Identify AI concepts and capabilities
Every scenario here gets one name, drawn from one of three vocabularies
A retailer wants an app that photographs a paper receipt and writes the total into an expense system. This domain never asks you to build that app; it asks you to name what you are looking at, and the naming happens in three vocabularies, one per page: the responsible AI principles a solution owes the people it affects, the model components and configurations that carry the work, and the AI workload families that describe the job itself. The receipt app is an information-extraction workload, and the trap is that its surface details pull elsewhere, because a photograph sounds like computer vision and "reads the total" sounds like text analysis. Decide which vocabulary the answer has to come from before you compare the names inside it, since a right-sounding name from the wrong vocabulary is how most of these questions are lost.
Three vocabularies, three pages: obligations, then the model, then the workload
Read the three pages in the order the objectives list them. "Describe principles of responsible AI" covers the six named principles and the scenario cues that keep them apart, and it applies to every solution no matter what the solution does. "Identify AI model components and configurations" opens the model itself: how a language model builds its answer one token at a time, which model class fits which requirement, what turns a catalog entry into something your code can call, and which request parameters change the result. "Identify AI workloads" is the routing catalog: given a described outcome, which of the six families the work belongs to — generative AI, agentic AI (acting through tools), text analysis, speech, computer vision, information extraction, which Microsoft's own list presents as five entries — and which Azure service usually serves it. The three pages together carry 40–45% of the exam, and the other domain, implementing AI solutions with Microsoft Foundry, carries 55–60% and is where the code lives.
When two answers both fit, go with what the solution must hand back
One instinct settles more questions in this domain than any other: read the requirement for the result that has to come out, not for the material going in or the technology the scenario mentions in passing. It picks the workload family, because "list the products on this shelf" and "create a promotional picture of this shelf" start from the same photograph and land in different families. It picks the model, because a requirement for a similarity score needs an embedding model while a requirement for a written reply needs a chat model. It even sorts the responsible AI principles, where the question becomes what the solution owes and to whom: telling users they are dealing with AI is transparency, while keeping their personal data out of prompts and logs is privacy and security.
The three vocabularies this domain names, and where each is taught
| Vocabulary | The question it answers | What you end up naming | Drill into |
|---|---|---|---|
| Responsible AI principles | What does this solution owe the people it affects? | One of the six principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, accountability | Describe principles of responsible AI |
| Model components and configurations | Which model carries the job, reached how, and tuned with what? | A model class, an access path such as a deployment, and the request parameters that shape a response | Identify AI model components and configurations |
| AI workloads | What kind of job is this, and what must it hand back? | One workload family, and the Azure service that usually serves it | Identify AI workloads |