Why AI-900 matters more in 2026 than it did at launch

Microsoft launched AI-900 in 2020 as a conceptual introduction to AI — a credential aimed at business stakeholders and non-technical professionals who needed to understand what AI could and could not do without writing any code. In its original form, AI-900 was a lightweight companion to the Azure Fundamentals (AZ-900) certification, occupying a similar market position: useful for demonstrating awareness, not particularly valued as a technical signal in a hiring context.

The generative AI wave of 2023–2026 changed that calculus. Microsoft’s investment of $13 billion into OpenAI, the launch of Azure OpenAI Service as an enterprise-grade API for GPT-4 and successors, and the rollout of Microsoft Copilot across Office 365, GitHub, Dynamics, and Power Platform have made Azure AI tooling pervasive in enterprise IT environments. Virtually every Azure-using organisation now has at least a pilot deployment of AI-powered services, and roles that previously had no AI surface area — system administrators, business analysts, data engineers, customer success engineers — are now expected to understand how these systems work and how to configure them responsibly. AI-900 has become the baseline literacy credential for that expectation.

The updated 2024 blueprint reflects this shift. The exam now dedicates a full domain to generative AI workloads — covering how large language models work, what the Azure OpenAI Service provides, and how responsible AI principles apply specifically to generative systems where outputs are probabilistic rather than deterministic. For any professional whose role now involves Copilot configuration, Azure AI Studio, or integrating OpenAI models into business workflows, the AI-900 exam content is directly applicable daily work, not abstract theory.

What AI-900 actually tests

The current AI-900 blueprint organises content into five domains, each weighted between 15 and 25% of the exam. The exam is deliberately non-coding: candidates are never asked to write Python, read API documentation, or configure Azure resources step-by-step. The testing format is conceptual and scenario-based — candidates must demonstrate that they understand what each AI service category does, when it is appropriate, what its limitations are, and how it maps to responsible AI principles. The most common question format presents a business scenario and asks candidates to identify the correct Azure AI service or the correct AI concept that applies.

Domain 1: Describe Artificial Intelligence workloads and considerations (15–20%)

This foundational domain establishes the vocabulary and conceptual framework the rest of the exam builds on. Candidates are expected to understand AI at the level of a technically informed business stakeholder — what the common AI workload categories are, what distinguishes them, and what responsible AI means in practice.

  • Common AI workloads: the exam distinguishes between machine learning (pattern recognition from labelled data), computer vision (image and video analysis), natural language processing (text and speech understanding and generation), document intelligence (extracting structured data from unstructured documents), knowledge mining (discovering insights from large information repositories), and generative AI (producing new content — text, images, code — from a prompt). Candidates must match business scenarios to the correct workload category — “a system that reads invoices and extracts line items” is document intelligence, not general machine learning.
  • Responsible AI principles: Microsoft defines six responsible AI principles — fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability — and the exam tests candidates’ ability to apply each principle to a scenario. The most commonly tested scenario type asks candidates to identify which principle is at risk in a given situation: a hiring algorithm that underperforms for one demographic group is a fairness concern; a medical diagnosis AI that cannot explain its reasoning is a transparency concern; a face recognition system deployed without user consent is a privacy and security concern.

Domain 2: Describe fundamental principles of machine learning on Azure (20–25%)

This is the highest-weighted domain and the one candidates most frequently underestimate. It does not require mathematical knowledge of ML algorithms, but it requires a precise conceptual understanding of supervised versus unsupervised learning, the ML workflow, and the Azure Machine Learning service’s capabilities.

  • Types of machine learning: the exam tests the distinction between supervised learning (training on labelled data to predict outputs for new inputs), unsupervised learning (discovering patterns in unlabelled data, typically for clustering or anomaly detection), and reinforcement learning (training an agent to maximise a reward signal through trial and error). Within supervised learning, candidates must distinguish classification (predicting a discrete category — “spam” or “not spam”) from regression (predicting a continuous numerical value — “forecast next month’s sales”). Time series forecasting is covered as a specialised regression use case.
  • The ML workflow: the exam tests the sequence of steps in a machine learning project: data collection and preparation, feature engineering (selecting and transforming inputs), model training, model evaluation (using metrics appropriate to the task — accuracy for classification, RMSE or MAE for regression), and model deployment. Candidates must understand what overfitting is (the model performs well on training data but poorly on new data) and the role of a validation set in detecting it.
  • Azure Machine Learning: the exam tests Azure ML’s no-code and low-code tools. Azure ML Designer is the visual drag-and-drop pipeline builder for constructing training workflows without writing code — the exam tests what Designer can and cannot do. Automated ML (AutoML) trains multiple model types and hyperparameter combinations automatically and selects the best performer — the exam tests what inputs AutoML requires and what metrics it optimises. Azure ML Studio is the workspace that hosts Designer, AutoML, and the notebook environment. Candidates must understand the relationship between these tools without needing to operate them.
  • Model evaluation metrics: AI-900 is the only Azure Fundamentals-level exam that tests quantitative model evaluation concepts. For classification models, the exam covers accuracy, precision, recall, and the F1 score — specifically, candidates must understand why accuracy alone is misleading for imbalanced datasets (a model that always predicts “not fraud” on a 99% non-fraud dataset achieves 99% accuracy while being useless). For regression models, the exam covers mean absolute error (MAE) and root mean squared error (RMSE), and the intuition that RMSE penalises large errors more heavily than MAE.

Domain 3: Describe features of computer vision workloads on Azure (15–20%)

This domain covers AI services that analyse images and video. Azure’s computer vision portfolio has consolidated significantly under the Azure AI Vision brand, and the exam tests the capabilities of the current service lineup.

  • Image analysis capabilities: Azure AI Vision (the successor to Cognitive Services Computer Vision) can classify images, detect and label objects within images with bounding box coordinates, read text from images and documents (OCR), detect faces and facial attributes, and generate image captions. The exam tests which capability applies to a given scenario and, importantly, which capabilities require a Custom Vision model (where the default service is insufficient for a specific domain) versus the pre-built general-purpose model.
  • Custom Vision: Azure Custom Vision allows organisations to train image classification and object detection models on their own labelled image datasets without writing model code. The exam tests the training data requirements (minimum labelled images per class), the distinction between the training endpoint (used during model development) and the prediction endpoint (used in production applications), and the threshold setting that controls the precision-recall trade-off when the model returns a confidence score.
  • Azure AI Face: the Azure AI Face service detects faces in images, analyses facial attributes (approximate age, emotion, head pose, glasses), and can verify whether two face images are of the same person. The exam tests the distinction between face detection (locating faces in an image), face verification (comparing two specific faces), and face identification (matching a face against a known-person database) — three separate operations the service supports. The exam also tests the responsible AI constraints Microsoft has applied to Face: certain capabilities including face identification require approval through a Limited Access programme, and the service may not be used for surveillance of individuals without consent.
  • Azure AI Video Indexer: a managed service that extracts insights from video content: transcription and translation, speaker identification, OCR from on-screen text, object and scene detection, sentiment analysis of dialogue, and detection of branded content. The exam tests what Video Indexer outputs and in what scenarios it is the appropriate service choice compared to applying individual AI services frame-by-frame.

Domain 4: Describe features of Natural Language Processing workloads on Azure (15–20%)

This domain covers the Azure AI services that process text and speech. The exam tests the Azure AI Language service (the consolidated successor to the individual Text Analytics, LUIS, and QnA Maker services) and Azure AI Speech.

  • Azure AI Language capabilities: Azure AI Language provides a unified API surface for named entity recognition (identifying people, organisations, locations, and dates in text), key phrase extraction (pulling the most important concepts from a text body), sentiment analysis (determining whether a text expresses a positive, negative, neutral, or mixed sentiment), language detection, and linked entity recognition (connecting entities to a knowledge base like Wikipedia). The exam tests which capability applies to a given scenario and how the service returns confidence scores rather than binary judgements, allowing applications to handle ambiguous or borderline cases explicitly.
  • Conversational language understanding (CLU): CLU is the replacement for LUIS (Language Understanding Intelligent Service) and enables applications to understand natural language commands by mapping user utterances to intents and extracting entities. The exam tests the core concepts: intents (what the user wants to do), entities (the specific values relevant to that intent — dates, locations, product names), and utterances (example sentences used to train the model). Candidates must understand the training and deployment cycle for a CLU model and the distinction between built-in entity types (dates, numbers, currencies — recognised without training data) and custom entity types (domain-specific values that require labelled examples).
  • Azure AI Question Answering: the successor to QnA Maker, this service creates a queryable knowledge base from FAQ documents, web pages, or manually entered question-answer pairs. The exam tests what inputs the service accepts, how it ranks candidate answers by confidence, and how it integrates with Azure Bot Service to power customer support bots. The distinction between question answering (answering questions from a fixed knowledge base) and conversational language understanding (interpreting arbitrary commands) is a common exam question type.
  • Azure AI Speech: the Azure AI Speech service covers speech-to-text (real-time and batch transcription, speaker diarisation, and custom acoustic and language models for domain-specific vocabulary), text-to-speech (synthesis with dozens of built-in neural voices and custom voice creation), speech translation (real-time translation between spoken languages), and speaker recognition (verifying or identifying a speaker from audio). The exam tests which Speech capability applies to a given scenario and the distinction between real-time transcription (low-latency, streaming) and batch transcription (high-throughput, offline audio file processing).

Domain 5: Describe features of generative AI workloads on Azure (15–20%)

Added to the blueprint in 2024, this domain reflects the central importance of large language models and the Azure OpenAI Service in the current Azure AI landscape. It is the domain most directly relevant to the AI-powered tooling that IT professionals encounter daily through Microsoft Copilot integrations.

  • Large language models (LLMs): the exam tests foundational concepts about how LLMs work without requiring mathematical depth. Candidates must understand that LLMs are trained on large text corpora and learn to predict the next token in a sequence, that this training produces models capable of text generation, summarisation, translation, question answering, and code generation, and that LLM outputs are probabilistic — the same prompt can produce different responses. The exam tests the concept of a transformer architecture at the level of understanding attention mechanisms (the model attends to relevant parts of the input when generating each output token) without requiring knowledge of implementation details.
  • Azure OpenAI Service: Azure OpenAI Service provides enterprise-grade access to OpenAI models (including GPT-4 and its successors, DALL-E for image generation, and Whisper for speech transcription) via Azure’s managed infrastructure, with Azure-standard security, compliance, and data residency guarantees. The exam tests what Azure OpenAI Service provides compared to using the OpenAI API directly (the answer is enterprise security controls, private networking, Azure Active Directory integration, and data processing agreements) and how Azure AI Studio serves as the development workspace for building and testing generative AI applications on Azure.
  • Prompt engineering concepts: the exam introduces prompt engineering at a conceptual level: the system prompt (instructions that define the model’s persona and constraints), the user prompt (the specific request), few-shot prompting (providing examples within the prompt to guide the model’s output format), and grounding (anchoring model responses to specific retrieved documents to reduce hallucination). Candidates must understand what hallucination is in the context of LLMs (the model generating plausible-sounding but factually incorrect information) and why it occurs — because the model optimises for token prediction probability, not factual accuracy.
  • Responsible AI for generative systems: the exam extends the responsible AI principles from Domain 1 to generative AI-specific risks. Jailbreaking (adversarial prompts designed to bypass safety filters), prompt injection (malicious instructions embedded in retrieved documents that redirect model behaviour), and content filtering (Azure OpenAI Service’s built-in filters that block harmful content categories) are all covered. Candidates must understand the four harm categories Azure’s content filtering addresses (hate speech, sexual content, violence, and self-harm) and what mitigations apply at the model level versus the application level.

Exam format: 45 questions, 60 minutes, Pearson VUE

AI-900 is delivered exclusively through Pearson VUE, either at an authorised test centre or via online remote proctoring. The exam consists of approximately 45 questions, predominantly single-answer multiple choice, with a small number of multiple-select questions and drag-and-drop matching questions pairing scenarios to AI service categories. The pass mark is a scaled score of 700 out of 1000. Microsoft uses scaled scoring, meaning the raw pass threshold varies slightly between exam versions based on statistical calibration, but candidates can plan for approximately 70% of questions correct.

The exam registration fee is $165 USD. Microsoft offers a 50% discount on most Azure Fundamentals exams for students through its academic programme, accessible via a Microsoft Learn student verification. Retake policy: candidates who fail may retake after 24 hours; after a second failure, there is a 14-day wait before each subsequent attempt, with a maximum of five attempts per year. Exam results and the pass/fail decision are available immediately upon completing the exam. Successful candidates receive a badge through Credly and the certification is recorded in their Microsoft Learn profile.

The most common AI-900 failure pattern is treating Domain 2 as conceptually simple because it is labelled “fundamentals.” The machine learning domain has the highest weight (up to 25%) and tests specific knowledge about Azure Machine Learning Designer, AutoML, and model evaluation metrics — candidates who skim this domain expecting intuitive multiple-choice questions and encounter precision/recall questions or Designer-versus-AutoML scenario comparisons are frequently caught off guard. Treat Domain 2 as the exam’s technical core and allocate proportionally more preparation time to it than the domain weight alone suggests.

AI-900 in the 2026 job market: signal vs. noise

AI-900 sits in a class of credentials where its value depends heavily on role context. For purely technical roles — machine learning engineer, data scientist, AI developer — AI-900 is too foundational to serve as a differentiator. Hiring managers in those roles expect AI-102 (Azure AI Engineer Associate) or AZ-104 combined with AI-specific project experience, not an entry-level conceptual survey. Carrying AI-900 as a standalone credential in an experienced ML engineer’s CV can actually signal a mismatch between claimed experience and demonstrated knowledge depth.

Where AI-900 carries genuine signal value in 2026:

Preparing for AI-900: what works in 2026

Microsoft provides the complete official learning path for AI-900 through Microsoft Learn at no cost. The learning path consists of approximately eight modules covering all five exam domains, with interactive exercises and knowledge checks throughout. The total estimated completion time is around ten hours of guided content, making it one of the most efficient study materials available for any Azure certification. For most candidates with a basic technology background, the Microsoft Learn path alone is sufficient preparation for AI-900 — the exam does not go deeper than the official content covers.

The area where candidates benefit most from supplementary preparation is Domain 2 (machine learning fundamentals). The official Microsoft Learn modules cover the Azure ML service tools clearly, but the model evaluation metrics content (precision, recall, F1, RMSE) is relatively brief. Candidates who want confidence on those question types benefit from one additional resource: Microsoft’s own “Introduction to Machine Learning” modules on Learn provide slightly more depth on evaluation metrics than the AI-900-specific path. Spending one to two additional hours on those modules before sitting the exam is worth the time for candidates without a prior ML background.

For Domain 5 (generative AI), the Microsoft Learn “Introduction to Azure OpenAI Service” module is the most relevant supplementary resource. It covers the Azure OpenAI Service architecture, the relationship between models and deployments, and the responsible AI content filtering configuration in more operational detail than the AI-900 path alone, while still remaining accessible without an engineering background.

Preparation tip

The highest-leverage study activity for AI-900 is working through the official Microsoft Learn AI-900 practice assessment, available free in the Microsoft Learn certification exam page. The practice assessment uses the same question format and style as the live exam and covers all five domains. Work through it at the end of your study period, then revisit any domain where you score below 70% with the relevant Learn modules. Most candidates who complete the Microsoft Learn path and score above 75% on the practice assessment pass the live exam on their first attempt. Treat any score below 80% on Domain 2 in the practice assessment as a signal to spend additional time on the ML fundamentals modules before booking the exam.

Next credential after AI-900

Candidates who pass AI-900 and want to move into implementation roles should target AI-102 (Azure AI Engineer Associate) as the next credential. AI-102 covers the same service categories at a configuration and integration depth — building custom speech models, deploying document intelligence solutions, configuring Azure OpenAI deployments, and implementing responsible AI content filters programmatically. AI-102 is the credential that unlocks the Azure AI Engineer job title and the associated salary premium. Candidates pursuing a data science track should consider DP-100 (Azure Data Scientist Associate) instead, which covers Azure Machine Learning at experiment and deployment depth. Both follow naturally from AI-900’s conceptual foundation.

Practice Azure AI and cloud certification questions for free on CertQuests — AI-900, AZ-900, and more.

Browse All Certifications →