Article
FREE AI-102 Exam Guide 2026: The Retired Azure AI Engineer Exam (GenAI, Azure OpenAI, Vision, NLP) and AI-103
Key Facts
- Exam AI-102: Designing and Implementing a Microsoft Azure AI Solution earned the Microsoft Certified: Azure AI Engineer Associate credential.
- Microsoft retired Exam AI-102 and the Azure AI Engineer Associate certification on June 30, 2026.
- Exam AI-103: Developing AI Apps and Agents on Azure is the replacement, earning the Azure AI Apps and Agents Developer Associate credential.
- The AI-102 exam fee was $165 USD in the United States, set by Microsoft and delivered through Pearson VUE.
- AI-102 used a scaled passing score of 700 out of 1000, with no published per-item weighting.
- The AI-102 exam gave candidates 100 minutes via Pearson VUE test centers or OnVUE online proctoring; Microsoft did not publish its question count.
- The final AI-102 skills outline (December 23, 2025) covered six domains: plan and manage (20-25%), generative AI (15-20%), agentic (5-10%), vision (10-15%), NLP (15-20%), knowledge mining (15-20%).
- AI-102 had no formal prerequisite; Microsoft's audience profile expected experience building solutions in languages such as Python or C#.
- LUIS (Language Understanding Intelligent Service) was fully retired on March 31, 2026, and Microsoft directs users to Conversational Language Understanding (CLU) in Azure AI Language.
AI-102 Exam Guide 2026: The Azure AI Engineer Credential Built Around Azure OpenAI and Generative AI
Retirement Notice (updated July 2026): Microsoft retired Exam AI-102 and the Azure AI Engineer Associate certification on June 30, 2026 . You can no longer schedule or sit AI-102. Its replacement is Exam AI-103: Developing AI Apps and Agents on Azure , which earns the Microsoft Certified: Azure AI Apps and Agents Developer Associate credential. This guide stays useful because (a) professionals who earned the credential before retirement keep it until it expires (it can no longer be renewed), and (b) AI-103's outline also covers generative AI, RAG, agents, and Microsoft Foundry, with generative AI and agentic solutions weighted 30-35%. Check the AI-103 study guide for its exact objectives before you rely on this material.
Exam AI-102: Designing and Implementing a Microsoft Azure AI Solution changed substantially before it retired. Microsoft revised the skills outline several times to add Azure OpenAI Service , retrieval-augmented generation (RAG) , Microsoft Foundry (formerly Azure AI Studio, then Azure AI Foundry), and agentic workflows, and to rename Azure Cognitive Search to Azure AI Search . The May 2023 outline still had a conversational AI domain; the final outline had dedicated generative AI and agentic domains that older study guides miss.
This guide is written for the final AI-102 exam window (skills measured as of December 23, 2025 , the last revision before the June 30, 2026 retirement). It maps every skill domain against that final outline, calls out the Azure OpenAI / RAG / agentic material, and gives you an 8-week plan and a hands-on lab checklist for the Azure AI Engineer Associate credential - plus transition notes for AI-103.
AI-102 Exam At-a-Glance (2026)
- Item Detail (2026)
- Full Name Exam AI-102: Designing and Implementing a Microsoft Azure AI Solution
- Credential Earned Microsoft Certified: Azure AI Engineer Associate
- Delivery Pearson VUE (online-proctored via OnVUE or test center)
Questions Not published (most Microsoft exams have 40-60); Microsoft's exam sandbox demos question types such as multiple choice, drag and drop, hot area, build list and case studies
- Time Limit 100 minutes of exam time (~120 minutes total seat time)
- Passing Score 700 out of 1000 (scaled)
- Exam Fee $165 USD in the US (varied by country)
Prerequisites None required; Microsoft expected experience building solutions in Python or C# with REST APIs and SDKs
Languages English, Japanese, Chinese (Simplified), Korean, German, French, Spanish, Portuguese (Brazil), Chinese (Traditional), Italian
- Certification Validity 1 year; no renewal (renewal assessment retired June 30, 2026)
- Retake Policy 24-hour wait after 1st fail; 14-day wait for attempts 2-5; max 5 attempts per 12 months
Final Skills Outline December 23, 2025 (minor renames only); six domains, including the agentic solution domain added April 30, 2025 (see below)
Status (July 2026) RETIRED June 30, 2026 ; replaced by AI-103: Developing AI Apps and Agents on Azure (Azure AI Apps and Agents Developer Associate)
Related AI-900 (AI Fundamentals, retired June 30, 2026; replaced by AI-901), DP-100 (Data Scientist, retired June 1, 2026), AZ-305 (Solutions Architect), DP-700 (Fabric Data Engineer), AI-103 (successor)
Source: Microsoft Learn exam page (learn.microsoft.com/credentials/certifications/exams/ai-102), AI-102 skills measured outline (final revision December 23, 2025), retirement announcement, and Pearson VUE scheduling portal.
Start Your FREE AI-102 Prep Today
Jump into scenario-based Azure AI questions covering every skill domain - including the generative AI and Azure OpenAI chapters - with instant AI explanations. Start FREE AI-102 practice questions for the Azure AI Engineer Associate exam . 100% FREE.
Who AI-102 Was For (and Who AI-103 Is For)
AI-102 is retired, and its target audience overlaps with the replacement AI-103 exam. The final AI-102 skills outline targeted developers who build AI solutions using Azure services , including the Azure OpenAI Service, Azure AI services (formerly Cognitive Services), Azure AI Search, Azure AI Document Intelligence, Azure AI Vision, Azure AI Language, Azure AI Speech, and Microsoft Foundry (formerly Azure AI Studio / Azure AI Foundry) for generative and agentic workloads.
Candidate Profile Why AI-102 Fits
Python or C# developers adding AI to existing apps Core target audience: the audience profile named Python and C# with REST APIs and SDKs
Full-stack engineers building RAG apps on Azure OpenAI Generative AI chapter covers the exact patterns you ship
ML engineers with Databricks/PyTorch experience Validates the Azure-specific deployment, security, and responsible AI layer
- Cloud engineers moving into AI roles Adds AI to your Azure credential stack (complements AI-200, AZ-305)
- Data engineers integrating AI with pipelines Pairs with DP-700 (Fabric Data Engineer) for data pipelines
- Solution architects designing Azure AI platforms Provides hands-on depth beyond AI-900 fundamentals
- Recommended Prerequisites (Not Required)
AI-102 had no formal prerequisites . Microsoft's audience profile described an Azure AI engineer who:
Has experience developing solutions in languages such as Python or C#
Can use REST APIs and SDKs to build secure image processing, video processing, natural language processing, knowledge mining, and generative AI solutions
- Understands the components of the Azure AI portfolio and the available data storage options
- Can apply responsible AI principles
This guide also recommends basic Azure portal, Azure CLI, and Bicep/ARM navigation, and a conceptual grasp of embeddings and vector search.
If you are brand new to Azure AI, start with AI-901 (Azure AI Fundamentals, $99 in the US), which replaced AI-900, before moving on to AI-103.
Build AI-102 Mastery with FREE Practice Questions
Access FREE AI-102 practice questions covering the final AI-102 skills outline -Azure OpenAI, AI Search, Vision, NLP, Document Intelligence, and generative AI - with AI-powered explanations. 100% FREE .
AI-102 Skills Measured (2026 Domain Weights)
The final Microsoft Learn skills outline for AI-102 (as of December 23, 2025 ) splits the exam into six weighted domains . The domains and weights date from the April 30, 2025 revision, which added the agentic solution domain; the December 23, 2025 revision changed only names, such as Azure AI Foundry to Microsoft Foundry .
Domain Final Weight (Dec 23, 2025) What It Covers
1. Plan and manage an Azure AI solution 20-25% Microsoft Foundry service selection, deployment, security (Entra ID), responsible AI, content moderation, monitoring, cost
2. Implement an agentic solution 5-10% Microsoft Foundry Agent Service, Microsoft Agent Framework, multi-agent orchestration, testing and deploying agents
3. Implement computer vision solutions 10-15% Azure AI Vision (image analysis, OCR/Read), Custom Vision, Video Indexer, Spatial Analysis (retired March 30, 2025)
4. Implement natural language processing solutions 15-20% Azure AI Language (key phrases, entities, sentiment, language detection, PII, CLU, Custom QA), Translator, Speech (STT/TTS/SSML/translation)
5. Implement knowledge mining and information extraction solutions 15-20% Azure AI Search (vector + hybrid + semantic), Document Intelligence, Content Understanding
6. Implement generative AI solutions 15-20% Azure OpenAI in Foundry Models, prompt engineering, RAG, prompt flow (retiring April 20, 2027), fine-tuning, evaluation, Microsoft Foundry hubs/projects
Sections below are presented in a study-friendly sequence; Microsoft's official outline lists generative AI as Domain 2 and agentic as Domain 3. Weights are identical either way.
IMPORTANT 2026 CAVEAT: AI-102 is now retired (June 30, 2026). The weights above are from the final December 23, 2025 outline, whose change log marks all six domains "No change" : the edits were minor renames (Azure AI Foundry to Microsoft Foundry; Azure AI Agent Service, Semantic Kernel and AutoGen to Microsoft Foundry Agent Service and Microsoft Agent Framework; Document Intelligence to "Azure Document Intelligence in Foundry Tools"). The big change came earlier, in the April 30, 2025 outline: compared with the October 31, 2024 outline, it added the "Implement an agentic solution" domain (5-10%), raised generative AI from 10-15% to 15-20% , and removed the separate "Implement content moderation solutions" domain (10-15%); content moderation moved under Domain 1's responsible AI objectives. For the successor exam, see the AI-103 study guide on learn.microsoft.com.
Source: Microsoft AI-102 Study Guide, final revision dated December 23, 2025.
Domain 1: Plan and Manage an Azure AI Solution (20-25%)
This domain is architecture and operations , not model code.
Select the appropriate Azure AI service
Choose between single-service vs multi-service (Azure AI services) resources for cost and key management.
Decide between Azure AI services, Azure OpenAI , Azure Machine Learning , and Microsoft Foundry based on workload (prebuilt vs custom vs generative).
Understand that model availability varies by region and deployment type; Microsoft publishes it on the Foundry Models availability pages.
Plan, configure, and secure an AI solution
Configure authentication : key-based vs Microsoft Entra ID (formerly Azure AD) with Managed Identity and Key Vault.
Implement private endpoints and VNet integration for data exfiltration protection.
Apply Customer-Managed Keys (CMK) with Key Vault for encryption at rest.
Enable diagnostic logging to Log Analytics, Event Hub, or storage.
Rotate keys programmatically and use Key Vault references in apps.
Implement responsible AI
Apply the six principles of the Microsoft Responsible AI Standard (fairness, reliability and safety, privacy and security, inclusiveness, transparency, accountability).
Configure content filtering in Azure OpenAI (default medium-severity threshold, blocklists, Prompt Shields, protected material detection).
Use Azure AI Content Safety for image/text moderation (hate, sexual, violence, self-harm) and prompt shield.
Implement responsible AI metrics and human review workflows.
Manage, monitor, and optimize cost
Use Azure Monitor and Application Insights for telemetry.
Track token usage and throughput units (PTUs) in Azure OpenAI.
Choose between pay-as-you-go and Provisioned Throughput Units (PTUs) for Azure OpenAI.
Understand deployment types: Global Standard , Data Zone Standard and Standard (pay per token); Global , Data Zone and Regional Provisioned (reserved PTUs); Global and Data Zone Batch (asynchronous, 50% discount); and Developer for evaluating fine-tuned models.
Deploy Azure AI services in containers
Deploy eligible services (Text Analytics, Language, Translator, Speech, Document Intelligence, AI Vision Read) as Docker containers on-prem or in Azure Kubernetes Service for data residency.
Domain 2: Implement an Agentic Solution (5-10%) - Added in the April 30, 2025 Outline
This was the newest domain : it first appeared in the April 30, 2025 outline (with Azure AI Agent Service, Semantic Kernel and AutoGen as its tools) and kept its name and 5-10% weight in the final December 23, 2025 revision, which renamed that tooling Microsoft Foundry Agent Service and Microsoft Agent Framework. It reflects Microsoft's shift toward agent-based generative AI workloads. Expect items on the Microsoft Foundry Agent Service and Microsoft Agent Framework.
Microsoft Foundry Agent Service : managed agents with built-in tools such as web search, file search and code interpreter, plus custom tools through functions, OpenAPI specs and MCP servers. Microsoft retired the Azure OpenAI Assistants API and points users to Foundry Agent Service.
Microsoft Agent Framework : build complex, multi-agent solutions with orchestration across multiple users and autonomous capabilities.
Agent design : understand the role and use cases of an agent, configure the necessary resources, and create an agent in Foundry.
Test, optimize, and deploy an agent including orchestration for multi-agent solutions.
Note (content moved from the old content moderation domain): The October 31, 2024 AI-102 outline had a 10-15% "Implement content moderation solutions" domain. From the April 30, 2025 outline onward, that content sits inside Domain 1 ("Implement AI solutions responsibly"). The key items to know:
Azure AI Content Safety : text moderation (hate/sexual/violence/self-harm categories with severity 0/2/4/6), image moderation, Prompt Shields (jailbreak + indirect prompt injection detection), Protected Material detection (known text such as song lyrics, articles and recipes, plus code from known GitHub repositories), Groundedness detection (flags response text the grounding sources don't support).
Personalizer - contextual bandit reinforcement learning for content ranking; no new resources since September 20, 2023, and Microsoft retired the service on August 25, 2026.
Anomaly Detector - no new resources since September 20, 2023; Microsoft set its retirement for October 1, 2026, and it is not in the final outline.
- Domain 3: Implement Computer Vision Solutions (10-15%)
- Azure AI Vision (formerly Computer Vision; now Azure Vision in Foundry Tools)
Image Analysis 4.0 : captioning, dense captioning, OCR/Read, tags, object detection, people detection, smart crops, image retrieval via multimodal embeddings (background removal was retired on March 31, 2025). Microsoft has deprecated Image Analysis and will retire it on September 25, 2028.
Read API / OCR 4.0 for printed + handwritten text.
Face API : face detection works without registration; Face identification and verification are Limited Access features that need an approved registration.
Spatial Analysis (container) for detecting the presence and movement of people in video: still an objective in the final outline, but Microsoft retired Spatial Analysis on March 30, 2025.
Custom Vision
Build custom classification (multi-class or multi-label) and object detection models; Custom Vision needs at least 5 labeled images per tag for classification and 15 for object detection, and Microsoft recommends 50 or more.
Export compact-domain models to ONNX, TensorFlow, CoreML or a Docker container for edge deployment.
Microsoft will retire Custom Vision on September 25, 2028.
Understand iterations and performance metrics (precision, recall, mAP).
Azure AI Video Indexer
Extract insights from video (speakers, topics, OCR, faces, sentiment, keywords, labels).
Use custom models for language, brands, people.
Integrate via widget embedding or API-driven pipelines.
- Domain 4: Implement Natural Language Processing (15-20%)
- Azure AI Language (formerly Text Analytics + LUIS + QnA Maker)
Prebuilt features: sentiment analysis , opinion mining , key phrase extraction , language detection , PII detection + redaction, Named Entity Recognition (NER) , text summarization (extractive + abstractive), healthcare text analytics .
Custom models : custom NER, custom text classification (single + multi label), Conversational Language Understanding (CLU) (the successor to LUIS), Custom Question Answering (successor to QnA Maker), orchestration workflow for chaining CLU + QA.
Azure AI Translator
Text translation (100+ languages), document translation (with optional glossaries), Custom Translator for domain-specific translation models, transliteration.
Azure AI Speech
Speech-to-Text (STT) : real-time, batch, fast transcription, custom speech models, pronunciation assessment.
Text-to-Speech (TTS) : neural voices, custom neural voice (Limited Access), SSML for prosody.
Speech translation : speech-to-text and speech-to-speech translation.
Retired Speech features : Speaker Recognition (voice identification and verification) was retired on September 30, 2025. Intent recognition in Azure Speech was retired the same day, although the final outline still lists "intent and keyword recognition"; Microsoft points to Azure Language or Azure OpenAI for intents.
Note: LUIS was fully retired on March 31, 2026 (its portal closed on October 31, 2025). Any study guide mentioning LUIS as current content is out of date. Microsoft directs LUIS users to Conversational Language Understanding (CLU) . CLU and Custom Question Answering themselves retire from Azure Language on March 31, 2029, and Microsoft directs new projects to Microsoft Foundry.
- Domain 5: Implement Knowledge Mining and Information Extraction (15-20%)
- Azure AI Search (formerly Azure Cognitive Search)
Create indexes , indexers , data sources , skillsets for AI enrichment.
Built-in cognitive skills : OCR, key phrase extraction, entity recognition, translation, image analysis.
Custom skills : call Azure Functions or custom APIs from a skillset.
Vector search : create vector fields, embed documents via Azure OpenAI text-embedding-3-large / text-embedding-ada-002 , perform hybrid search (keyword + vector), use semantic ranker for reranking.
Integrated vectorization : chunking and embedding inside the indexer pipeline, using the AzureOpenAIEmbedding skill (text-embedding-ada-002, text-embedding-3-small or text-embedding-3-large).
Security trimming using user group membership or Entra ID.
Azure AI Document Intelligence (formerly Form Recognizer)
Prebuilt models : invoices, receipts, ID documents, W-2s, 1098s, 1099s, health insurance cards, US bank checks and bank statements, pay stubs, contracts and US mortgage forms (the business card model is deprecated in v4.0).
Layout model for tables + structure.
Custom models : custom template (consistent visual template; one model per layout variation; trains in 1 to 5 minutes) vs custom neural (structured, semi-structured and unstructured documents; one model for all variations; trains in 30 minutes to 12 hours). Microsoft recommends starting with custom neural when it meets your needs.
Composed models to group up to 500 custom models behind one model ID (v4.0 routes documents with an explicitly trained classifier).
Microsoft says you need only five labeled examples of the same form or document type to get started with either custom extraction model.
Domain 6: Implement Generative AI Solutions (15-20%) - The Highest-Value Chapter
This was one of the growing areas of AI-102: its official weight rose from 10-15% to 15-20% in the April 30, 2025 outline and stayed there in the final December 23, 2025 revision. Generative AI concepts also bleed into Domains 1 (responsible AI, content filtering, deployment types), 2 (agents), and 5 (RAG with Azure AI Search).
Azure OpenAI Service
Provision and deploy models : chat and reasoning models (such as the GPT-4o family), embeddings (text-embedding-3-large, text-embedding-3-small, text-embedding-ada-002) and image models (the final outline names DALL-E; Azure now lists gpt-image models). Check the model retirement schedule before you pick a model version.
Choose a deployment type : Global Standard (Microsoft's default starting point: lowest price, highest default quota, broadest model availability), Data Zone Standard or regional Standard for data residency, Provisioned (reserved PTU capacity), Global or Data Zone Batch (24-hour target turnaround at 50% less cost than Global Standard).
Apply content filtering (hate, sexual, violence, self-harm with severity thresholds, plus prompt shields and protected material detection).
Integrate via Azure OpenAI SDK (Python, C#, Java, JavaScript) or REST; use Entra ID with Managed Identity for keyless authentication.
Prompt engineering
System messages for role priming, grounding, few-shot examples.
Parameter tuning: temperature (0-2), top_p , presence_penalty , frequency_penalty , max_tokens , stop sequences , seed (best-effort reproducible outputs).
Chain-of-thought (CoT) prompting.
Function / tool calling : JSON schema definitions, tool choice, parallel calls.
Structured outputs via JSON mode or strict JSON schema.
Retrieval-Augmented Generation (RAG)
End-to-end pattern: chunk documents -> embed -> index in Azure AI Search -> retrieve with hybrid search + semantic ranker -> pass top-K chunks into a grounded prompt -> generate response with citations.
Azure OpenAI On Your Data ("Add your data") wired RAG to Azure AI Search, Blob Storage, uploaded files, URLs, Azure Cosmos DB for MongoDB, Elasticsearch or MongoDB Atlas. Microsoft retired it on October 14, 2026 and points Azure AI Search users to Foundry IQ knowledge bases.
Chunking strategies : fixed size, sentence-based, semantic chunking.
Evaluation : groundedness, relevance, coherence and fluency evaluators in Microsoft Foundry.
Microsoft Foundry (formerly Azure AI Studio / Azure AI Foundry)
Create projects for AI development teams. The final outline still says "deploy a hub, project"; Microsoft now documents hub-based projects as the classic option and suggests moving to Foundry projects.
Use the model catalog (Azure OpenAI + Meta Llama + Mistral + Cohere + custom) with serverless API deployments.
Build agents with Microsoft Foundry Agent Service (formerly Azure AI Agent Service), using tools such as web search, file search, code interpreter and functions.
Build prompt flows (an outline objective). Microsoft will retire prompt flow on April 20, 2027, no longer recommends it for new development, and points to Microsoft Agent Framework.
Monitor with AI-assisted metrics and safety evaluators .
Fine-tuning
Supervised fine-tuning (SFT), direct preference optimization (DPO) and reinforcement fine-tuning (RFT); support varies by model (for example SFT and DPO on GPT-4o and GPT-4.1, SFT on GPT-4o mini, RFT on o4-mini).
Data format: JSONL with messages arrays.
Evaluate vs base model, deploy as dedicated model.
Azure OpenAI Quick Reference (High-Yield for 2026)
Concept What to Know
PTU vs Standard PTU = reserved capacity billed per PTU per hour whether or not you send tokens, with predictable latency; a fully used PTU deployment returns 429s (spillover can send overflow to a standard deployment). Standard = pay-per-token with TPM/RPM quotas
Global Standard May process requests in any Azure region; Microsoft's default starting point, with the lowest price, highest default quota and broadest model availability
Global Batch Async file-based jobs, 24-hour target turnaround, 50% less cost than Global Standard - for large jobs that are not time-sensitive
Content filters Default on, blocking medium and high severity; any customer can set the threshold to low, medium, or high per category; turning filters off on completions requires approval
Prompt Shield Detects user-injected jailbreaks and indirect prompt injection in grounded documents
Protected Material Flags known text such as song lyrics, articles and recipes (text) or code from known GitHub repositories (code)
Groundedness Checks whether a response is supported by the provided sources and flags ungrounded segments; reasoning mode explains them, and an optional correction feature (preview) rewrites them
Entra ID auth Keyless alternative to API keys; use DefaultAzureCredential or Managed Identity
Responsible AI Limited Access Face identification and verification, custom neural voice and personal voice require approved registration (Speaker Recognition was retired on September 30, 2025)
Languages and SDKs on the Exam
AI-102's audience profile expected Python or C# experience with REST APIs and SDKs, and the outline asked you to "install and utilize the appropriate SDKs and APIs". The main packages:
Python ( azure-ai-textanalytics , azure-ai-vision-imageanalysis , azure-ai-documentintelligence , openai (Azure OpenAI), azure-search-documents , azure-ai-language-conversations , azure-ai-ml ).
C# / .NET ( Azure.AI.OpenAI , Azure.AI.Vision.ImageAnalysis , Azure.AI.TextAnalytics , Azure.AI.Language.Conversations , Azure.AI.DocumentIntelligence , Azure.Search.Documents ).
Practice reading SDK code until you can spot the right client, parameter and authentication pattern quickly.
- Example Code Snippet (Azure OpenAI - Python) from openai import AzureOpenAI
- from azure.identity import DefaultAzureCredential, get_bearer_token_provider
- token_provider = get_bearer_token_provider(
- DefaultAzureCredential(),
- "https://cognitiveservices.azure.com/.default",
- )
- client = AzureOpenAI(
- azure_endpoint="https://my-aoai.openai.azure.com/",
- api_version="2024-10-21",
- azure_ad_token_provider=token_provider,
- )
- response = client.chat.completions.create(
- model="gpt-4o", # deployment name
- temperature=0.2,
- response_format={"type": "json_object"},
- messages=[
- {"role": "system", "content": "Return strictly valid JSON."},
- {"role": "user", "content": "Extract invoice fields from: ..."},
- ],
- )
Example Code Snippet (Azure AI Search - Hybrid Vector Query, Python) from azure.search.documents import SearchClient
from azure.search.documents.models import VectorizedQuery
- vector_query = VectorizedQuery(
- vector=embedding, # 3072-dim from text-embedding-3-large
- k_nearest_neighbors=10,
- fields="content_vector",
- )
- results = search_client.search(
- search_text="privacy policy for minors",
- vector_queries=[vector_query],
- query_type="semantic",
- semantic_configuration_name="default",
- top=5,
- )
- 8-Week AI-102 Study Plan (Working Developer)
AI-102 can no longer be booked. The plan below is kept as a hands-on skills plan for the final AI-102 outline; if you are preparing for AI-103, compare each week with the AI-103 study guide first. It assumes ~8-10 hours per week and that you already know Python or C# plus basic Azure.
Week Focus Deliverable
Week 1 Azure AI services overview, resource creation, auth (key + Entra ID), Key Vault, private endpoints Deploy an Azure AI services multi-service + Azure OpenAI resource with Managed Identity
Week 2 Azure AI Vision (Image Analysis 4.0, OCR, Face), Custom Vision, Video Indexer Build an image classifier + OCR pipeline on sample receipts
Week 3 Azure AI Language (NER, sentiment, PII, summarization), Custom NER, CLU, Custom QA, Speech, Translator Build a CLU bot + custom QA knowledge base
Week 4 Document Intelligence (prebuilt + custom template + custom neural), composed models Train a custom neural model for mixed-layout invoices
Week 5 Azure AI Search (vector + hybrid), embeddings, integrated vectorization, semantic ranker Index 500 docs, run hybrid semantic vector query
Week 6 Azure OpenAI + prompt engineering + function calling + structured outputs Build a JSON-schema-grounded assistant with tool calls
Week 7 RAG end-to-end: Azure AI Search retrieval + Foundry evaluators + Foundry Agent Service Build + evaluate a RAG chatbot with groundedness metrics
Week 8 Responsible AI, content filters, cost/PTU, full-length timed mocks, remediation Score consistently >75% on two timed mocks before test day
Time Allocation (Match the Final Blueprint)
- Domain Weight Share of Study Time
- Plan and manage AI solution 20-25% 20%
- Generative AI (includes RAG) 15-20% 22%
- Natural language processing 15-20% 16%
- Knowledge mining / information extraction 15-20% 16%
- Computer vision 10-15% 13%
- Agentic solution 5-10% 13%
- Recommended AI-102 Resources (FREE-First)
Resource Type Why It Helps
ExamFinch AI-102 Practice (FREE) Free Scenario items written against the final December 23, 2025 skills outline, with explanations
Microsoft Learn self-paced training Free Learning paths and modules linked from the AI-102 and AI-103 study guides
Microsoft AI-102 Study Guide (aka.ms/AI102-StudyGuide) Free web page Final skills measured and change log
Azure free account Free $200 credit to use within 30 days plus 12 months of free monthly amounts for 20+ popular services
John Savill's Technical Training (YouTube) Free An older AI-102 Study Cram and an AI-103 Study Cram
MicrosoftLearning lab repos on GitHub (mslearn-ai-vision, mslearn-ai-language, mslearn-ai-studio, mslearn-ai-agents, mslearn-ai-information-extraction) Free Lab files for the Microsoft Learn AI modules
- Azure OpenAI samples (Azure-Samples/openai) Free Microsoft's collection of Azure OpenAI code samples
- Azure-Samples/azure-search-openai-demo Free Sample RAG app running in Azure with Azure AI Search
- Hands-On Labs Are Non-Negotiable
AI-102 was not a memorization exam: its objectives asked you to build, deploy, and configure solutions, and the same is true of AI-103. Use an Azure free account ($200 credit within 30 days + 12 months of free monthly amounts for popular services) or your employer's subscription to build:
An Azure OpenAI chat app with Entra ID auth, content filters, function calling, and structured JSON outputs.
An Azure AI Search index populated via integrated vectorization from a Blob container of 100+ PDFs.
A RAG app that retrieves from that search index and generates cited answers with Azure OpenAI, checked with Foundry's groundedness evaluator (Azure OpenAI On Your Data was retired on October 14, 2026).
A Document Intelligence custom neural model trained on mixed-layout invoices, plus a composed model that uses a trained classifier to route documents to it or to a second custom model.
A Custom Vision object detection model exported to ONNX.
An Azure AI Language CLU + Custom Question Answering orchestration bot deployed to a web app.
A Speech translation app using the Speech SDK with a prebuilt neural voice for output (custom neural voice needs Limited Access approval).
Common Pitfalls That Sink First-Time Scores
Studying an old (pre-2024) outline. Guides that still focus on LUIS, QnA Maker (old), Anomaly Detector as first-class, or skip Azure OpenAI entirely are out of date. For AI-103, check its study guide on learn.microsoft.com before you book.
Under-studying Azure OpenAI. With the final outline weight of 15-20% for generative AI plus its overlap into Domains 1 (responsible AI, content filtering), 2 (agents), and 5 (RAG), budget at least 25% of study time here.
Skipping Azure AI Search vector + hybrid + semantic. Integrated vectorization, text-embedding-3-large, hybrid queries, and semantic ranker configuration are high-yield in Domain 5.
Confusing custom template vs custom neural in Document Intelligence. Template = one consistent visual layout, a model per layout variation, 1 to 5 minutes to train. Neural = structured, semi-structured or unstructured documents, one model for all variations, 30 minutes or more to train, and Microsoft's suggested starting point. Both can start with five labeled examples, and in v4.0 both extract selection marks, tables and signatures.
Missing Entra ID / Managed Identity auth. Key-based auth still works, but the outline lists both "Manage and protect account keys" and "Manage authentication for a Microsoft Foundry Service resource", so know Managed Identity, Entra ID and Key Vault.
Ignoring Limited Access registrations. Face identification and verification and custom neural voice require Microsoft approval (Speaker Recognition was retired on September 30, 2025). Know which features need registration.
Case study time management. Read the question first, then skim the case for the specific detail it asks about.
Weak on PTU vs Standard vs Global Standard vs Global Batch. The outline asks you to deploy models using the appropriate deployment options, so know the cost, latency, and capacity trade-offs and when each deployment type wins.
Not knowing content filter tiers and Prompt Shield. Default medium-severity threshold, customer-configurable severity, Prompt Shield vs Protected Material vs Groundedness. Easy points if memorized.
No timed full-length practice. 100 minutes is tight when case studies appear; two timed mocks minimum before test day.
Test-Day Logistics and Strategies
AI-102 can no longer be booked. The steps below follow Microsoft's current policies for its role-based exams, including AI-103.
Before you sit:
Confirm your Microsoft Learn Profile matches your government ID exactly.
If online-proctored, run the OnVUE system test on the computer and network you will use, before exam day.
Clear your desk: you cannot access unauthorized materials, even during a break.
During the exam:
You can mark questions for review, but once you start a break you cannot return to questions you saw before it, and the exam clock keeps running during breaks.
For case studies, read the question first, then skim the case for the relevant detail.
When two answers look defensible, pick the Microsoft-native, managed-service option with least operational overhead (e.g., integrated vectorization beats hand-rolled embedding code; Managed Identity beats rotating keys).
Role-based exams let you open Microsoft Learn in a split screen, but no extra time is added, so use it only for a lookup you need.
After the exam:
Microsoft says results for most exams arrive within minutes of finishing, with a score report (scaled 1-1,000, pass = 700) and performance feedback.
Your score appears in your Microsoft Learn profile within 24 hours.
If you fail: 24-hour wait for first retake, then 14-day wait for attempts 2-5, max 5 per 12 months.
- Deep Dives on Five Topics Competitor Guides Skim
- 1. Azure OpenAI Deployment Types (Standard vs PTU vs Global)
Microsoft Foundry groups deployment types into three categories - standard (pay per token), provisioned (reserved capacity) and batch (discounted asynchronous processing) - plus a Developer type for evaluating fine-tuned models. Within standard and provisioned, you choose global, data zone or Azure-geography processing. Knowing when each wins is a core Domain 1 / Domain 6 cross-over concept.
Global Standard : pay per token; requests may be processed in any Azure region. Microsoft's recommended starting point for most workloads, with the lowest price, highest default quota and broadest model availability. At high volume, latency can vary more.
Data Zone Standard / Standard : pay per token, with processing kept inside a US, EU or APAC data zone, or inside one Azure geography, for data-residency needs. Quotas are measured in TPM (tokens per minute) and RPM (requests per minute).
Provisioned (Global, Data Zone or Regional) : you reserve PTUs (each PTU is a fixed amount of model processing capacity), billed per PTU per hour whether or not you use them, or through Azure reservations. You get predictable throughput and latency, but a fully used deployment returns 429 errors; spillover can route the overflow to a standard deployment. Good for steady, high-volume production apps.
Global Batch / Data Zone Batch : async file-based jobs with a 24-hour target turnaround at 50% less cost than Global Standard. Good for large jobs that are not time-sensitive, such as document summarization backfills or embedding generation.
2. Retrieval-Augmented Generation (RAG) End-to-End
RAG is the core generative pattern in the outline ("Implement a RAG pattern by grounding a model in your data"). The canonical pipeline:
Ingest documents from a source an Azure AI Search indexer supports, such as Azure Blob Storage, Azure Cosmos DB, Microsoft OneLake or SharePoint in Microsoft 365 (preview).
Chunk using fixed size or semantic chunking. Microsoft suggests starting with 512-token chunks and a 25% (128-token) overlap.
Embed each chunk with text-embedding-3-large (3072 dims) or text-embedding-3-small (1536 dims). Integrated vectorization does this automatically.
Index in Azure AI Search with keyword + vector + semantic fields.
Retrieve the top results with a hybrid query (BM25 + vector, merged with Reciprocal Rank Fusion), then rerank the top 50 with the semantic ranker.
Compose a grounded prompt with retrieved chunks as context and explicit citation instructions.
Generate with GPT-4o / GPT-4o mini.
Evaluate groundedness, relevance, coherence and fluency with Foundry evaluators.
Azure OpenAI On Your Data used to wire steps 4-7 automatically, but Microsoft retired it on October 14, 2026. For an Azure AI Search-backed app, Microsoft's migration path is a Foundry IQ knowledge base (optionally with Foundry Agent Service), or classic RAG in which your application generates the answer.
3. Content Filtering, Prompt Shields, and Groundedness
Azure OpenAI and Azure AI Content Safety give you four families of safety checks:
Content filters across hate, sexual, violence, self-harm with severity 0/2/4/6; medium and high severity are blocked by default, while low severity passes. Any customer can adjust the threshold (low, medium, or high); approval is required only to turn filters partially or fully off on completions.
Prompt Shield - detects user prompt attacks (jailbreaks) and document attacks (indirect prompt injection hidden in documents, emails or web pages). In Foundry, user prompt attack detection is on by default and document attack detection is off until you enable it.
Protected Material detection - flags known text such as song lyrics, articles and recipes (text) or code from known GitHub repositories (code).
Groundedness detection - checks whether a response is supported by the provided sources and flags ungrounded segments; reasoning mode explains them, and an optional correction feature (preview) rewrites them.
Azure AI Content Safety is the standalone service that exposes these capabilities plus image moderation and custom category models.
4. Microsoft Foundry, Agents, and Prompt Flow
Microsoft Foundry (rebranded from Azure AI Studio, then Azure AI Foundry) is the unified workspace for AI development:
Hubs are the container for hub-based projects, now needed only for selected use cases; Microsoft now documents them as the classic option and suggests moving to Foundry projects.
Projects are the unit of AI development (evaluations, agents, deployments).
Model catalog exposes Azure OpenAI, Meta Llama, Mistral, Cohere, Phi, and custom registered models with one-click serverless or managed deployment.
Prompt flow is a visual DAG-based tool for prompt orchestration, evaluation, and deployment. Nodes call LLMs, Python scripts, Azure AI Search, etc. Microsoft will retire prompt flow on April 20, 2027 and points to Microsoft Agent Framework for new work.
Microsoft Foundry Agent Service (formerly Azure AI Agent Service) provides managed agents with tools such as web search, file search, code interpreter and functions; Microsoft retired the Assistants API and points users here.
5. Fine-Tuning vs Prompt Engineering vs RAG
Candidates frequently over-apply fine-tuning. The exam rewards picking the right tool:
Prompt engineering (system prompts + few-shot examples) is the first thing to try: it needs no training job and is fast to iterate.
RAG solves knowledge freshness and grounding - the model cites recent proprietary data it was never trained on.
Fine-tuning solves style, tone, structure, and token efficiency - when you need a compressed instruction footprint or a specific output format. Use SFT or DPO on supported models such as GPT-4o and GPT-4.1 (GPT-4o mini supports SFT). Requires high-quality JSONL training data, and Microsoft notes that fine-tuning doesn't replace retrieval for current information.
The canonical exam answer to "how should I reduce hallucinations about our internal HR policies" is RAG , not fine-tuning. "How should I enforce a strict company-specific JSON schema output" can go either way but often starts with structured outputs + prompt engineering before fine-tuning.
How AI-102 Fits into the Broader Microsoft AI / Data Certification Path
- Exam Role When to Sit
- AI-900 Azure AI Fundamentals Entry Retired June 30, 2026; replaced by AI-901
- AI-102 Azure AI Engineer Associate AI Engineer Retired June 30, 2026 ; see AI-103 below
AI-103 Azure AI Apps and Agents Developer Associate AI App/Agent Developer Successor to AI-102; Microsoft Foundry + agents focus
- DP-100 Azure Data Scientist Associate Data scientist Retired June 1, 2026; replaced by AI-300
- DP-900 Azure Data Fundamentals Entry Optional foundation
- DP-700 Fabric Data Engineer Associate Data engineer Fabric data engineering alongside an AI credential
AZ-305 Azure Solutions Architect Expert Architect Requires the Azure Administrator Associate (AZ-104) certification
After AI-102, the current Microsoft AI path runs through AI-103 , AI-300 (which replaced DP-100) and, for data work, DP-700 .
Renewal and Continuing Competency
The certification was valid for one year. Microsoft retired the renewal assessment with the certification on June 30, 2026, so an earned credential stays valid until it expires and cannot be renewed; to hold a current Azure AI credential, pass AI-103.
Total Cost of AI-102 Certification (2026)
- Item Cost Notes
- Exam fee (US) $165 Varied by country
Azure free account $0 ($200 credit + free monthly amounts) Azure OpenAI is not in the free-services list, so it draws on the credit; after the credit or 30 days, move to pay-as-you-go (billed per token) and set a budget
- Microsoft Learn training $0 Free official path + sandbox labs
- ExamFinch practice $0 Free scenario bank
- Paid practice tests, video courses, bootcamps (optional) Varies Check current prices with each provider
- Azure Free Tier Strategy for AI-102 Labs
Azure's free account and free tiers have limits that matter for AI labs:
Azure AI service free amounts - the Azure free-services list includes Speech, Vision, Document Intelligence and AI Search; check each service's pricing page for the monthly limits.
Azure OpenAI - not in the free-services list. Usage is billed per token, so experiment with a small model and set a budget.
Azure AI Search - one free service per subscription (50 MB of storage, up to 3 indexes, shared hardware with no replicas or partitions to scale) supports small RAG labs. Upgrade to Basic only if you need more than 50 MB; you can also delete and recreate the free service between labs.
Document Intelligence F0 - 500 pages free per month across all models, but only the first two pages of each document are analyzed; custom model training is free.
Content Safety F0 - 5,000 text records and 5,000 images free per month.
Keep Training with FREE AI-102 Practice
Practice AI-102 scenario questions across Azure OpenAI, RAG, AI Search, Vision, NLP, Document Intelligence, and responsible AI - all 100% FREE .
Frequently Missed 2026 Details (Competitor Guides Get These Wrong)
AI-102 retired June 30, 2026. Its replacement is AI-103: Developing AI Apps and Agents on Azure (Azure AI Apps and Agents Developer Associate).
QnA Maker is retired. Custom Question Answering replaced it, and Custom Question Answering itself retires from Azure Language on March 31, 2029.
LUIS was fully retired on March 31, 2026 ; its portal closed on October 31, 2025. Microsoft directs LUIS users to Conversational Language Understanding (CLU).
Cognitive Services became Azure AI services, now called Foundry Tools. The multi-service resource is now the Microsoft Foundry resource (kind AIServices), which reaches Azure OpenAI and Foundry Tools through one endpoint and key.
Azure AI Studio rebranded to Azure AI Foundry, then to Microsoft Foundry (late 2024 / 2025). The final AI-102 outline uses the "Microsoft Foundry" name; Microsoft now documents hub-based projects as the classic option.
text-embedding-3-large (3,072 dimensions) and text-embedding-3-small (1,536) are the newer embedding models; text-embedding-ada-002 (1,536) is the older one. Embedding models follow their own retirement timelines.
Integrated vectorization in Azure AI Search automates embedding via the Azure OpenAI embedding skill inside a skillset.
Anomaly Detector accepted no new resources after September 20, 2023, and Microsoft set its retirement for October 1, 2026. Personalizer followed the same path and retired on August 25, 2026.
Azure OpenAI Global Batch costs 50% less than Global Standard for async workloads with a 24-hour target turnaround.
Prompt Shield is the official Microsoft term for jailbreak + indirect prompt injection detection.
Groundedness detection (Content Safety) and the Groundedness evaluator (Foundry evaluations) are distinct features - know both.
Microsoft Foundry Agent Service (formerly Azure AI Agent Service) replaces the retired Assistants API, with tools such as web search, file search, code interpreter and functions.
Azure OpenAI On Your Data was retired on October 14, 2026 , and prompt flow retires on April 20, 2027.
Speaker Recognition and Speech intent recognition were retired on September 30, 2025; Spatial Analysis was retired on March 30, 2025.
Official Sources Used
Microsoft Learn - Exam AI-102: Designing and Implementing a Microsoft Azure AI Solution (skills outline, final revision December 23, 2025; exam retired June 30, 2026)
- Microsoft Learn - Exam AI-103: Developing AI Apps and Agents on Azure study guide (successor exam)
- Microsoft Learn - Microsoft Certified: Azure AI Apps and Agents Developer Associate credential page
- Microsoft Learn - AI-102 Study Guide (aka.ms/AI102-StudyGuide)
- Microsoft Certified: Azure AI Engineer Associate credential page
- Azure AI services documentation (learn.microsoft.com/azure/ai-services)
- Azure OpenAI Service documentation and model availability pages
- Azure AI Search documentation (vector + hybrid + semantic ranker)
- Azure AI Foundry / Microsoft Foundry documentation
- Microsoft Learn - What is Responsible AI (the six principles of the Microsoft Responsible AI Standard)
- Microsoft Learn - exam retake policy, exam scoring and exam duration pages; Microsoft exam pricing (US fee)
- Microsoft Learn credential renewal policy (6-month window, free assessment)
Microsoft Learn retirement notices for LUIS, Azure OpenAI On Your Data, prompt flow, the Assistants API, Custom Vision, Image Analysis, CLU and Custom Question Answering; Speech SDK release notes and Microsoft Q&A answers for Speaker Recognition and Spatial Analysis
Certification details, fees, and skills measured may be revised by Microsoft. Always confirm current requirements directly on learn.microsoft.com before scheduling.