NodnRegression
Predicts a number: a house price, tomorrow's temperature, or units sold. The label is numeric.
Study guide · Core notes
Azure AI Fundamentals Notes: workloads and services
Reference notes for the retired AI-900 exam: which workload a scenario is, how machine learning is trained, and which Azure service handles vision, language, speech, or a prompt.
How to read these notes: Items tagged Extra are outside the May 2, 2025 skill list. Learn the unmarked items first. Exam AI-900 retired on June 30, 2026. These notes do not cover AI-901. New to the exam? Start with the overview for the format and domain weights.
Module 1: Common AI workloads
An AI workload is the kind of job you are asking a model to do. AI-900 starts by matching a scenario to one of four families. Name the family before you name a product.
| Workload | What it does | A typical ask |
|---|---|---|
| Computer vision | Interprets images and video. | What is in this photo, where is the object, or what words are printed on it? |
| Natural language processing | Interprets or produces human language, including speech. | What is this review about, who is named, or what did the caller say? |
| Document processing | Pulls structured fields out of forms, receipts, and invoices. | What is the total, the date, and the vendor on this scan? |
| Generative AI | Creates new content from a prompt. | Write a reply, draft code, or generate an image that did not exist. |
Document processing is more than reading characters. Optical character recognition returns text. Document processing returns fields you can put in a database, such as a line-item table. Azure AI Document Intelligence is the usual Azure service for that job. The skill bullet names the workload.
A caption that describes a photo you already have is a vision task. A paragraph that invents a product description from three bullets is generative. The difference is whether the content is new.
Module 2: Responsible AI principles
Microsoft's six principles. A scenario usually tests one of them. Do not treat the list as six names for the same idea.
| Principle | The question it answers | A failure it is meant to catch |
|---|---|---|
| Fairness | Are people in similar situations treated similarly? | A hiring model scores one group lower for the same qualifications. |
| Reliability and safety | Does the system behave as intended, including when it is unsure or under stress? | A support bot gives a dangerous instruction, or it keeps answering when it should stop. |
| Privacy and security | Is personal data protected, and is access controlled? | Patient notes are sent to a public endpoint, or a prompt leaks another customer's data. |
| Inclusiveness | Can people with different abilities use it? | A voice-only flow has no text alternative for someone who cannot hear the reply. |
| Transparency | Can people understand what the system does and where it is limited? | Users cannot tell that a photo or a paragraph was generated. |
| Accountability | Are people responsible for the system and its decisions? | A denied claim has no owner who can explain or override the model. |
Fairness is about the decision across groups. Inclusiveness is about whether people can use the experience. Privacy is about the data. Transparency is about understanding. Accountability is about human ownership. Reliability and safety is about the system doing the right thing and failing safely.
Generative AI adds a few sharper versions of the same principles. A hallucination is fluent output that is not supported by facts. Content filters reduce harmful text and images. They do not make an answer true. Telling the user that content was generated is a transparency control. A person who can review and override the output is an accountability control.
Module 3: Machine learning techniques
Supervised learning uses examples that already have the answer. Unsupervised learning looks for structure when you do not have that answer.
Classification
Predicts a category: spam or not, a species, or a churn flag. Two categories is binary classification. More than two is multiclass.
Clustering
Groups similar rows when there is no label. Customer segments discovered from behavior are the usual example. You did not tell the model the segment names in advance.
Deep learning uses neural networks with many layers. It is useful when the useful features are hard to write by hand, such as pixels in a photo or samples in an audio clip. A decision tree on a small table of columns is not deep learning.
A Transformer is a neural network design that uses attention. Attention lets the model weigh which tokens matter to each other, and it can process a sequence without reading only one word at a time the way older recurrent networks did. Many large language models are built on Transformers. Recognizing that feature is the AI-900 ask. Training one is not.
Module 4: Features, labels, training, and validation
The columns and the split decide whether a score means anything.
| Term | Role |
|---|---|
| Feature | An input the model may use. Square footage and neighborhood are features when you predict a price. |
| Label | The value you want to predict in supervised learning. The sale price is the label in that example. Clustering has no label. |
| Training dataset | The rows the model learns from. It adjusts to these examples. |
| Validation dataset | Held-out rows used to measure how the model does on data it did not learn from, and to compare choices such as algorithms. |
If you score the model on the same rows you trained on, the number looks better than real use. That gap is overfitting: the model memorized the training rows and does not generalize. The validation set is how you notice.
Extra A third split, the test set, is a final check you do not use while choosing the model. The May 2, 2025 bullets name training and validation, not a separate test set.
Module 5: Azure Machine Learning
Azure Machine Learning is the platform for the data-science workflow: data, compute, training, a registered model, and a deployment. AI-900 asks what it can do, not which button to click.
- Automated machine learning tries algorithms and settings for a task you define, then ranks models. You still supply data and say whether the task is regression, classification, or something similar. It does not invent the business question.
- Data can come from datastores registered to the workspace, including cloud storage. The workspace is the boundary for the experiment, the data, and the models.
- Compute instance is a development machine for notebooks. A compute cluster is scalable compute for training. They are not the same bill or the same job.
- Model management registers a trained model so you can version it. Deployment publishes a model to an endpoint. A real-time endpoint scores one request at a time. Batch deployment scores a file of rows.
Registering a model does not make it available to an application. Deployment does. Training on a cluster does not replace validation.
Module 6: Computer vision
Four solution types, then two Azure services.
| Solution | Output | Not this |
|---|---|---|
| Image classification | A label for the whole image, such as "mountain" or "cat." | It does not draw a box around each object. |
| Object detection | Each object, plus a bounding box that says where it is. | A single scene label is classification, not detection. |
| Optical character recognition | Printed or handwritten text extracted from the image. | It does not judge whether the sentence is positive. |
| Facial detection | Whether a face is present and where it is. | It does not by itself say who the person is. |
| Facial analysis | Attributes of a detected face, such as whether the person is smiling or wearing glasses. | The skill list does not ask you to identify a person by name. |
Azure AI Vision analyzes images: tags, captions, object detection, and reading text. Azure AI Face is the service for detecting faces and analyzing facial attributes. Do not send a receipt to Face because you want the total. Do not send a face-location job to a sentiment model.
Extra Custom Vision and identifying a specific person are outside the May 2, 2025 bullets. Microsoft has also limited some facial recognition scenarios. Stay with detection and analysis unless a newer exam guide says otherwise.
Module 7: Language understanding
These tasks all start from text. They answer different questions, which is why the distractors on a practice item look alike.
Key phrase extraction
The main talking points, as short phrases. It is not a full summary, and it is not a list of people and places.
Entity recognition
Named things: people, organizations, locations, dates, quantities. "Paris" as a place is an entity. "on-time delivery" as a theme is a key phrase.
Sentiment analysis
Whether the text is positive, negative, neutral, or mixed. It does not translate the text and it does not extract the invoice total.
Language modeling
Learning how likely a sequence of words is, which is how a model predicts or generates the next text. It is a language task, not OCR.
Azure AI Language covers text analysis: sentiment, key phrases, entities, language detection, and related understanding features. It does not turn a microphone recording into text. That is speech.
Module 8: Speech and translation
Direction matters. Speech recognition and speech synthesis are opposites.
| Task | Direction | Service |
|---|---|---|
| Speech recognition | Audio in, text out. | Azure AI Speech |
| Speech synthesis | Text in, spoken audio out. | Azure AI Speech |
| Speech translation | Spoken language in, another language out, as text or speech. | Azure AI Speech |
| Text translation | Written text in one language, written text in another. | A translation feature. Do not use sentiment analysis for this. |
Azure AI Speech is the service for recognition, synthesis, and speech translation. Azure AI Language stays on text. A call-center recording that must become a transcript is speech recognition, even if you later run sentiment on the transcript. Those are two steps and two services.
Module 9: Generative AI
A generative model creates new content: text, code, images, or other media. A discriminative model, such as a classifier, chooses a label for content that already exists. AI-900's largest domain is the generative side.
Common scenarios are a chat assistant, a copilot inside an application, drafting or summarizing text, generating code, and generating images. The input you give the model is the prompt. The text it returns is the completion or response. The model reads and writes text in tokens, which are pieces of words rather than whole sentences.
Responsible AI considerations show up here more sharply than in a simple classifier:
- Hallucinations. The answer sounds confident and is not grounded in a source. A content filter does not fix that. Supplying relevant source text with the prompt, often called grounding, reduces unsupported claims. It is not the same thing as retraining the model.
- Harmful content. Filters can block violent, hateful, or sexual output. They are a reliability and safety control, not a fairness audit.
- Transparency. Users should be able to tell that content was generated and what the system cannot be trusted to do.
- Accountability. A person or team owns the application, including when the model is wrong.
- Privacy. Prompts can contain customer data. Where that data is sent, and who can read it, is a privacy and security decision.
Extra Temperature and other sampling settings change how varied the output is. The May 2, 2025 bullets do not name them. Do not treat a slider as a required objective.
Module 10: Azure AI Foundry, Azure OpenAI, and the model catalog
Three names the May 2, 2025 guide uses. Later AI-901 material calls the platform Microsoft Foundry. Keep the AI-900 names when you study this exam.
| Name | What it is for | What it is not |
|---|---|---|
| Azure OpenAI | OpenAI models deployed in your Azure subscription. Chat completions, embeddings, and image generation, depending on the model, plus content filters and Azure security controls. | Not the catalog of every provider, and not a classifier for an existing photo library. |
| Azure AI Foundry | The place to explore models, try prompts, evaluate results, and build a generative AI application or agent. | Not itself a single model. A playground is a way to test a prompt before you write an application. |
| Model catalog | A list of foundation models from Microsoft, OpenAI, and other providers that you can review and deploy from Foundry. | Not limited to one vendor's models. |
Use Azure OpenAI when the requirement is an OpenAI model with Azure identity, networking, and content filtering. Use the model catalog when you need to compare or deploy models from more than one provider. Use Foundry as the workspace around that work. Use Azure Machine Learning when the job is classical or custom model training, automated machine learning, and your own training data, rather than a chat completion against a foundation model.
Extra Azure AI Content Safety, Azure Bot Service, and Anomaly Detector appear in broader Microsoft AI documentation. The May 2, 2025 skill list does not name them as measured objectives. Content filters on generative output are in scope. The product name Content Safety is the extra part.
Frequently asked questions
What is the difference between regression, classification, and clustering?
Regression predicts a number, such as a price. Classification predicts a category, such as spam or not spam. Clustering groups similar items when you have no labels. Features are the inputs. The label is the value you want to predict, and clustering does not use one.
Which service reads text in a photo, and which one judges the tone of a sentence?
Azure AI Vision can read text in an image with optical character recognition. Azure AI Language judges sentiment, extracts key phrases, and finds entities in text. Azure AI Speech turns speech into text or text into speech. Azure AI Face locates faces and can describe facial attributes.
What is the difference between Azure OpenAI and the model catalog?
Azure OpenAI gives you OpenAI models inside Azure, with Azure identity, networking, and content filters. The Azure AI Foundry model catalog lists foundation models from Microsoft, OpenAI, and other providers. Azure AI Foundry is the place you explore, test, and build a generative AI app.
Continue the Azure AI Fundamentals study path
Drill the distinctions, then test them under a timer.
Overview
Retirement date, the 700 passing score, domain weights, and a study plan.
Core Notes
Reference notes for workloads, machine learning, vision, language, and generative AI.
Flashcards & Memory Notes
40 recall cards in four decks, plus memory notes for principles, models, and services.
Practice Exams
200 original questions with custom exams, explanations, and a score report by domain.
Related Tools
Useful companions while you study.
Password Generator
Generate strong random passwords with custom length, characters, symbols, and security options.
JSON YAML Converter
Convert structured data between JSON and YAML with formatting and validation.
Base64 Converter
Encode text and files to Base64 or decode Base64 data with UTF-8 support.
Diff Checker
Compare two text blocks and highlight added, removed, and changed content.
Legal Disclaimer
NodnWebTools provides general informational, educational, and convenience resources. Calculations, conversions, estimates, and learning materials may contain errors or become outdated. Financial, tax, medical, legal, and travel information is not professional advice. Verify important results and current requirements with qualified professionals or authoritative sources. Protect sensitive files and personal information, review each tool’s privacy limitations, and use only content you are authorized to process. You are responsible for how you use and share results. Study resources are independent and do not guarantee exam success or imply certification-provider endorsement. Microsoft, Azure, and related marks are trademarks of Microsoft Corporation. NodnWebTools is not affiliated with, endorsed by, or sponsored by Microsoft.