AI-901 identify AI concepts and capabilities explained
Identify AI concepts and capabilities is worth 40–45% of AI-901 and it is the half you can genuinely learn from reading. It has three parts: the principles of responsible AI, how models and their configuration work, and recognising which AI workload a scenario calls for. None of it requires you to build anything — but it does require precision, because the questions are built on distinctions that sound similar until you look closely.
Responsible AI: six principles, and how they are tested
Microsoft names six. Learn them as a list and then learn to tell them apart:
| Principle | The question it answers |
|---|---|
| Fairness | Does the system treat comparable people comparably? |
| Reliability and safety | Does it behave consistently and fail safely? |
| Privacy and security | Is personal data protected and access controlled? |
| Inclusiveness | Can people of all abilities and backgrounds use it? |
| Transparency | Can people understand how it reached a decision? |
| Accountability | Is it clear who is answerable for the system? |
The exam almost never asks “name the six principles”. It gives a short scenario and asks which principle it primarily concerns. The trap is that several usually apply. Read for the primary harm described: an unequal outcome between groups is fairness; a user who cannot tell why they were refused is transparency; an unlogged system with no owner is accountability; a captioning gap for deaf users is inclusiveness.
Model components and configuration
Three things sit here.
How generative models work. At the level tested: a model predicts likely continuations of text, works in tokens rather than words, and has a limited context window. You should be comfortable with what a prompt is and why the same prompt can give different answers.
Choosing a model by capability. Large language models for broad reasoning, small language models where latency and cost dominate, multimodal models where images or audio are part of the input. The exam gives you a constraint and expects the cheapest model that still satisfies it.
Deployment options and parameters. What a deployment is, and what the common parameters do — temperature being the one most likely to appear, where lower means more predictable output and higher means more varied.
Recognising AI workloads
The final part is vocabulary matching, and it is the easiest section of the exam to secure marks in:
- Generative and agentic AI — producing content; an agent additionally uses tools and takes steps.
- Text analysis — keyword extraction, entity detection, sentiment analysis, summarisation.
- Speech — recognition, which is speech to text, and synthesis, which is text to speech.
- Computer vision — analysing images, plus image-generation models.
- Information extraction — pulling structured content out of text, images, audio and video.
Know which technique produces which shape of output. Sentiment gives you a score or label. Entity detection gives you names of things. Summarisation gives you shorter prose. Extraction gives you fields. A question that asks for a field is never answered by summarisation.
Sample questions
Question 1. A recruitment tool produces a shortlist but cannot show which factors led to a candidate being excluded, and applicants ask why they were rejected. Which responsible AI principle is most directly at stake?
- A. Fairness
- B. Privacy and security
- C. Transparency
- D. Reliability and safety
Show answer
Answer: C
Transparency is about people being able to understand how a system reached its decision. Fairness would be the answer if the concern were an unequal outcome between groups; privacy concerns data handling; reliability concerns consistent behaviour. Here the described problem is that the reasoning is not visible.
Question 2. A mobile app must classify short product reviews as positive or negative. It runs at very high volume, the task is simple, and per-call cost matters more than nuance. Which model choice is most appropriate?
- A. A large language model
- B. A small language model
- C. A multimodal model
- D. An image-generation model
Show answer
Answer: B
A small language model gives the lowest cost and latency, which is what the requirement optimises for on a simple, well-defined task. A large model adds reasoning the task does not need, a multimodal model adds image handling that is irrelevant, and an image-generation model does not classify text at all.
Question 3. A logistics firm wants to turn recorded driver phone calls into a searchable list of mentioned locations and delivery reference numbers. Which two capabilities does the workload need?
- A. Speech recognition and entity detection
- B. Speech synthesis and summarisation
- C. Image generation and entity detection
- D. Speech recognition and image captioning
Show answer
Answer: A
The audio must first become text through speech recognition, then entity detection identifies the locations and reference numbers within that text. Speech synthesis produces spoken output rather than consuming it, image generation is unrelated, and summarisation would shorten the call rather than list specific entities.
What to practise
Write the six principles on one sheet with a one-line scenario beside each, in your own words. Then, for each of the five workload types, write down what goes in and what comes out. That single page covers most of what this domain can ask, and it is the cheapest revision in the whole exam.