arrow_back Cert

Microsoft AI-900 Azure AI Fundamentals

Study & Mastery Mode

🔥 0 streak

Score 0% /70%

timer Mock Exam lock Pro menu_book Course description 3-Page download Free

menu_book

AI-900 Azure AI Fundamentals

Compressed Course

AI-900 Microsoft Azure AI Fundamentals — Compressed Exam-Preparation Course

Target exam: AI-900: Microsoft Azure AI Fundamentals
Blueprint used: Skills measured as of May 2, 2025
Passing score: 700 or greater
Important timing note: Microsoft states that AI-900 will retire on June 30, 2026 and will be replaced by AI-901. This course is intentionally aligned to AI-900, not AI-901. Do not mix AI-901 implementation topics into your AI-900 revision unless you are changing exams.
Branding note: The AI-900 blueprint uses the name Azure AI Foundry. Newer Microsoft documentation may use Microsoft Foundry. For AI-900 questions, recognize both names but answer using the concept requested by the question.

1. Exam Overview

What the exam is testing

AI-900 is a fundamentals exam. It tests whether you can recognize:

the correct AI workload for a business requirement;
the most suitable Azure service for that workload;
the difference between closely related AI capabilities;
the basic logic of machine learning;
responsible AI principles;
common generative AI scenarios, risks, and safeguards.

It is not primarily a coding exam. You should be able to read a short business scenario, identify the required output, eliminate unrelated capabilities, and choose the simplest correct option.

The source question bank contains 1,104 original practice questions organized into easy, medium, and hard sections. Its repeated patterns reveal the most valuable exam skills:

Match the data type to the correct workload.
Match the requested output to the correct capability.
Distinguish a managed Azure service from an unrelated service.
Recognize when a plausible answer solves a nearby but different problem.
Apply responsible AI controls to public-facing or high-impact systems.
Choose the correct model type, dataset role, or endpoint in Azure Machine Learning.
For generative AI, distinguish generation from extraction and add grounding, evaluation, safety controls, and human review where appropriate.

How to think like the exam

Use this four-step method for almost every scenario:

Identify the input type.
Is the input text, audio, image, video, a document, or tabular data?
Identify the required output.
Does the user need a category, a number, extracted text, a detected object location, generated content, a transcription, a translation, or a risk-control decision?
Select the narrowest correct capability.
Prefer the workload or service that directly produces the requested output. Do not choose a broad service when the question asks for a specific capability.
Reject adjacent distractors.
Many wrong answers are not random. They are nearby concepts that operate on the wrong data type or produce the wrong output.

How to use this course

Study the guide in four passes:

Pass 1: Learn the workload map and service map.
Pass 2: Memorize the confusing pairs.
Pass 3: Practice scenario elimination.
Pass 4: Use the rapid-review and exam-day checklist.

2. Exam Domains

Official AI-900 domain list

Official AI-900 domain	Official weighting	Priority
Describe Artificial Intelligence workloads and considerations	15–20%	High
Describe fundamental principles of machine learning on Azure	15–20%	High
Describe features of computer vision workloads on Azure	15–20%	High
Describe features of Natural Language Processing (NLP) workloads on Azure	15–20%	High
Describe features of generative AI workloads on Azure	20–25%	Highest

Priority notes

Generative AI has the largest official weighting. However, the other four domains are each large enough that you cannot safely skip any of them.

A practical revision split:

Revision time	Recommended focus
Generative AI	25%
AI workloads and responsible AI	19%
Machine learning fundamentals	19%
Computer vision	18%
NLP and speech	19%

What matters most

The most exam-relevant decision rules are:

Images or video: think computer vision.
Written language: think NLP.
Spoken audio: think speech recognition, speech synthesis, or speech translation.
Scanned document fields: think document processing and OCR.
New text, summaries, answers, or code: think generative AI.
Numeric prediction: think regression.
Known category prediction: think classification.
Unlabeled grouping: think clustering.
Real-time model prediction API: think managed online endpoint.
Scheduled large-scale scoring: think batch endpoint.
Trusted-answer generation: think grounding with approved data.
High-impact generated output: think human review and safeguards.

3. Start-to-Finish Study Path

Foundation

Master these before moving on:

AI workload families:
- computer vision;
- NLP;
- document processing;
- machine learning;
- generative AI.
ML techniques:
- regression;
- classification;
- clustering.
Responsible AI principles:
- fairness;
- reliability and safety;
- privacy and security;
- inclusiveness;
- transparency;
- accountability.
Core service map:
- Azure AI Vision;
- Azure AI Face detection service;
- Azure AI Language;
- Azure AI Speech;
- Azure Machine Learning;
- Azure OpenAI Service;
- Azure AI Foundry;
- Azure AI Foundry model catalog.

Intermediate

Practice choosing between confusing options:

image classification vs object detection;
OCR vs document processing;
speech recognition vs speech synthesis;
entity recognition vs key phrase extraction;
classification vs clustering;
online endpoint vs batch endpoint;
Azure OpenAI Service vs Azure AI Foundry;
grounding vs unrestricted generation;
transparency vs accountability;
fairness vs reliability and safety.

Advanced

Work through scenario logic:

Find the exact output requested.
Check whether the input type matches the service.
Eliminate answers that return a different output.
Add governance controls when the workload affects people or uses sensitive information.
Add evaluation, content safety, grounding, and human review to generative AI where required.
Avoid overengineering: a fundamentals question usually rewards the simplest directly aligned solution.

Final review

On your last revision day:

Read the rapid service-selection table.
Recite the six responsible AI principles.
Compare the confusing pairs.
Review the “If you see X, think Y” rules.
Confirm that you can explain why each top distractor is wrong.

4. Core Concepts by Domain

Domain 1 — Describe Artificial Intelligence Workloads and Considerations

4.1 Common AI workloads

Workload selection table

Requirement	Think	Do not confuse with
Analyze images or video	Computer vision	NLP
Analyze written language	NLP	Computer vision
Extract structured fields from forms or invoices	Document processing	Image classification
Extract visible text from an image or scan	OCR	Object detection
Convert spoken audio into text	Speech recognition	Speech synthesis
Convert text into spoken audio	Speech synthesis	Speech recognition
Generate a draft, answer, summary, image, or code	Generative AI	OCR
Predict a number	Regression	Classification
Predict a known category	Classification	Clustering
Discover natural groups without known labels	Clustering	Classification

Computer vision workload

Computer vision interprets visual information such as images and video.

Examples:

identify damaged products in photographs;
locate vehicles in a camera frame;
read serial numbers from equipment labels;
detect whether a face appears in an image.

Exam rule: If the input is pixels and the question is about visible content, start with computer vision.

NLP workload

Natural language processing works with text and language.

Examples:

determine whether a review is positive or negative;
extract organization names from reports;
identify important phrases in feedback;
translate support messages.

Exam rule: If the input is written human language, start with NLP.

Document processing workload

Document processing extracts structured information from documents such as forms, invoices, or receipts.

Examples:

vendor name;
invoice number;
total amount;
form fields;
document-specific structured values.

Trap: OCR reads text. Document processing is broader: it can organize extracted information into meaningful fields.

Generative AI workload

Generative AI creates new content based on instructions or context.

Examples:

generate a response;
summarize a conversation;
answer a question using approved documentation;
draft product descriptions;
generate starter code.

Trap: OCR extracts existing content. Generative AI creates new output.

4.2 Responsible AI principles

Principle	Core question	Typical scenario
Fairness	Are outcomes equitable across groups?	Check whether a screening model produces unjustified differences in outcomes
Reliability and safety	Does the solution behave consistently and avoid harmful failures?	Test unexpected, ambiguous, or adversarial input
Privacy and security	Is sensitive data protected?	Restrict access, reduce unnecessary collection, prevent data leakage
Inclusiveness	Can people with varied abilities and circumstances use the solution?	Provide alternatives for users who cannot rely on audio
Transparency	Do users understand that AI is used and know important limitations?	Tell users an answer was AI-generated and may require verification
Accountability	Is a person or organization responsible for the system and its outcomes?	Assign an owner, escalation path, approval process, or appeal route

Fairness

Choose fairness when the main issue is unequal treatment or unjustified differences in outcomes.

Example:

A loan-screening system has lower approval rates for one demographic group even when relevant factors are similar.

Correct focus: fairness.

Do not choose transparency merely because explanations are useful. Transparency does not replace testing for equitable outcomes.

Reliability and safety

Choose reliability and safety when the concern is robust behavior, failure handling, or harm prevention.

Example:

Test how a chatbot responds to ambiguous, unsafe, or adversarial prompts.

Correct focus: reliability and safety.

Do not choose fairness unless the scenario is about unequal outcomes across groups.

Privacy and security

Choose privacy and security when sensitive data, access controls, confidential prompts, or information exposure are involved.

Typical controls:

data minimization;
access restriction;
secure handling of prompts and retrieved content;
protection of training data;
protection of customer records.

Inclusiveness

Choose inclusiveness when the requirement is to support diverse users, abilities, or circumstances.

Example:

A voice-enabled application must also work for users who cannot rely on audio.

Correct focus: inclusiveness.

Transparency

Choose transparency when users need to know:

AI is being used;
output may require verification;
a model has limitations;
how a recommendation was generated at an appropriate level.

Accountability

Choose accountability when the organization must assign responsibility:

business owner;
approval authority;
review process;
escalation path;
appeal mechanism.

High-impact decision rule

When an AI system can materially affect people, favor:

human review;
escalation;
auditability;
appeal routes;
assigned ownership;
monitoring.

Avoid:

fully automated adverse decisions with no appeal;
removal of human oversight;
removal of logging;
assuming AI output is always correct.

4.3 Domain 1 exam traps

Trap	Why it fails
Choose NLP for image pixels	NLP works with language, not visual content
Choose image classification for document fields	Image classification labels an image; it does not extract invoice totals or form fields
Choose object detection for printed text	Object detection locates objects; OCR extracts visible text
Choose transparency for unequal outcomes	Fairness is the primary principle for equity across groups
Choose privacy for accessibility needs	Inclusiveness is about accommodating varied users
Choose the largest model for a governance problem	Model size does not replace oversight, safety controls, or accountability

Domain 2 — Describe Fundamental Principles of Machine Learning on Azure

4.4 Common machine learning techniques

Technique-selection table

Required output	Correct technique	Example
Numeric value	Regression	Predict delivery time in minutes
One of two known categories	Binary classification	Fraud or not fraud
One known category from several categories	Multiclass classification	Route a request to billing, technical support, or accounts
Multiple applicable categories	Multilabel classification	Tag an article as finance and technology
Natural groupings with no predefined labels	Clustering	Segment customers by purchasing behavior

Regression

Use regression when the output is a continuous number.

Examples:

house price;
energy consumption;
delivery time;
expected sales amount.

Fast clue: If the answer is measured, counted, priced, timed, or forecast as a number, consider regression.

Classification

Use classification when the output is a category.

Examples:

spam or not spam;
fraud or not fraud;
ticket category;
disease class.

Binary classification

Exactly two categories.

Examples:

yes or no;
churn or no churn;
fraud or legitimate.

Multiclass classification

One category from more than two categories.

Example:

route a support ticket into billing, technical support, accounts, or sales.

Multilabel classification

Multiple categories can apply at the same time.

Example:

tag an article with finance, technology, and policy.

Clustering

Use clustering to discover groups when labels are not already defined.

Example:

discover naturally occurring customer segments.

Trap: Classification requires known target categories. Clustering does not.

4.5 Deep learning and Transformer architecture

Deep learning

Deep learning uses neural networks with multiple layers. It is commonly associated with complex patterns in:

images;
audio;
language.

You do not need to design neural networks for AI-900. Recognize the concept and the workloads where it is useful.

Transformer architecture

Transformers use attention mechanisms and are foundational to many modern language and generative AI models.

Exam rule: If the question asks which architecture is associated with modern large language models, choose Transformer architecture, not linear regression or clustering.

4.6 Features, labels, training, and validation

Concept	Meaning	Exam clue
Feature	Input variable used to make a prediction	The model uses it
Label	Known target value or category the model learns to predict	The model predicts it
Training dataset	Data used to fit or train the model	Learn patterns
Validation dataset	Separate data used to estimate performance on unseen examples	Check generalization

Example

Suppose you want to predict a home's sale price.

Column	Role
Floor area	Feature
Number of bedrooms	Feature
Neighborhood	Feature
Sale price	Label

Overfitting

Overfitting happens when a model performs well on training data but poorly on new data.

Exam rule: A validation dataset helps detect whether a model generalizes beyond the examples it learned from.

Do not confuse overfitting with underfitting:

Overfitting: memorizes training-specific patterns.
Underfitting: fails to learn enough useful patterns.

4.7 Azure Machine Learning capabilities

Automated machine learning

Automated machine learning can test candidate algorithms and parameter configurations to help identify a suitable model.

Use it when:

the task is a supported ML problem;
the team wants to compare candidate models;
the goal is to reduce manual experimentation.

Do not confuse automated ML with OCR, speech synthesis, or object detection services.

Compute instance

A compute instance is a managed development environment, often used for:

notebooks;
experimentation;
interactive data science work.

Trap: It is not the production prediction interface.

Compute cluster

A compute cluster provides scalable compute for training workloads.

Use it when:

training jobs need scalable resources;
workloads may scale up or down;
multiple jobs require compute capacity.

Model registration

Registering a model supports model management and version tracking.

Use it to:

track trained model versions;
manage deployment workflows;
improve reproducibility.

Managed online endpoint

Use a managed online endpoint for real-time inference.

Examples:

an application sends a request;
the model returns a low-latency prediction;
predictions are needed on demand.

Batch endpoint

Use a batch endpoint for asynchronous scoring of a large set of records.

Examples:

overnight scoring;
scheduled processing;
bulk predictions for a file.

Online vs batch decision table

Requirement	Choose
Real-time request from an application	Managed online endpoint
Low-latency prediction on demand	Managed online endpoint
Score a large file overnight	Batch endpoint
Scheduled bulk scoring	Batch endpoint

4.8 Domain 2 exam traps

Trap	Why it fails
Choose classification for a numeric value	Classification predicts categories
Choose regression for a yes-or-no result	Regression predicts a number
Choose classification when no labels exist	Use clustering to discover groups
Call the target column a feature	The target is the label
Use training data as the only performance check	Validation data is needed for unseen examples
Use a compute instance as an inference API	Use a managed online endpoint
Use an online endpoint for a large overnight file by default	Use a batch endpoint
Choose linear regression as the architecture behind LLMs	Modern LLMs are commonly associated with Transformers

Domain 3 — Describe Features of Computer Vision Workloads on Azure

4.9 Computer vision capability map

Requirement	Correct capability	Output
Assign a label to the whole image	Image classification	Image-level category
Identify each object and its location	Object detection	Object labels and bounding boxes
Extract visible text	OCR	Machine-readable text
Find faces and their locations	Facial detection	Face presence and position
Analyze image content more broadly	Image analysis	Descriptive visual information

Image classification

Image classification categorizes an entire image.

Examples:

healthy plant or diseased plant;
clear image or blurry image;
damaged package or undamaged package.

Exam clue: The result is an image-level label.

Object detection

Object detection identifies individual objects and where each object appears.

Examples:

find vehicles in a traffic image;
count forklifts in a warehouse frame;
identify visible products on a shelf.

The location is commonly represented by a bounding box.

Exam clue: If the requirement includes “where,” “locate,” “count each,” or “identify every instance,” think object detection.

OCR

Optical character recognition extracts text from images or scanned documents.

Examples:

serial number from a label;
handwritten notes from forms;
storefront hours from a photograph.

Exam clue: If the required result is visible text, choose OCR.

Facial detection

Facial detection identifies whether faces are present and where they appear.

Do not automatically infer identity recognition. A scenario may require detection without identifying the person.

Exam clue: “Find faces” is not the same as “verify identity.”

Azure AI Vision

Azure AI Vision is the managed Azure service associated with common computer vision tasks such as image analysis and OCR.

Azure AI Face detection service

Use the Azure AI Face detection service when a question specifically asks for a managed Azure service for face detection scenarios.

4.10 Computer vision comparison table

Scenario	Best answer	Strong distractor	Why distractor fails
Classify a photo as damaged or undamaged	Image classification	Object detection	The question only needs an overall category
Locate every forklift in an image	Object detection	Image classification	Classification does not return each object's location
Read a serial number from a photo	OCR	Object detection	The desired output is text
Find faces in a photograph	Facial detection	Sentiment analysis	Sentiment analysis works with language
Count products on a shelf	Object detection followed by counting	Image classification	You need individual detections to count instances

4.11 Domain 3 exam traps

Do not choose image classification when the requirement asks for object locations.
Do not choose object detection when the requirement is text extraction.
Do not choose OCR when the question asks whether a whole image belongs to a category.
Do not choose speech recognition for visible text.
Do not choose Azure AI Language for image pixels.
Do not assume face detection means identifying a person.

Domain 4 — Describe Features of Natural Language Processing Workloads on Azure

4.12 NLP capability map

Requirement	Correct capability
Identify the main concepts in text	Key phrase extraction
Find names of people, places, organizations, products, or other entities	Entity recognition
Determine whether a review is positive, neutral, or negative	Sentiment analysis
Work with text sequences or language-generation patterns	Language modeling
Convert spoken audio into text	Speech recognition
Convert text into spoken audio	Speech synthesis
Convert content from one language to another	Translation
Translate spoken language in near real time	Speech translation

Key phrase extraction

Key phrase extraction identifies important phrases or concepts in a text.

Example:

Extract the main topics mentioned in meeting notes.

Correct capability: key phrase extraction.

Do not choose sentiment analysis. Sentiment tells you attitude, not the central topics.

Entity recognition

Entity recognition identifies named entities such as:

people;
places;
organizations;
products;
locations;
identifiers, depending on the scenario.

Example:

Extract organization names and locations from incident reports.

Correct capability: entity recognition.

Do not choose key phrase extraction if the question specifically asks for named entity types.

Sentiment analysis

Sentiment analysis evaluates attitude or opinion.

Examples:

positive review;
negative feedback;
neutral comment;
prioritize strongly negative complaints.

Do not choose entity recognition. It extracts names, not attitude.

Language modeling

Language modeling works with patterns in language sequences. It is associated with predicting or generating text based on context.

Speech recognition

Speech recognition converts audio into text.

Examples:

call transcription;
captions;
transcribe a recorded interview.

Speech synthesis

Speech synthesis converts text into audio.

Examples:

read an accessibility notice aloud;
generate spoken output for an application.

Translation

Translation converts language content from one language into another.

Speech translation

Speech translation combines spoken-language handling with translation.

Example:

Provide near-real-time captions in another language during a presentation.

Correct capability: speech translation.

4.13 Azure NLP services

Azure AI Language

Use Azure AI Language for managed text-analysis scenarios such as:

sentiment analysis;
key phrase extraction;
entity recognition.

Azure AI Speech

Use Azure AI Speech for audio-focused language scenarios such as:

speech recognition;
speech synthesis;
speech translation.

Service split rule

Input and requirement	Service
Written text analysis	Azure AI Language
Audio transcription	Azure AI Speech
Spoken audio output	Azure AI Speech
Speech translation	Azure AI Speech

4.14 NLP comparison table

Scenario	Best answer	Strong distractor	Why distractor fails
Find main ideas in feedback	Key phrase extraction	Sentiment analysis	Sentiment measures attitude
Find company names in reports	Entity recognition	Key phrase extraction	The question asks for named entities
Flag negative reviews	Sentiment analysis	Entity recognition	Entity recognition extracts names
Transcribe a call	Speech recognition	Speech synthesis	The direction is audio to text
Read a notice aloud	Speech synthesis	Speech recognition	The direction is text to audio
Convert Spanish text into English	Translation	Sentiment analysis	Sentiment does not change language
Caption spoken content in another language	Speech translation	Speech synthesis only	Synthesis alone does not translate incoming speech

4.15 Domain 4 exam traps

Speech recognition and speech synthesis are opposite directions.
Entity recognition finds names; key phrase extraction finds important phrases.
Sentiment analysis identifies attitude, not topics or named entities.
Translation changes language; it does not summarize or classify sentiment.
OCR extracts text from images; it is not an NLP answer for ordinary written text.
Azure AI Vision handles visual data; Azure AI Language handles written language; Azure AI Speech handles audio.

Domain 5 — Describe Features of Generative AI Workloads on Azure

4.16 What generative AI does

Generative AI creates new content based on prompts and context.

Common scenarios:

Scenario	Generative AI use
Draft a customer response	Content generation
Produce a concise overview of a long conversation	Summarization
Answer questions using approved documents	Question answering with grounding
Create starter source code from a request	Code generation
Transform or rewrite content	Content transformation

Prompt and completion

Term	Meaning
Prompt	Input instruction or context sent to the model
Completion	Generated output returned by the model

A prompt can include:

the user's question;
instructions;
examples;
context;
retrieved approved content.

A completion is the model's generated response.

4.17 Large language models and Transformers

Large language models can:

generate text;
summarize content;
transform text;
answer questions;
generate code;
reason over provided context.

Modern LLMs are commonly based on Transformer architecture.

Exam trap: OCR is not a language model. OCR extracts visible text; it does not create new prose.

4.18 Azure generative AI services

Azure OpenAI Service

Use Azure OpenAI Service when the scenario asks for access to Azure-hosted generative AI models for:

chat;
summarization;
generation;
question answering;
code generation.

Azure AI Foundry

Use Azure AI Foundry when the scenario focuses on the broader generative AI development environment:

build;
evaluate;
manage;
compare;
develop AI applications.

New Microsoft documentation may call the evolved platform Microsoft Foundry, but the AI-900 blueprint uses Azure AI Foundry.

Azure AI Foundry model catalog

Use the model catalog when the question asks you to:

browse models;
discover models;
compare available models;
choose an appropriate foundation model.

Generative AI service-selection table

Requirement	Best answer
Call a managed Azure-hosted generative model for chat or summarization	Azure OpenAI Service
Build, evaluate, and manage an AI application in an Azure-focused environment	Azure AI Foundry
Browse and choose a suitable model	Azure AI Foundry model catalog
Answer from trusted internal documentation	Azure OpenAI Service plus grounding with approved content
Compare outputs before selecting a model	Azure AI Foundry evaluation capabilities

4.19 Grounding and hallucination risk

Hallucination risk

Generative AI can produce fluent, plausible, but unsupported statements. This is commonly described as hallucination.

Example:

A policy assistant invents a leave-entitlement rule that is not present in the approved handbook.

The answer may sound confident and still be wrong.

Grounding

Grounding supplies trusted context to help a model answer using approved information.

Examples:

product documentation;
internal policy documents;
approved knowledge bases;
current support documentation.

Grounding decision rule

If the application must answer factual questions using trusted organizational content:

choose a generative AI service;
add approved grounding data;
evaluate the output;
apply safety controls;
use human review where risk is high.

Do not assume grounding guarantees perfect accuracy. It reduces risk and improves relevance, but evaluation is still required.

4.20 Responsible generative AI

Generative AI requires controls because output may be:

inaccurate;
inappropriate;
unsafe;
biased;
misleading;
privacy-sensitive.

Use these safeguards:

Risk	Recommended control
Unsupported factual answer	Ground with approved content and evaluate responses
Unsafe or inappropriate output	Add content-safety controls
Sensitive prompts or retrieved documents	Protect data and restrict access
High-impact recommendation	Add human review before irreversible action
User confusion about generated content	Tell users that AI is involved and explain limitations
No clear owner for harm or escalation	Assign accountability and a review process

Human-in-the-loop rule

Use human review when generated output could:

affect a customer's account;
provide legal or financial information;
trigger an irreversible action;
materially affect a person;
be published publicly in a sensitive context.

Avoid fully autonomous publication or decision-making for high-impact content.

4.21 Domain 5 exam traps

Trap	Why it fails
Choose OCR for drafting text	OCR extracts existing text; generative AI creates new content
Increase randomness to improve factual accuracy	More randomness does not make answers more reliable
Remove human review for high-impact output	Generated answers can be plausible but wrong
Use Azure AI Vision for a policy assistant	Vision analyzes images
Choose a compute instance as the generative AI service	A compute instance is a development environment, not the managed generative service
Assume fluent output is accurate	Hallucinations can sound convincing
Share prompts without controls	Prompts and retrieved content can contain sensitive data
Choose the model catalog when the requirement is simply to call a hosted model	The catalog is for discovery and selection; Azure OpenAI Service supplies generative model access

5. Service Selection Guide

5.1 Master service map

Azure service or platform	Use it for	Do not choose it for
Azure AI Vision	Image analysis, OCR, common visual tasks	Text sentiment, audio transcription, LLM chat
Azure AI Face detection service	Detect faces in images	Text extraction, sentiment analysis, speech
Azure AI Language	Text analysis: sentiment, entities, key phrases	Image analysis, audio transcription
Azure AI Speech	Speech recognition, speech synthesis, speech translation	Object detection, invoice field extraction
Azure Machine Learning	Train, manage, and deploy ML models	Simple managed OCR or sentiment questions unless a custom model is explicitly needed
Azure OpenAI Service	Chat, summarization, generation, Q&A, code generation	OCR, object detection, tabular clustering
Azure AI Foundry	Broader AI app development, evaluation, and management	A narrow image-only or speech-only task where a dedicated service is the direct answer
Azure AI Foundry model catalog	Discover and select models	Real-time scoring of a tabular ML model
Managed online endpoint	Real-time inference for a deployed ML model	Bulk overnight scoring
Batch endpoint	Asynchronous bulk inference	Low-latency request-response prediction
Compute instance	Interactive data science development	Production prediction API
Compute cluster	Scalable training compute	Text sentiment analysis as a managed API

5.2 Workload-first decision tree

Use this sequence:

Is the input an image or video?
- Need an overall category? Image classification.
- Need object locations or counts? Object detection.
- Need visible text? OCR.
- Need face presence or location? Facial detection.
- Need a managed Azure visual service? Azure AI Vision or Azure AI Face detection service depending on the requirement.
Is the input written text?
- Need attitude? Sentiment analysis.
- Need named items? Entity recognition.
- Need important topics? Key phrase extraction.
- Need another language? Translation.
- Need a managed text-analysis service? Azure AI Language.
Is the input spoken audio?
- Audio to text? Speech recognition.
- Text to audio? Speech synthesis.
- Speech into another language? Speech translation.
- Need the managed service? Azure AI Speech.
Is the output a prediction from tabular or historical data?
- Number? Regression.
- Known category? Classification.
- Unknown grouping? Clustering.
- Real-time endpoint? Managed online endpoint.
- Bulk scheduled scoring? Batch endpoint.
Is the output newly generated content?
- Generation, summary, Q&A, or code? Azure OpenAI Service.
- Build, evaluate, and manage the application? Azure AI Foundry.
- Discover models? Azure AI Foundry model catalog.
- Need trusted factual answers? Add grounding data and evaluation.

6. Architecture Patterns