This lesson introduces nothing new. It puts together what Nordlicht Reisen decided in lessons 1.4 and 2.1 to 2.9, and shows where to go from here. There is no code and there are no portal steps here.
The solution at a glance
The case keeps the assignment from lesson 1.4: one input, one suitable service. The table adds what has come in since.
| Input | Service (lesson) | Model deployment? | Human step |
|---|---|---|---|
| Reviews as text | Azure Language, sentiment analysis (2.6) | no, a resource of its own according to the quickstarts | customer service reads the negative reviews |
Call as audio, anruf-hotline.wav |
Azure Speech, speech recognition (2.7); alternative: Content Understanding (2.9) | Speech no, Content Understanding yes | customer service reads transcript and summary |
Room photo, zimmerfoto.jpg |
vision-enabled model in the prompt (2.8) | yes, reise-chat, if the model understands images |
a person judges the finding |
Invoice, rechnung-seeblick.pdf |
Content Understanding, prebuilt-invoice (2.9) |
yes, chat and embedding model | staff check fields with low confidence |
| Questions from customer service | prompt agent reise-helfer (2.4, 2.5) |
yes, reise-chat |
a person checks every answer before it goes to customers |
Across these lie the foundations: foundry-nordlicht and nordlicht-ki (2.1), good prompts (2.2), the chat client (2.3). Sign-in works with Entra ID and the role Foundry User, not with keys in the code. Learn's examples in 2.6, 2.7 and 2.9 use keys; for production, Learn recommends Entra ID there.
What the lessons changed about 1.4
No assignment of the case has changed. Two sentences of the overview have shifted on purpose:
- Lesson 1.4 named Azure Vision for image processing, alongside models with image input. Lesson 2.8 says: Learn marks Image Analysis as deprecated and announces its retirement. For new applications, the case chooses the model with image input, as already in 1.4.
- For calls, Azure Speech remains the case's choice. Lesson 2.9 shows that Content Understanding evaluates calls as well. That is an alternative, not a switch.
Responsible AI at the places in the pilot
The six principles (lesson 1.1) are not in a slide, but in decisions:
| Place in the pilot | Principle | Decision |
|---|---|---|
| A person checks every answer and every finding | Accountability | Ms. Berg is named responsible as head of customer service |
| Sentiment analysis with a threshold, tested only on real data | Reliability and safety | the threshold is open and is measured, not estimated |
| Dialects and languages, English too | Fairness and inclusiveness | specify the language, do not accept English as the default (2.6) |
| Recordings and guest photos | Privacy and security | consent and legal questions with legal counsel, no public addresses for photos (2.7, 2.8) |
| Notice that a helper is involved; labeling of generated images | Transparency | no generated images for customers in the pilot; labeling stays open (2.8) |
What is not implemented in the pilot
The case leaves out:
- Decisions by the model about refunds, cancellations or contract questions.
- Answers without a check and speech synthesis to customers.
- Image generation for customers.
- Recognizing or naming people in photos.
- Recording and evaluating calls, as long as legal counsel has not settled that.
- Preview features in production. That is a decision of the case, not a requirement of Learn: What Learn lists as a preview is used only after general availability.
What stays open
Availability of the models in the region (also reise-chat with images and with Content Understanding); how long sentiment analysis and summarization are supported as "legacy"; the thresholds for "negative" and for low confidence; the best analyzer for the hotel invoice; whether audio models and image solutions are still a preview; whether the Foundry Tools keep data in the EU data zone (for models: lesson 1.3; not read for the tools); the questions to legal counsel on recording and labeling.
What you can do now
According to the study guide, the AI-901 exam has two areas. Read the lists as taking stock, not as a promise to pass the exam:
- AI concepts and capabilities: name the six principles; explain how generative models work; weigh models and deployment types; tell the workloads text analysis, speech, computer vision, information extraction and generative and agentic AI apart and assign scenarios to them.
- AI solutions with Microsoft Foundry: deploy and use a model in the portal; write prompts; build and call a chat client and an agent; bring text analysis, speech, images and information extraction into an application with the services from Learn.
The exam assumes Python syntax; this course showed C# (note in lesson 2.3). Anyone who wants to take it reads the Python tabs of the same Learn pages. According to the study guide, Microsoft offers a practice assessment in the AI Skills Navigator; it is an offer from Microsoft, not from us.
Where to go from here
- AI track (the course "AI track: AI engineering with .NET"): the same material in depth and in C#. Suitable entry points: T1.4 "Responsible AI and the EU AI Act", T2.1 "The IChatClient", T4.1 "Embeddings" and T5.1 "Grounding answers", T6.1 "Prompt injection", T7.1 "The first agent" (there with the Microsoft Agent Framework instead of the Foundry Agent Service) and T10.1 "Foundry projects and deployments". The AI track does not teach text analysis, speech, image analysis and Content Understanding; they are in this course.
- AZ-900 compact: Anyone who does not yet know resource group, subscription, Entra ID and roles well finds the Azure foundations there.
- AI-103, the developer certification, is a course of its own and not part of this course.
Exercise 1
Nordlicht wants a new helper: Travelers should be able to send voice messages to customer service that appear as text and are briefly summarized for the advisers. Design the solution as a table row: service, deployment yes or no, human step, two principles and at least two points that are not implemented in the pilot.
Solution 1
One possible answer, there are others: Service: Azure Speech, speech recognition (2.7), then text analysis (2.6); alternatively Content Understanding for audio (2.9). Deployment: no for Speech and Language, yes for Content Understanding. Person: Advisers check transcript and summary against the original audio before they reply. Principles: Privacy and security (consent and retention open) and accountability (Ms. Berg). Not in the pilot: automatic replies, decisions about refunds, evaluating without a settled legal basis.