What counts as an AI workload

Azure AI Fundamentals groups AI workloads into six areas you must recognise: machine learning, computer vision, natural language processing (NLP), knowledge mining/document intelligence, generative AI, and conversational AI (bots/agents). Exam questions often give you a scenario and ask which workload it maps to - learn the trigger words. 'Read text from a scanned form' = document intelligence. 'Detect objects in a photo' = computer vision. 'Summarise customer reviews' = NLP. 'Write a product description from a prompt' = generative AI.

Responsible AI - the six Microsoft principles

Microsoft defines six responsible AI principles, and the exam expects you to match a scenario to the right one:

• Fairness - the system treats all groups of people equitably, no bias in outcomes (eg a loan model shouldn't favour one demographic).
• Reliability and safety - the system performs consistently and safely, including under edge cases, and is tested rigorously before release.
• Privacy and security - personal data is protected, used only with consent, and kept secure.
• Inclusiveness - AI empowers everyone, including people with disabilities, and considers diverse needs.
• Transparency - people understand how and why a system makes decisions; users are told they're interacting with AI.
• Accountability - people and organisations are responsible for how AI systems operate, with governance and human oversight in place.

Common mistakes

Students mix up fairness (bias/equity) with inclusiveness (access/diverse needs) - fairness is about outcomes across groups, inclusiveness is about who can use and benefit from the system. Also watch transparency vs accountability: transparency is about explaining decisions to users, accountability is about who is responsible when things go wrong (governance, human-in-the-loop). Reliability and safety often gets confused with security - reliability is about consistent correct performance, security is about protecting data and systems from harm.

Human-in-the-loop and governance

Expect scenario questions where a high-stakes decision (medical, legal, financial) needs human review before acting - that's accountability plus reliability and safety working together. Microsoft's guidance stresses that responsible AI isn't a single checkbox; it spans the whole lifecycle from design through deployment and monitoring.

Exam tip

When a question describes a real-world AI scenario, first identify the workload type, then separately identify which responsible AI principle(s) apply if asked. Don't assume every question wants both - read carefully whether it's asking 'what kind of AI is this' or 'which principle does this violate/support'.

What machine learning actually is

Machine learning (ML) is a subset of AI where a model learns patterns from data instead of being explicitly programmed with rules. The model is trained on historical data, then used to make predictions on new data. In Azure, the core service for building ML models is Azure Machine Learning (Azure ML), accessed via Azure ML Studio.

The core ML workflow

Every ML project follows roughly the same steps: collect and prepare data, split it into training and validation (test) datasets, select an algorithm, train the model, evaluate its performance with metrics, then deploy it as an endpoint others can call. A common mistake is evaluating a model on the same data it was trained on - this gives falsely high accuracy, which is why a held-out test set matters.

Types of machine learning

• Supervised learning: trains on labelled data (inputs plus known correct outputs). Covers regression (predicting a numeric value, eg house price) and classification (predicting a category, eg spam or not spam).
• Unsupervised learning: trains on unlabelled data to find structure, most commonly clustering (grouping similar items, eg customer segments).
• Reinforcement learning: an agent learns by trial and error, receiving rewards or penalties for actions taken in an environment.

Features and labels

A feature is an input variable used to make a prediction (eg square footage). A label is the known output value used in supervised training (eg the actual sale price). The exam expects you to identify which is which in a scenario.

Key evaluation metrics

• Regression: uses metrics like MAE (mean absolute error) and RMSE (root mean squared error) - lower is better for both.
• Classification: uses accuracy, precision, recall, and a confusion matrix that shows true/false positives and negatives.
• Clustering: no labels to check against, so metrics judge how tight and separated the clusters are.

Azure Machine Learning Studio tools

Azure ML Studio offers a no-code Designer (drag-and-drop pipelines), Automated ML (AutoML, which automatically tries multiple algorithms and picks the best), and Notebooks for full code control using Python SDK. AutoML is the go-to answer for 'build a model quickly with minimal code'.

Common mistakes to avoid

• Confusing classification (categories) with regression (numbers) - a very common trick in exam scenarios.
• Forgetting that clustering is unsupervised - there is no 'correct answer' label to train against.
• Thinking Azure ML Studio requires coding - Designer and AutoML are code-free.
• Mixing up training data (used to build the model) with validation/test data (used to check it works on unseen data).

What computer vision covers on Azure

Azure's computer vision services let you extract information from images and video without training your own models. The two main services are Azure AI Vision (formerly Computer Vision) and Azure AI Custom Vision. There's also Face for face detection/analysis and Azure AI Video Indexer for video.

Azure AI Vision

This is the general-purpose, pre-trained service. Key capabilities:

• Image analysis: tags, captions, dense captions, objects, categories, brands and adult/racy content ratings
• Optical Character Recognition (OCR) using the Read API - extracts printed and handwritten text, supports many languages
• Spatial analysis: understands people's movement in physical space via video, e.g. counting how many people entered a zone
• Smart-cropping and background removal
• Face detection is included but the older Face service gives more detailed facial attributes

Custom Vision

Use this when the built-in tags don't match your business need - for example spotting a specific defect on a production line, or identifying a particular product. Custom Vision has two project types:

• Image classification: assigns one or more labels to a whole image
• Object detection: identifies objects AND returns bounding box coordinates for each one

You upload and tag your own training images, then train and publish a model via the Custom Vision portal or SDK. It needs far fewer images than building a model from scratch because it uses transfer learning on top of a pre-trained backbone.

Face service

Detects human faces in an image and returns attributes like bounding box, head pose and (with limited access approval) age and emotion. Face identification and verification (matching a face to a known identity) are restricted-access features due to responsible AI concerns - you must apply for approval.

Common exam traps

• Don't confuse Image Analysis (general tagging/captions) with Custom Vision (your own labels)
• OCR reads text FROM images; it does not translate it - that needs Azure AI Translator afterwards
• Object detection returns a bounding box; image classification does not
• Spatial analysis is about counting/tracking people in physical space, not identifying who they are
• Face identification/verification need Limited Access approval - this is tested because of Microsoft's Responsible AI commitments
• These are all part of Azure AI Services (the multi-service resource) OR can be provisioned as single-service resources - know both options exist

Provisioning

You can create a single Vision resource (billed only for that service) or an Azure AI Services multi-service resource (one key/endpoint covers Vision, Language, Speech etc, useful for prototyping and consolidated billing).

What Azure NLP covers

Natural language processing (NLP) is how Azure AI services understand and generate human language - text and speech. For AI-900 the two headline services are Azure AI Language and Azure AI Speech, both part of Azure AI Services (the multi-service resource).

Azure AI Language features

• Key phrase extraction - pulls out the main talking points from a block of text.
• Sentiment analysis - scores text as positive, negative, neutral or mixed, with a confidence score between 0 and 1 for each label.
• Named Entity Recognition (NER) - finds and categorises entities such as people, places, dates, quantities and organisations.
• PII detection - a specific type of entity recognition that flags and can redact personal data like names, phone numbers and card numbers.
• Language detection - identifies the language a text is written in and returns an ISO code plus a confidence score.
• Conversational Language Understanding (CLU) - the successor to the old LUIS service, used to build custom intent and entity models for chatbots and voice apps. Remember: LUIS is retired, CLU is the current tool.
• Question Answering - turns a knowledge base (often a FAQ document) into an API that answers natural-language questions.

Azure AI Speech features

• Speech to text - transcribes spoken audio into text, real-time or batch.
• Text to speech - converts text into natural-sounding spoken audio; supports both prebuilt neural voices and custom voices.
• Speech translation - converts spoken audio in one language into text or speech in another, in near real time.
• Speaker recognition - identifies or verifies who is speaking from their voice.

Common mistakes to avoid

• Do not confuse sentiment analysis (whole-document tone) with key phrase extraction (topic words) - they answer different questions.
• Language detection returns a confidence score; a value of (Unknown) means the service could not reliably tell.
• CLU replaces LUIS - if an exam item mentions building custom intents, the current answer is CLU, not LUIS.
• Translator (text translation) is a separate Azure AI service from Speech Translation, but both sit under the NLP umbrella - Translator does text-to-text, Speech does audio-involving translation.
• PII detection is not the same as content moderation; PII is about personal data, not offensive content.

Quick recap

Azure AI Language handles written text tasks (sentiment, key phrases, NER, PII, language detection, CLU, Question Answering). Azure AI Speech handles spoken language tasks (speech-to-text, text-to-speech, translation, speaker recognition). Both can be accessed through a single multi-service Azure AI Services resource or as standalone resources.

What is generative AI?

Generative AI creates new content - text, images, code, audio - rather than just classifying or predicting from existing data.

Azure OpenAI Service gives you managed, enterprise-grade access to OpenAI's models (GPT, embeddings, DALL-E) inside your own Azure subscription, with Azure's security, networking and compliance wrapped around it.

Key model families

• GPT models (e.g. GPT-4 class) - generate and understand natural language and code, used for chat, summarisation, writing, reasoning.
• Embeddings models - turn text into numeric vectors for search, clustering and comparing meaning (used in RAG systems).
• DALL-E models - generate images from text prompts.
• Whisper - speech-to-text transcription and translation.

Prompts and prompt engineering

• A prompt is the input you give the model; the completion is its output.
• System message sets behaviour/persona/rules for the whole conversation.
• Few-shot prompting gives examples in the prompt to steer output format and style - no retraining needed.
• Zero-shot prompting asks the model to do a task with no examples at all.
• Prompt engineering (wording, examples, structure) is the cheapest, fastest way to improve output quality - try this before fine-tuning.

Key generation parameters

• Temperature (0-1 typically) - controls randomness; low = focused/deterministic, high = creative/varied.
• Top P - nucleus sampling, alternative way to control randomness by limiting the probability mass considered.
• Max tokens - caps the length of the response; a token is roughly 4 characters or 0.75 of a word in English.
• Stop sequences - strings that tell the model to stop generating.

Grounding and responsible use

• Base models can 'hallucinate' - producing plausible but false or unverifiable content - because they generate the statistically likely next token, not verified facts.
• Grounding techniques (e.g. Retrieval Augmented Generation, RAG) reduce hallucination by feeding the model relevant, trusted data (like your own documents) at prompt time via Azure AI Search plus 'your data' features in Azure OpenAI Studio.
• Fine-tuning trains a model further on your own labelled examples to change its style or specialise it for a task - used only when prompt engineering isn't enough.
• Azure OpenAI includes built-in content filtering for harmful content, and Microsoft's Responsible AI principles (fairness, reliability & safety, privacy & security, inclusiveness, transparency, accountability) apply throughout.

Common exam trap

Don't confuse Azure OpenAI Service (specific OpenAI models via Azure) with Azure AI Studio / Azure AI Foundry (the broader platform for building generative AI apps, including model catalog beyond OpenAI). Also remember DALL-E is for image generation, not Whisper (that's speech) - the exam likes to mix these up.

Azure AI services overview

Azure AI services (formerly Cognitive Services) are pre-built, pre-trained AI models exposed as APIs or SDKs. You call them without training your own model, which makes them the fastest route to adding AI capability to an app.

The main service categories

• Vision: image classification, object detection, OCR (Read API), spatial analysis via Azure AI Vision.
• Language: sentiment, key phrase extraction, entity recognition, translation, question answering via Azure AI Language.
• Speech: speech-to-text, text-to-speech, speech translation via Azure AI Speech.
• Decision: anomaly detection, content moderation, personalizer for recommendations.
• Azure OpenAI Service: access to GPT, embeddings and DALL-E models with Azure security and compliance wrapped around them.
• Azure AI Search: adds AI enrichment (OCR, entity extraction) to search indexes over your own data.

Resource types — the exam favourite

When you provision in the Azure portal you choose between:

• A single-service resource — dedicated endpoint and key for just one service, e.g. only Language. Gives its own billing meter and separate access control.
• A multi-service resource (the Azure AI services resource) — one endpoint and one key for most services (Vision, Language, Speech, Decision, etc.). Simplifies billing into one line and one set of credentials.

Note: Azure OpenAI and some newer/preview services still need their own dedicated resource — they are not bundled into the multi-service resource.

Authentication and endpoints

• Every resource gives you an endpoint URL plus two keys (a primary and secondary) for key rotation without downtime.
• You can also authenticate with Azure Active Directory (Microsoft Entra ID) tokens for more secure, keyless access in production.
• Regional endpoint: the resource is tied to an Azure region; calls must go to that region's endpoint.

Pricing tiers

• Free tier (F0): limited calls per month/minute, good for learning and testing — one F0 per subscription per service.
• Standard tier (S0 and above): pay-as-you-go, scales for production, higher rate limits.

Responsible AI

Microsoft's six Responsible AI principles apply across every Azure AI service: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. Expect exam questions matching a scenario to one of these six.

Common mistakes

• Confusing 'multi-service resource' with 'getting every Azure AI service in one place' — Azure OpenAI is usually separate.
• Forgetting that a free (F0) tier is capped and only one is allowed per subscription per service.
• Mixing up the key (authentication) with the endpoint (the URL you send requests to) — you need both.
• Assuming these services require you to train a model — they are pre-trained; you only fine-tune in specific cases (e.g. Custom Vision, Azure OpenAI fine-tuning).