The responsible AI block in AI-901: from six principles to four questions

The responsible AI block in AI-900 and AI-901 covers six official principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. For the exam, you memorise them, match them to the right question, and pass at 700. But what do you actually do with them on Monday, when you hand Copilot a client email to summarise, point a Foundry agent at a dataset, or use ChatGPT to classify something?

Why six principles don’t survive in practice

Six items is fine for an exam checklist. As a working checklist, it’s different: before you use an AI tool, you don’t have space to walk through six separate principles one by one. What you want are a few questions fast enough to ask before you go live, and concrete enough that you know what to do when the answer is “not really.”

The four questions below are my didactic simplification of the six official Microsoft principles — not a new Microsoft framework. The official framework stays at six, and you need all six for a 700 on AI-901. The cluster to four is a practical shortcut so that you actually ask them.

The four questions

Question 1 — Who does this work well for, and who does it work less well for?

This covers fairness and inclusiveness. An AI tool that pre-sorts leads or screens applications can be biased if the training data wasn’t representative of all the groups the system will judge. Before each use, ask: are all customer groups, regions, or roles fairly represented in what this system learned? And are there people it will consistently underserve?

Question 2 — What does this do when the input is strange, or something goes wrong?

This covers reliability and safety. Say Copilot summarises a client report that gets forwarded without review. What happens if the source data is incomplete, or the summary draws a conclusion the data doesn’t support? No tool performs evenly on unexpected input. Do you know what this one does — and how you would catch it?

Question 3 — What data goes in, and where does it go?

This covers privacy and security. The moment you put customer data, personnel records, or sensitive business information into a prompt, the relevant question is: where does this go? Does the system comply with applicable privacy law? Has the person or client given consent? This is the question people skip when they are in a hurry — and the one that causes the most legal and reputational damage afterward.

Question 4 — Who can explain what came out, and who is responsible?

This covers transparency and accountability. A Foundry agent that independently completes a step in an internal process: who explains what happened if something goes wrong? Who owns that decision? If you cannot answer that now, the system is not ready for production.

Can you make the mapping yourself? The six official principles come by one at a time below; click the question each one belongs to.

Accountability is not a question

Of the four, the last one is not really a question you ask before each use — it is an attitude you take. Accountability is not a checkbox item; it is the way you take the other three seriously and follow through. For the AI-901 exam, it is a principle to describe. In practice, it is the posture the other five rest on: if no one takes responsibility for asking the other questions, the five principles are well-written paperwork.

Back to AI-901

The four questions are an operational filter, not a replacement for the official framework. When an answer gives you pause, you know exactly which of the six Microsoft principles to dig into further. On the exam, there are six, you need all six, and they are all correct.

Candidates who read the responsible AI block in an evening and remember it for the exam will likely get their 700. Candidates who also ask the questions before every AI use are the professionals who will have the fewest problems when something goes wrong — and who know what to explain when it does.

More on AI-900 and AI-901 at trainerbjorn.nl, including the current course catalogue and training dates.