Common AI workloads and Azure services
The first exam objective is recognising which kind of AI problem a scenario describes. Once you know the workload, the matching Azure service is usually obvious.
Workloads
| Workload | What it does | Example |
|---|---|---|
| Computer vision | Understands images and video | Tag photos, read text, detect objects |
| Natural language processing (NLP) | Understands and produces text and speech | Sentiment, translation, chatbots |
| Anomaly detection | Finds unusual data points | Fraud, faulty sensors |
| Forecasting / prediction | Predicts future numbers from history | Sales or sea level next year |
| Document intelligence | Extracts fields from forms, invoices, receipts | Invoice number, total |
| Knowledge mining | Makes large amounts of content searchable | Azure AI Search with AI enrichment |
| Generative AI | Creates new text, code or images | Copilots, summaries, drafts |
| Conversational AI | Dialogue with users through bots | Support chatbot |
Services
Azure AI services were previously called Azure Cognitive Services, and the exam may still use that name. They are now part of Azure AI Foundry Tools, prebuilt APIs you call without training a model:
- Vision (image analysis, OCR, Face) for images.
- Language for sentiment, key phrases, entities, question answering and conversation understanding.
- Speech for speech to text, text to speech, translation of speech.
- Translator for text translation.
- Document Intelligence for forms.
- Azure AI Search for knowledge mining.
- Azure OpenAI, part of Foundry, for generative AI.
- Azure Machine Learning when you need to train your own model.
Prebuilt or custom
If a prebuilt service does the job (read receipts, translate text), use it. If you need a model tuned to your own data, use a customisable option such as custom image models, custom question answering or Azure Machine Learning.
Speech examples that appear often
- Voice commands to control devices use speech to text followed by language understanding.
- Reading a recipe aloud, or narrating a silent video from a script, is speech synthesis (text to speech).
- Spoken captions are speech recognition.
Exam traps
- Reading text from an image is computer vision (OCR), not NLP, because the input is an image.
- Sentiment analysis on typed text is NLP, not anomaly detection.
- Predicting a number is regression, a category is classification, and unusual values are anomaly detection.
- Document Intelligence extracts fields. Azure AI Search indexes content. They are different services.
- A bot needs a bot service plus a language capability. A bot alone does not understand language.
Practice questions
Choose an answer, then review the explanation to check your understanding.
1. A website must detect when a customer typing in a chat window is upset. Which workload fits?
Show answer
Answer: B.Sentiment of typed text is an NLP task, handled by the Language service.
2. A factory wants an alert when a sensor reading suddenly differs from its normal pattern. Which workload is this?
Show answer
Answer: B.Anomaly detection finds unusual values or patterns in time series data.
3. An app must create a first draft of a product description from a short prompt. Which workload is this?
Show answer
Answer: A.Producing new text from a prompt is generative AI.
4. A company wants to pull merchant, date and total from scanned receipts. Which workload fits best?
Show answer
Answer: B.Extracting structured fields from forms and receipts is done by Azure Document Intelligence (formerly Form Recognizer).