AI-103 generative AI and agentic solutions explained
Implement generative AI and agentic solutions is the largest domain on AI-103 at 30–35% of the score, and it is the reason the exam exists. Where the older track spread attention across individual Azure AI services, this domain concentrates on one thing: building an application or an agent on Microsoft Foundry that uses a model, grounds its answers in your data, calls tools, and keeps working once real users touch it.
Building generative applications
The starting point is deploying and consuming models — large, small, code and multimodal — and connecting an application to a Foundry project through the SDKs and connectors. Expect the exam to assume Python.
Two things carry disproportionate weight:
Retrieval-augmented generation. RAG appears across the whole exam, not just here. You need to understand the shape of it: content is chunked and indexed, a query retrieves the relevant chunks, and those chunks go into the prompt so the model answers from them. The consequence you must be able to reason about is that RAG keeps answers current without retraining, and that when an answer is wrong the fault is usually retrieval, not the model.
Evaluation. Detecting fabrications, and measuring relevance, quality and safety. Know what the common evaluators actually measure, because the exam tests the distinction:
| Evaluator | Answers the question |
|---|---|
| Groundedness | Is the answer supported by the retrieved content? |
| Relevance | Does the answer address what was asked? |
| Coherence | Does the answer hang together logically? |
| Fluency | Is the language well formed? |
A fluent, coherent, irrelevant answer is a real failure mode, and the exam likes it.
Beyond single calls, you are expected to design workflows: tool-augmented flows and multistep reasoning pipelines rather than one prompt doing everything.
Building agents
This is the part that most distinguishes AI-103. An agent, as the exam frames it, is a model given a role, a goal, a way to track conversation, and a set of tool schemas it may call.
You should be comfortable with:
- Tools. APIs, knowledge stores, search, content understanding and custom functions. Given a requirement, pick the tool — and know that a function call is the answer when the agent needs to do something, while retrieval is the answer when it needs to know something.
- Memory. Conversation memory versus retrieved knowledge. They solve different problems and the exam will offer you the wrong one.
- Multi-agent orchestration. When one agent delegates to specialised agents, and how they are coordinated.
- Safeguards. Autonomous and semi-autonomous workflows need approval flows and constraints. Any scenario involving an irreversible action — a payment, a deletion, an email to a customer — wants a human approval step in the answer.
- Monitoring and error analysis. Evaluating agent behaviour after deployment, not just before.
Optimising and operationalising
The third group is about what you do once it runs: tuning generation behaviour through prompt engineering and model parameters, reflection and self-critique loops, observability through tracing, token analytics, safety signals and latency breakdowns, and orchestrating several models or a hybrid of a model and a rules engine.
A useful instinct for the exam: when a scenario mixes a deterministic requirement with a fuzzy one, the expected design is hybrid. Rules handle what must be exact; the model handles what must be flexible.
Sample questions
Question 1. An agent books meeting rooms. It must be able to cancel a booking, but a cancellation can never happen without a person confirming it. What should you build?
- A. Remove the cancellation tool and tell users to cancel manually
- B. Give the agent the cancellation tool and instruct it in the system prompt to be careful
- C. Give the agent the cancellation tool behind an approval flow that requires human confirmation
- D. Give the agent read-only access to the booking system
Show answer
Answer: C
A cancellation is an irreversible action, so the agent needs a tool it can call plus an approval step that a human completes before the call executes. Removing the tool makes the agent unable to do its job, and describing the action in the prompt does not enforce anything.
Question 2. A RAG assistant over product manuals answers fluently but often cites the wrong product model. Retrieval returns five chunks per query. What is the most likely fix?
- A. Improve chunking and switch to hybrid or semantic search so the correct chunks are retrieved
- B. Deploy a larger model
- C. Lower the temperature to zero
- D. Add an instruction to the system prompt telling the model not to make mistakes
Show answer
Answer: A
Fluent but wrong-source answers point at retrieval, not generation. Improving the index and query strategy so the right chunks are retrieved addresses the cause. A larger model, a lower temperature or a longer prompt all still receive the wrong context and will keep producing the same error confidently.
Question 3. You must give stakeholders a number that shows how often your assistant's answers are actually supported by the documents it retrieved. Which evaluator do you report?
- A. Fluency
- B. Coherence
- C. Relevance
- D. Groundedness
Show answer
Answer: D
Groundedness measures whether the response is supported by the retrieved source content, which is exactly the claim being made. Relevance only checks that the answer is on topic, and fluency and coherence measure language quality.
What to practise
Build one agent end to end and give it at least two tools, one of which is a custom Python function. Ground it with RAG over a handful of your own documents, then deliberately break retrieval — bad chunk size, no semantic search — and watch the answers degrade while staying perfectly fluent. Run an evaluation over the good and the broken version and compare the groundedness scores. That single exercise covers a large share of this domain.