AI-901 Video Guide: John Savill's Study Cram
John Savill, a colleague here at Microsoft, has been teaching Azure to the community for years, and his free study crams are among the best exam preparation anywhere. His AI-901 Microsoft Azure AI Fundamentals Study Cram covers the whole exam in 66 minutes: core AI concepts and responsible AI, Microsoft Foundry models and agents, how applications call Foundry, generative AI and embeddings, and the language, speech, vision, and information extraction services, with live demos in the Foundry portal throughout.
How to use this guide
Watch the whole cram once, then use the sections below to revisit topics as you study each domain. Select any timestamp to jump the player to that point. The badges show which exam objectives a section supports and link to those skills on the domain pages. John also shares the whiteboard from the video, which makes an excellent one-page review sheet.
John’s study advice#
- If Python is new to you, start with John’s beginner videos on Python and on AI development. The exam expects you to understand what short code samples do.
- Use the official exam page, the study guide, and the self-paced training and hands-on labs to check each skill measured.
- Get an Azure trial or another Azure subscription so you can try Microsoft Foundry and the AI services yourself.
- If you do not pass on your first attempt, use the score report to focus on your weakest areas before you try again.
The cram, section by section#
The key points below are a study summary of what John explains in each section, written to help you find and review topics. They are not a substitute for watching him explain them.
- The updated AI-901 exam still tests core AI concepts, but it also expects you to understand how apps use AI and to recognize simple Python code.
- John recommends separate beginner videos on Python and AI development so non-developers can build enough grounding to follow the examples.
- The official AI-901 page links to the study guide, exam scheduling, self-paced learning modules, and hands-on labs that map directly to the skills measured.
- The exam is fundamentals level, so the focus is choosing the right AI capability, knowing the main concepts, and understanding what code samples are doing rather than building complex systems.
- John defines AI as software that imitates human abilities such as prediction, evaluation, vision, hearing, speech, language understanding, summarization, and creative generation.
- Machine learning differs from hand-coded rules because the system learns patterns from data and can use labeled examples for prediction or classification.
- Assistants respond to user-controlled turns to improve productivity, while agents can act more autonomously, respond to triggers, plan multiple steps, and work toward a user goal.
- Responsible AI must be designed into solutions through fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability.
- Transparency and accountability matter because users should understand why an AI system made a decision and someone must be responsible for outcomes.
- Azure is the Microsoft cloud platform used for AI-901 scenarios, and Entra ID provides identities for users, groups, devices, and agents.
- A subscription primarily acts as a cost boundary and can also hold access permissions, and it trusts a specific Entra tenant for identity.
- Resource groups hold Azure resources such as storage accounts, virtual machines, networks, and Microsoft Foundry instances, but resource groups are not nested.
- Microsoft Foundry is the pro-developer AI app and agent resource with a portal at ai.azure.com organized around discovering, building, operating, and using AI resources.
- The Foundry model catalog is model-diverse, so learners should expect many models from Microsoft, OpenAI, Anthropic, Hugging Face, and others rather than a single provider.
- John describes a model as the brain of an AI application, and you must deploy a model before an app can use it.
- Deployment choices include global, data zone, regional, model version, token limits, provisioned throughput, priority processing, and guardrails.
- Guardrails provide safety controls such as jailbreak protection, content filtering, and protected-material handling, and their thresholds can be adjusted for legitimate scenarios.
- Foundry supports prompt-based agents for no-code instruction-driven behavior and hosted agents for pro-code implementations packaged and run by Foundry.
- An agent always uses a model, can optionally use tools and knowledge, and uses instructions to define how it should behave.
- Every Foundry instance has an HTTPS endpoint that applications call, commonly by using REST with JSON underneath.
- Apps must authenticate to that endpoint, and John prefers Entra ID because managed identities can avoid storing secrets in code.
- API keys can work for simple scenarios, but they must never be placed in source code, should be protected with a safe mechanism such as Key Vault, and should be regenerated if compromised.
- SDKs hide much of the REST plumbing, and Foundry can generate code samples for different languages, API styles, and authentication methods.
- In the custom app example, the system prompt sets overall behavior, while the user prompt contains the question or task for that turn.
- Large and small language models are both generative, but larger models usually have more parameters and may be more capable, slower, and more expensive than smaller distilled models.
- A request is the input and the response is the output, and the model performs inference as it works through the request.
- Models can handle modalities such as text, audio, images, video, and code, and multimodal means a model supports more than one input or output type, while multimodel means an app uses several models.
- Prompts are converted into tokens and then embeddings, which are high-dimensional vectors that represent meaning and help search by semantics rather than exact words.
- John cautions that generative models are not the answer to every scenario, because Foundry also includes specialized services for language, speech, vision, generation, and extraction.
- Natural language processing focuses on understanding and inferring meaning from human language, including key phrase extraction, named entity recognition, text classification, and summarization.
- Traditional NLP may tokenize, lowercase, remove punctuation or stop words, and tag parts of speech before analytics, while modern NLP also uses embeddings and transformers for context.
- A generative model can demonstrate NLP steps, but Azure Language is more specialized and deterministic for tasks such as language detection and returns structured output with confidence scores.
- John contrasts deterministic language services with nondeterministic generative models, noting that specialized services can be more consistent, predictable, and cheaper when they fit the need.
- Speech workloads go both directions, speech to text for transcription and text to speech for generated audio, with uses such as meetings, service agents, accessibility, notifications, training, and entertainment.
- Computer vision includes image classification for a whole-image label, object detection for what and where, semantic segmentation for pixel-level regions, and contextual analysis, with models that may use labeled data or vision embeddings.
- Image and video generation use diffusion, where training teaches a model to reverse added noise and generation repeatedly denoises toward the concept in the prompt.
- In the Foundry demo, image generation deployments expose practical controls such as resolution, quality, compression, format, number of variations, and optional source images for editing.
- Information extraction turns content into usable data, often starting with OCR and then mapping fields from receipts, invoices, contracts, or other documents through Azure Content Understanding.
- The summary reinforces that exam success depends on matching scenarios to services, understanding deployment choices and guardrails, knowing agents and authentication, and practicing in the training labs.
Keep going#
- The two exam domains for the official skills measured and Microsoft Learn training
- The exam simulator to test yourself on every skill
- John Savill’s Technical Training on YouTube for his other certification crams and Azure deep dives