Career Roadmap
Azure AI Engineer: Zero to Hero
This roadmap reflects the 2026 Microsoft AI certification transition. AI-102 (Azure AI Engineer Associate) retires June 30, 2026 and is replaced by AI-103 (Azure AI Apps and Agents Developer Associate). AI-900 retires simultaneously and is replaced by AI-901. AI-103 is a fundamentally different exam from AI-102 — it shifts from traditional Azure AI service implementation to building generative AI applications, agentic workflows, and production-ready AI systems on Microsoft Foundry. The five AI-103 domains are - Plan and Manage Azure AI Solutions (25-30%), Implement Generative AI and Agentic Solutions (35-40%), Implement Computer Vision Solutions (10-15%), Implement Text Analysis Solutions (10-15%), and Implement Information Extraction Solutions (10-15%). AI-103 beta launched April 21, 2026 with GA targeted June 2026. Use ExamOS practice quizzes at every step to make progress measurable before each exam attempt.
Embark on your career roadmap by setting a target and staying accountable
Set targetStep 0 - Programming and AI engineering foundations
Build the programming and conceptual foundations that every Azure AI engineering task depends on. AI-103 is a developer exam — it assumes Python fluency and the ability to build real applications, not just call APIs.
3-4 weeks3-4 weeks
Step 0 - Programming and AI engineering foundations
Build the programming and conceptual foundations that every Azure AI engineering task depends on. AI-103 is a developer exam — it assumes Python fluency and the ability to build real applications, not just call APIs.
- Python proficiency — functions, classes, async/await, virtual environments, pip, exception handling, reading SDK documentation
- REST API fundamentals — HTTP methods, authentication headers (API keys, Bearer tokens), request/response patterns, error handling
- JSON and data handling — parsing, serializing, working with nested structures, schema validation
- Git fundamentals — version control for AI projects, branching, managing configuration files securely
- Azure basics — resource groups, subscriptions, the Azure portal, Azure CLI, managed identities, Key Vault for secrets
- AI concepts — what large language models are, tokens, context windows, temperature, system prompts, hallucination
- Responsible AI — Microsoft's six responsible AI principles (fairness, reliability, privacy, inclusiveness, transparency, accountability)
💡 Python fluency is a hard requirement for AI-103. The exam tests whether you can reason about Python SDK code that builds Foundry agents, configures Azure AI Search, and calls evaluation APIs. Candidates who are new to Python should invest 4-6 weeks here rather than 2.
💡 Managed identities are the authentication pattern the exam consistently rewards over API key authentication. Know how to assign a managed identity to an Azure resource and how to grant it role assignments to access AI services without credentials in code.
💡 The responsible AI principles are tested in Domain 1 (Plan and Manage) and are woven through the agentic solutions domain. Know each principle and be able to identify which principle is violated or supported in a described scenario.
Step 1 - AI fundamentals (AI-901)
Build the foundational understanding of generative AI, agents, and Microsoft Foundry that AI-103 builds on. AI-901 is the updated replacement for AI-900 and is built around Foundry rather than individual Azure AI services.
2-3 weeks2-3 weeks
Step 1 - AI fundamentals (AI-901)
Build the foundational understanding of generative AI, agents, and Microsoft Foundry that AI-103 builds on. AI-901 is the updated replacement for AI-900 and is built around Foundry rather than individual Azure AI services.
- Generative AI concepts — how foundation models work, tokens, embeddings, context windows, system prompts
- Microsoft Foundry overview — the unified AI development platform, hubs, projects, model catalog, deployments
- AI agents at a conceptual level — what makes a system agentic, the reasoning loop (think-act-observe), tools, memory
- Multimodal AI — vision models, speech-to-text, text-to-speech, combining modalities in AI applications
- Azure Content Understanding — extracting structured information from documents, images, and audio
- Responsible AI in practice — bias and fairness, explainability requirements, content filtering, human oversight
- AI lifecycle — from model selection through deployment to monitoring and improvement
Certifications
💡 AI-900 retires June 30, 2026 and is replaced by AI-901 (Azure AI Fundamentals, same certification name). If you are starting this path after June 30, prepare for AI-901 directly. AI-901 beta launched April 21, 2026 with GA expected in June 2026.
💡 AI-901 is substantially redesigned around Microsoft Foundry rather than individual Azure AI services as AI-900 was. This makes it more useful as preparation for AI-103 than AI-900 was, because both exams share the Foundry platform context.
💡 40-60 questions, 45 minutes, 700/1000 passing score, no prerequisites, Fundamentals certifications never expire.
💡 Candidates with existing generative AI or Azure AI experience can move through this step in 1-2 weeks. The goal is to confirm conceptual clarity and Foundry familiarity before the deeper technical work begins.
💡 Use ExamOS quizzes to test AI fundamentals and Foundry conceptual understanding before beginning AI-103 domain-specific preparation.
Step 2 - Microsoft Foundry platform and Azure AI architecture
Build hands-on fluency with Microsoft Foundry as the primary development platform for Azure AI engineering. AI-103 is Foundry-centric throughout all five domains — candidates who haven't used Foundry directly will find the exam significantly harder.
3-4 weeks3-4 weeks
Step 2 - Microsoft Foundry platform and Azure AI architecture
Build hands-on fluency with Microsoft Foundry as the primary development platform for Azure AI engineering. AI-103 is Foundry-centric throughout all five domains — candidates who haven't used Foundry directly will find the exam significantly harder.
- Microsoft Foundry architecture — hubs versus projects, connections, the relationship between Foundry and Azure resources
- Azure OpenAI Service in Foundry — model deployments, deployment types (standard versus provisioned), API versions, quota management
- Model catalog in Foundry — selecting between GPT-4o, GPT-4o-mini, Phi-4, Phi-4-mini, Mistral, Llama, Cohere and the criteria determining correct model selection
- Foundry connections — connecting to Azure AI Search, Storage, Cognitive Services, and external data sources from within a project
- Azure AI Foundry SDK — the azure-ai-projects, azure-ai-inference, and azure-ai-evaluation Python packages
- Prompt Flow — prompt engineering workflows, flow types (standard versus chat versus evaluation), flow deployment
- Azure Content Safety — harm categories (hate, violence, sexual, self-harm), severity thresholds, content filter configuration, custom blocklists
- Managed identity authentication — assigning identities to Foundry projects, role assignments for AI Search, Storage, and OpenAI access
💡 AI-103 is the replacement for AI-102, representing Microsoft's strategic shift from traditional Azure AI service implementation toward building modern AI-powered applications, agent workflows, RAG solutions, and production-ready generative AI systems on Azure.
💡 The hardest part of AI-103 is choosing the correct Foundry component for a described scenario. Model deployment versus agent versus Prompt Flow versus AI Search versus Document Intelligence versus Content Understanding — each solves a different problem and the exam is specifically designed to test whether candidates can distinguish between them.
💡 Hands-on time in Foundry is essential and cannot be substituted by documentation study. Create a Foundry hub, deploy a model, build a Prompt Flow, create a Foundry agent with tools, and run an evaluation before your exam.
💡 The Microsoft Spring Skills Challenge 2026 offered an 80% beta exam discount for AI-103. Check Microsoft Learn events for any current discount opportunities.
Step 3 - AI-103 Domain 1 — Plan and manage Azure AI solutions (25-30%)
Design, plan, secure, and govern Azure AI solutions at the project and organization level. Domain 1 at 25-30% covers the strategic and operational layer of AI engineering.
3-4 weeks3-4 weeks
Step 3 - AI-103 Domain 1 — Plan and manage Azure AI solutions (25-30%)
Design, plan, secure, and govern Azure AI solutions at the project and organization level. Domain 1 at 25-30% covers the strategic and operational layer of AI engineering.
- Solution planning — selecting the right Azure AI services for described business requirements, cost versus capability trade-offs
- Azure AI service selection — when Azure OpenAI Service versus Azure AI Services APIs versus custom model versus Foundry agent is the correct choice
- Cost management for AI solutions — token pricing models, provisioned versus consumption-based throughput, estimating costs for different usage patterns
- AI solution security architecture — network isolation for AI services, private endpoints, VNet integration for Foundry projects
- Responsible AI implementation — content filtering configuration, bias testing, impact assessment, model cards
- Monitoring AI solutions — Azure Monitor for AI service metrics, Application Insights for distributed tracing, custom metrics for AI-specific KPIs
- AI solution governance — role-based access in Foundry (owner, contributor, reader), audit logging with Azure Monitor
- Compliance considerations — data residency requirements, GDPR implications for AI data processing, EU AI Act risk tiers
💡 Service selection questions in Domain 1 are specifically designed to test whether candidates default to Azure OpenAI Service when a different service is more appropriate. Azure AI Language for NLP tasks on existing text, Azure AI Vision for pre-built image analysis, Azure AI Speech for audio processing — these are tested against scenarios where a less expensive and more appropriate managed service exists than Azure OpenAI.
💡 Cost management scenarios test token budgeting decisions. Provisioned throughput (reserved capacity, predictable latency, higher baseline cost) versus consumption throughput (pay-per-token, variable latency, no baseline cost) is a design decision that the exam tests against described usage patterns.
💡 The EU AI Act's risk classification (unacceptable, high-risk, limited risk, minimal risk) is specifically tested in responsible AI scenarios. Know which AI application categories fall into each risk tier and what compliance obligations they create.
💡 Use ExamOS for planning and management scenario practice that tests service selection and governance design for described organizational requirements.
Step 4 - AI-103 Domain 2 Part A — RAG and information extraction (10-15% + 10-15%)
Build Retrieval-Augmented Generation systems and information extraction pipelines that ground AI responses in real organizational data. RAG is the foundational pattern tested across both Domain 2 and Domain 5.
4-5 weeks4-5 weeks
Step 4 - AI-103 Domain 2 Part A — RAG and information extraction (10-15% + 10-15%)
Build Retrieval-Augmented Generation systems and information extraction pipelines that ground AI responses in real organizational data. RAG is the foundational pattern tested across both Domain 2 and Domain 5.
- RAG architecture — why RAG exists, the retrieve-then-generate pattern, when RAG versus fine-tuning versus pure prompting
- Azure AI Search — creating indexes, configuring fields, analyzers, vector fields for semantic search
- Vector indexes in AI Search — embedding model selection, embedding dimensions, similarity metrics (cosine versus dot product)
- Hybrid search — combining vector search with BM25 keyword search, reciprocal rank fusion, semantic reranking with semantic configurations
- Document ingestion pipelines — Document Intelligence for PDF and Office extraction, chunking strategies (fixed-size, semantic, sentence-based)
- Foundry RAG integration — connecting Azure AI Search as a grounding tool in Foundry, configuring data source connections
- Azure AI Document Intelligence — prebuilt models (invoice, receipt, contract, identity document), custom models, layout analysis for complex documents
- Azure AI Content Understanding — extracting structured data from unstructured documents with LLM-based extraction
- RAG evaluation — groundedness (is the answer supported by retrieved context?), relevance, coherence using Azure AI Evaluation SDK
💡 Azure AI Search is the primary vector store tested on AI-103. Know how to create a vector index, configure the embedding model, and set up hybrid search with semantic reranking. These specific configuration steps appear in implementation scenarios.
💡 Hybrid search with semantic reranking is the configuration that produces the best retrieval results for most enterprise RAG use cases. Know that semantic ranker re-scores results from the initial hybrid search using a cross-encoder model and that it requires the Standard tier or above in Azure AI Search.
💡 Information Extraction at 10-15% specifically tests Azure AI Document Intelligence prebuilt model selection. Invoice model for invoice data extraction, receipt model for receipt processing, identity document model for ID extraction, custom extraction models when prebuilt models don't match the document type.
💡 RAG evaluation using the Azure AI Evaluation SDK is specifically tested — not just that evaluation exists but what metrics it produces (groundedness score, relevance score, coherence score) and how to interpret them.
Step 5 - AI-103 Domain 2 Part B — Generative AI and agentic solutions (35-40%)
Build generative AI applications and autonomous AI agents using Microsoft Foundry, Semantic Kernel, and multi-agent orchestration patterns. This is the largest and most heavily weighted AI-103 domain at 35-40%.
5-6 weeks5-6 weeks
Step 5 - AI-103 Domain 2 Part B — Generative AI and agentic solutions (35-40%)
Build generative AI applications and autonomous AI agents using Microsoft Foundry, Semantic Kernel, and multi-agent orchestration patterns. This is the largest and most heavily weighted AI-103 domain at 35-40%.
- Azure AI Foundry Agent Service — creating agents, defining instructions, configuring tools (function tools, Code Interpreter, Azure AI Search grounding, Bing grounding)
- Agent memory — thread-based conversation memory in Foundry, external memory patterns for long-context scenarios, memory compression
- Function tools in agents — defining function signatures, JSON schema for tool parameters, handling tool call responses
- Code Interpreter tool — when to use, sandbox environment, file handling, supported operations
- Semantic Kernel — the Microsoft SDK for agent orchestration, plugins, kernel functions, planner patterns, kernel memory
- Multi-agent architectures — orchestrator agents that delegate to specialist sub-agents, agent handoff patterns, parallel agent execution
- AutoGen integration — multi-agent conversation patterns, group chat orchestration, user proxy agents
- Agent evaluation — measuring task completion rate, tool call accuracy, response grounding in multi-turn conversations
- Prompt engineering in Foundry — system prompt design, few-shot examples, chain-of-thought, structured output with JSON schema
- Agent safety and security — prompt injection risks in agentic systems (direct and indirect), input validation, human-in-the-loop controls, Entra Agent ID for non-human identity
💡 Generative AI and Agentic Solutions at 35-40% is the primary AI-103 domain and the one that most distinguishes it from AI-102. Candidates who treat this domain as "prompting and APIs" without understanding agent architecture, multi-agent orchestration, and tool design will miss a significant portion of the exam.
💡 Entra Agent ID is a new concept specific to AI-103 that has no AI-102 equivalent. It extends Azure's workload identity model to cover AI agents — non-human entities that need authenticated access to Azure resources. Know what Entra Agent ID is and why agents need managed identities rather than shared API keys.
💡 Prompt injection in agentic systems is specifically tested as a security topic. Direct prompt injection (user manipulates agent behavior through their input) and indirect prompt injection (agent is manipulated through content it retrieves from external sources) have different attack vectors and different mitigations.
💡 Semantic Kernel is the primary tested SDK for agent orchestration in AI-103. Know how plugins are defined, how kernel functions are invoked, and how the planner selects which functions to call for a given task.
💡 Build at least one multi-agent system in Foundry before your exam. A simple pattern — one orchestrator agent that delegates to a specialist search agent and a specialist summarization agent — is sufficient to develop the intuition the exam tests.
💡 Use ExamOS for agentic solutions scenario practice that tests tool design decisions, multi-agent architecture patterns, and agent safety configuration.
Step 6 - AI-103 Domain 3 and 4 — Computer vision and text analysis (10-15% each)
Implement computer vision and text analysis solutions using Azure AI Vision and Azure AI Language services. These two domains together represent 20-30% of the exam.
3-4 weeks3-4 weeks
Step 6 - AI-103 Domain 3 and 4 — Computer vision and text analysis (10-15% each)
Implement computer vision and text analysis solutions using Azure AI Vision and Azure AI Language services. These two domains together represent 20-30% of the exam.
- Azure AI Vision — image analysis (objects, tags, captions, dense captions), face detection and recognition, spatial analysis, background removal
- Azure AI Vision custom models — image classification, object detection, training data requirements, model evaluation metrics
- Multimodal vision with Azure OpenAI — GPT-4o for visual understanding, combining image and text inputs, video analysis workflows
- Azure AI Language — key phrase extraction, entity recognition (NER), sentiment analysis, opinion mining, language detection
- Azure AI Language custom capabilities — custom NER for domain-specific entities, custom text classification, training data preparation
- Conversational language understanding (CLU) — intents, entities, utterances, model training and evaluation
- Azure AI Translator — document translation, text translation, language detection, custom glossaries
- Azure AI Speech — speech-to-text transcription, text-to-speech synthesis, speaker recognition, pronunciation assessment, custom neural voice
- Question answering — Azure AI Language question answering capability, knowledge base creation, chit-chat responses, active learning
💡 Computer Vision at 10-15% tests service selection between Azure AI Vision prebuilt capabilities and multimodal AI with Azure OpenAI. Know when GPT-4o's visual understanding replaces or augments dedicated vision services — Azure AI Vision for structured feature extraction at scale, GPT-4o for complex visual reasoning requiring language understanding.
💡 Custom NER versus standard NER is a domain selection decision that the exam tests. Standard NER recognizes common entity types (Person, Organization, Location, Date). Custom NER recognizes domain-specific entities (product codes, internal terminology) that prebuilt models don't handle.
💡 Azure AI Speech combined with Azure AI Language is the typical architecture for voice-first applications. Speech-to-text transcribes the audio, Language understanding extracts intent and entities, Text-to-speech delivers the response. Know how these services connect and what the configuration looks like.
💡 Use ExamOS for vision and text analysis scenario practice that tests service selection between prebuilt, custom, and generative AI approaches for described application requirements.
Step 7 - Production AI systems, security, and responsible deployment
Build production-ready AI applications that are secure, monitored, cost-controlled, and compliant with responsible AI requirements. This step consolidates cross-domain production concerns that appear throughout the exam.
2-3 weeks2-3 weeks
Step 7 - Production AI systems, security, and responsible deployment
Build production-ready AI applications that are secure, monitored, cost-controlled, and compliant with responsible AI requirements. This step consolidates cross-domain production concerns that appear throughout the exam.
- Content Safety at production scale — configuring harm thresholds per use case, custom blocklists for domain-specific content, groundedness detection for RAG outputs
- AI application security — prompt injection defense patterns, input validation for LLM inputs, output validation before rendering
- Azure Key Vault integration — storing AI service keys and connection strings, managed identity access pattern versus key-based access
- Cost management for production — token budgeting, prompt compression techniques, model routing (expensive model for complex queries, smaller model for simple ones)
- Latency optimization — streaming responses, caching with Azure Cache for Redis, choosing smaller models for latency-sensitive paths
- AI monitoring in production — tracking token usage per user and per session, latency percentile monitoring, error rate alerts, cost per conversation
- Model version management — pinning API versions, testing before migration to new model versions, handling model deprecations
- Disaster recovery for AI applications — multi-region Azure OpenAI deployments, Foundry project replication, failover configuration
💡 Content Safety configuration appears in Domain 1 (planning) and Domain 2 (implementation) scenarios. Know the four harm categories, the four severity levels (safe, low, medium, high), what each threshold level permits and blocks, and when custom blocklists are needed beyond the built-in categories.
💡 Streaming responses using Server-Sent Events (SSE) appear in implementation scenarios about reducing perceived latency. Know that streaming delivers tokens as they are generated rather than waiting for the complete response, and what the SDK configuration looks like.
💡 Azure Cache for Redis as a semantic cache for LLM responses is tested in cost optimization scenarios. Know what semantic caching is (returning cached responses for semantically similar queries rather than exact-match queries), when it reduces costs effectively, and what cache hit rate implies about cost savings.
💡 Use ExamOS for production AI scenario practice that tests security configuration, cost optimization decisions, and monitoring design for described application requirements.
Step 8 - Exam readiness and follow-on paths
Consolidate AI-103 preparation through integrated scenario practice and identify the follow-on credentials that extend Azure AI engineering capability.
2-3 weeks2-3 weeks
Step 8 - Exam readiness and follow-on paths
Consolidate AI-103 preparation through integrated scenario practice and identify the follow-on credentials that extend Azure AI engineering capability.
- Full scenario practice — working through multi-service scenarios combining Foundry agents, AI Search, content safety, and evaluation
- Domain-weighted practice — ensuring preparation time reflects the 35-40% weight of the agentic solutions domain
- Exam technique — reading the full scenario before evaluating options, identifying the binding constraint first
- Follow-on credential paths — AI-200 (Azure AI Cloud Developer), AI-300 (Machine Learning Operations Engineer), SC-500 (Cloud and AI Security Engineer), AI-901 plus AI-103 plus AI-200 as the complete 2026 Azure AI developer credential stack
💡 AI-200 (Azure AI Cloud Developer Associate, replacing AZ-204) is the natural complement to AI-103. AI-103 covers the AI application and agent layer. AI-200 covers the cloud infrastructure those applications run on (containers, event-driven pipelines, vector data services, observability). Together they form the most complete Azure AI engineering credential profile available in 2026.
💡 AI-300 (Machine Learning Operations Engineer, replacing DP-100 which retires June 30, 2026) covers ML model training, pipeline automation, and MLOps — content that was previously partially covered in AI-102 but is now a separate credential targeted at ML operations roles.
💡 SC-500 (Cloud and AI Security Engineer) is specifically relevant for AI engineers whose role includes securing AI workloads. It covers Defender for AI Service, Entra Agent ID security, prompt injection defenses, and AI governance — directly complementary to the security topics in AI-103.
💡 Consistent performance above 80% on Legend mode across five or more consecutive ExamOS sessions is the clearest AI-103 readiness signal. The Generative AI and Agentic Solutions domain at 35-40% should receive disproportionate attention in the final preparation weeks.
Final step - Certification, validation, and the 2026 transition
The most important decision for Azure AI engineering candidates in 2026 is clear: AI-102 retires June 30, 2026 and should only be pursued if you are already well-prepared and can sit before the deadline. If you are starting fresh, prepare for AI-103 directly. AI-103 is a fundamentally different exam — it validates skills in building generative AI applications and agentic systems on Microsoft Foundry, not traditional Azure AI service implementation. The transition from AI-102 to AI-103 is not an update. It is a signal that Azure AI engineering has moved from services to intelligent systems. Before booking AI-103, ensure you have genuine hands-on Foundry experience across all five domains — especially Foundry Agent Service, Semantic Kernel orchestration, and Azure AI Search RAG integration. Use ExamOS scenario practice to measure readiness objectively. Consistent performance above 80% on Legend mode across multiple sessions, with strength in the agentic solutions domain, is the threshold that reflects genuine exam readiness.
Final step - Certification, validation, and the 2026 transition
The most important decision for Azure AI engineering candidates in 2026 is clear: AI-102 retires June 30, 2026 and should only be pursued if you are already well-prepared and can sit before the deadline. If you are starting fresh, prepare for AI-103 directly. AI-103 is a fundamentally different exam — it validates skills in building generative AI applications and agentic systems on Microsoft Foundry, not traditional Azure AI service implementation. The transition from AI-102 to AI-103 is not an update. It is a signal that Azure AI engineering has moved from services to intelligent systems. Before booking AI-103, ensure you have genuine hands-on Foundry experience across all five domains — especially Foundry Agent Service, Semantic Kernel orchestration, and Azure AI Search RAG integration. Use ExamOS scenario practice to measure readiness objectively. Consistent performance above 80% on Legend mode across multiple sessions, with strength in the agentic solutions domain, is the threshold that reflects genuine exam readiness.
Certifications
Realistic timeline
- 2 hours per day: approximately 5-7 months for the complete path including AI-901 and AI-103
- 3-4 hours per day: approximately 3.5-5 months
- Candidates who already hold AI-900 or have strong Azure AI experience: approximately 8-12 weeks for AI-103 specific preparation
- The Generative AI and Agentic Solutions domain (35-40%) should receive approximately 40% of total AI-103 preparation time — weight your effort proportionally
- Hands-on Foundry time building real agents, RAG pipelines, and evaluation workflows cannot be substituted by reading documentation
- AI-102 versus AI-103 decision: sit AI-102 before June 30, 2026 only if already close to ready; prepare for AI-103 if starting fresh
- AI-200 (Azure AI Cloud Developer, replacing AZ-204) is the recommended follow-on credential to complete the Azure AI developer profile
- Consistency across daily sessions produces better AI-103 outcomes than periodic marathon sessions
Embark on your career roadmap by setting a target and staying accountable
Set target