It can still miss; the source shows that

4. Someone in sales asks the document assistant about a personnel file they cannot open on the network drive. What should happen?

   • No answer from it: the existing rights apply
   • An answer, after a warning that it is confidential
   • A full answer, because the AI can read every folder
   • A summary only, because a summary does no harm

5. You let an agent prepare draft orders in your ERP. Which rights does it get?

   • Admin rights, so it does not get stuck halfway
   • The same rights as your IT department has
   • No more than the person doing it now
   • Read-only: an agent must not change anything

6. After three months the agent's standard status emails are right every time. What is a sensible next step?

   • Keep every email on approval: trust cannot grow
   • Give it more tasks first, and decide on approval after that
   • Remove all approval steps now, since it has proven itself
   • Let standard emails go; the exceptions stay with the planner

7. A supplier says its fine-tuned model is 'much better' for you. What do you ask?

   • Is it a newer model than the one we use now?
   • What did it score on examples kept out of training?
   • Do your users say it feels better in their daily work?
   • How long, and on how much data, was it trained?

8. You run an open model on your own server in the EU. What is still left to arrange?

   • Who has access, and how long chats are kept
   • Only a contract with the maker of the model
   • Nothing: a model on your own server is safe by default
   • Retraining it on your data so it stays private