Microsoft Developing AI Apps and Agents on Azure (AI-103)
Exam AI-102 and the Azure AI Engineer Associate certification retired June 30, 2026. The successor exam, AI-103 (Developing AI Apps and Agents on Azure), is more implementation-focused, built around Azure AI Foundry, with no automatic transition from the old credential. Five domains: Plan and manage an Azure AI solution (25-30%), Implement generative AI and agentic solutions (30-35%, the largest domain), Implement computer vision solutions (10-15%), Implement text analysis solutions (10-15%), Implement information extraction solutions (10-15%). 100 minutes, passing score 700.
Offered in: English, Chinese, French, German, Japanese, Korean, Italian, Portuguese, Spanish.
Owner: Microsoft · Read: · Source: Azure AI Apps and Agents Developer Associate (exam AI-103)
Exam fee at last check: not confirmed (the owner's page could not be read for the price on 2026-09-26; we hold $165 from an earlier check), per the vendor. The official objectives are at learn.microsoft.com/en-us/credentials/certifications/resources/study-guides/ai-103. Verified 2026-09-26.
Practice is free. The full analysis adds the right answer, the explanation and the source for every question, readiness by domain and a study plan, for one payment.
- Plan and Manage an Azure AI Solution25-30%
- Implement Generative AI and Agentic Solutions30-35%
- Implement Computer Vision Solutions10-15%
- Implement Text Analysis Solutions10-15%
- Implement Information Extraction Solutions10-15%
Source: Vendor exam objectives. Checked 2026-09-26. Bar length is relative to the heaviest domain; the figure beside each name is the published weight.
What's on the exam
Plan and Manage an Azure AI Solution 25-30%
Choosing the right Foundry services, setting up AI solutions, and managing/monitoring/securing them, plus responsible AI configuration.
- Choosing Foundry services and models (LLMs, small models, multimodal) for generative and agent tasks, and retrieval/indexing methods
- Designing Azure infrastructure for AI apps and agents, deployment options, and CI/CD integration with Foundry projects
- Managing quotas, scaling, and cost; monitoring model performance, drift, and grounding quality; securing with managed identity and private networking
- Responsible AI: safety filters, guardrails, evaluators, trace logging/provenance, and governing agent behavior with oversight modes
Primary sources: Microsoft Learn AI-103 study guide
Implement Generative AI and Agentic Solutions 30-35%
The largest domain: building generative applications and agents using Foundry, then optimizing and operationalizing them.
- Deploying/consuming models, implementing RAG, and building tool-augmented, multistep reasoning workflows via Foundry SDKs
- Defining agent roles, goals, and tool schemas; building agents with retrieval, function-calling, and conversation memory
- Multi-agent orchestration, autonomous/semiautonomous workflows with approval controls, and agent monitoring/error analysis
- Prompt engineering and parameter tuning, self-critique/reflection loops, and observability (tracing, token analytics, latency)
Primary sources: Microsoft Learn AI-103 study guide
Implement Computer Vision Solutions 10-15%
Image/video generation and editing, multimodal visual understanding, and responsible AI for visual content, all via Foundry.
- Generating and editing images/video from text prompts and reference media
- Multimodal visual understanding: captioning, visual question-answering, accessibility alt-text, and video analysis via Azure Content Understanding
- Responsible AI for visual content: unsafe-content filters, detecting indirect prompt injection embedded in images, watermarking/brand policy
Primary sources: Microsoft Learn AI-103 study guide, Azure AI Vision docs
Implement Text Analysis Solutions 10-15%
Language-model-driven text analysis and speech solutions using Foundry Tools.
- Extracting entities, topics, summaries, and structured JSON via generative prompting and Foundry Tools
- Sentiment/tone/safety detection, and translation via Azure Translator in Foundry Tools or LLM-powered flows
- Speech-to-text and text-to-speech for agentic interactions, custom speech models, and multimodal reasoning from audio
Primary sources: Microsoft Learn AI-103 study guide, Azure AI Language docs
Implement Information Extraction Solutions 10-15%
Retrieval/grounding pipelines and document extraction, connected directly to agent tools and RAG.
- Ingesting and indexing content (documents, images, audio, video); configuring semantic, hybrid, and vector search for grounding
- RAG ingestion flow including OCR, and connecting retrieval pipelines to agent tools
- Extracting structured or markdown output from documents via multimodal pipelines and Azure Content Understanding
Primary sources: Microsoft Learn AI-103 study guide, Azure AI Search vector search
Exam-day strategy
Implement Generative AI and Agentic Solutions is 30-35% of the exam, the largest domain. Do not under-study agent orchestration, tool integration, and evaluation just because the retired AI-102 barely tested agents.
Identify the Foundry capability the question is testing within the first read. Most questions resolve by matching the use case to the right named capability (function calling, multi-agent orchestration, Content Understanding, groundedness evaluation).
Watch for distractors that confuse Content Understanding (document/image/audio/video extraction) with a generic text-generation prompt with no structured schema.
For responsible-AI questions on agents, distinguish approval workflows/oversight modes (pre-action human control) from trace logging/provenance (post-action audit); both are tested but answer different questions.
PTUs vs pay-as-you-go and managed identity vs key-based auth are repeated patterns carried over from the retired exam. Memorize the canonical fit for each.
Pace at roughly 90 seconds per question across a 100-minute exam. Reserve time for case-study-style questions.
How to prepare for Microsoft Developing AI Apps and Agents on Azure (AI-103)
- EVIDENCE-BACKEDPractice in the same format as the exam. How well practice matches the final test shapes how much it helps. (Yang et al., 2021)
- EVIDENCE-BACKEDStart practice questions in week one, even on topics you have not studied. Failed attempts before seeing the answer still helped learning. (Kornell et al., 2009 and others)
- EVIDENCE-BACKEDSpace your reviews over weeks instead of cramming the week before. (Cepeda et al., 2006 and others)
- PLAUSIBLEDivide the minutes by the questions before you start, and keep a margin for a second pass if the exam allows one. (Sound arithmetic; no study compared time budgets on a licensing exam.)
Why this works
EVIDENCE-BACKED At least one peer-reviewed study or official source found the effect or states the fact. A study can be small.
PLAUSIBLE Reasonable and partly studied, but not tested on exams, or the results are mixed.
Practice that matches the final test format works better than practice that does not.
Failed attempts before seeing the answer still improved later learning of the material asked about; the benefit for material not asked about was virtually nil.
Spacing beat massing across hundreds of experiments and beat cramming with flashcards; combined with retrieval it raised scores on a real high-stakes exam.
Sound arithmetic; no study compared time budgets on a licensing exam.
Sources
- Yang, C., Luo, L., Vadillo, M. A., Yu, R., & Shanks, D. R. (2021). Testing (quizzing) boosts classroom learning: A systematic and meta-analytic review. Psychological Bulletin, 147(4), 399-435. doi.org/10.1037/bul0000309
- Kornell, N., Hays, M. J., & Bjork, R. A. (2009). Unsuccessful retrieval attempts enhance subsequent learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 35(4), 989-998. doi.org/10.1037/a0015729
- Richland, L. E., Kornell, N., & Kao, L. S. (2009). The pretesting effect: Do unsuccessful retrieval attempts enhance learning? Journal of Experimental Psychology: Applied, 15(3), 243-257. doi.org/10.1037/a0016496
- St. Hilaire, K. J., Chan, J. C. K., & Ahn, D. (2024). Guessing as a learning intervention: A meta-analytic review of the prequestion effect. Psychonomic Bulletin & Review, 31(2), 411-441. doi.org/10.3758/s13423-023-02353-8
- Cepeda, N. J., Pashler, H., Vul, E., Wixted, J. T., & Rohrer, D. (2006). Distributed practice in verbal recall tasks: A review and quantitative synthesis. Psychological Bulletin, 132(3), 354-380. doi.org/10.1037/0033-2909.132.3.354
- Kornell, N. (2009). Optimising learning using flashcards: Spacing is more effective than cramming. Applied Cognitive Psychology, 23(9), 1297-1317. doi.org/10.1002/acp.1537
- Kang, S. H. K. (2016). Spaced repetition promotes efficient and effective learning: Policy implications for instruction. Policy Insights from the Behavioral and Brain Sciences, 3(1), 12-19. doi.org/10.1177/2372732215624708
- Janes, J. L., Dunlosky, J., Rawson, K. A., & Jasnow, A. (2020). Successive relearning improves performance on a high-stakes exam in a difficult biopsychology course. Applied Cognitive Psychology, 34(5), 1118-1132. doi.org/10.1002/acp.3699
Additional resources
Explore Related Cybersecurity Resources
DecipherU is independent of Microsoft. Certified: Azure AI Engineer Associate (retired; successor: Developing AI Apps and Agents on Azure) is a trademark of Microsoft. DecipherU is not endorsed by, sponsored by, or affiliated with any certifying body.