AI-102 Retired: Microsoft AI-103 Study Guide (Azure AI Apps and Agents, 2026)

Microsoft AI-102 (Azure AI Engineer Associate) retired on June 30, 2026. You can no longer earn or renew it. The live associate credential is AI-103 — Azure AI Apps and Agents Developer Associate, built around Microsoft Foundry rather than a catalog of classic Cognitive Services.

If you were studying AI-102, do not finish that blueprint. AI-103 is not a rename. Generative AI and agents are now one combined domain worth about a third of the exam, planning is heavier, and vision/text/extraction are framed as inputs to apps and agents.

Start practicing: Free AI-103 practice questions → — 300+ questions across all five skills areas. Try 10 with no signup.

AI-103 Exam Overview

The exam is scenario-heavy. Microsoft tests whether you can choose Foundry services, ground a model with retrieval, wire tools into an agent, and operate the solution with security and responsible-AI controls — not whether you can recite old service names.

Official outline: Study guide for Exam AI-103.

Who Should Take AI-103?

AI-103 is the right next cert if you:

• Build AI apps or agents on Azure (RAG chat, tool-calling agents, multimodal workflows)
• Already passed AI-900 (or equivalent) and want an associate-level credential
• Were preparing for AI-102 and need the successor path
• Work with Microsoft Foundry — model catalog, deployments, evaluation, and agent runtimes
• Want a Microsoft-backed signal for Azure AI engineer / applied AI developer roles

Skip it (for now) if you only need AI literacy for a business role — take AI-900 first. If your stack is AWS, start with AWS AI Practitioner instead of forcing Azure.

AI-102 vs AI-103: What Actually Changed

If you already hold AI-102: the credential stopped being renewable after retirement. Treat AI-103 as the current proof of Azure AI app/agent skills. Overlap exists (RAG, safety, search, documents), but study Foundry agents, tool schemas, multi-agent orchestration, and evaluation — those are first-class on AI-103.

If you never sat AI-102: ignore archived AI-102 banks except as background. Practice against the AI-103 hub.

AI-103 Exam Domains (April 16, 2026)

Weight your study time to the two heavy domains. Together they are 55–65% of the exam.

1. Implement generative AI and agentic solutions (30–35%)

This is the exam. Budget the most hours here.

Generative apps

• Deploy and consume LLMs, small language models, code models, and multimodal models in Foundry
• Implement RAG: chunk → embed → retrieve → ground generation; measure fabrications and relevance
• Design tool-augmented and multi-step reasoning flows
• Evaluate quality, safety, and grounding; wire Foundry SDKs/connectors into an application

Agents

• Define roles, goals, conversation tracking, and tool schemas
• Integrate retrieval, function calling, and conversation memory
• Connect tools: APIs, knowledge stores, search, content understanding, custom functions
• Orchestrate multi-agent solutions with explicit handoffs
• Add safeguards and approval gates for risky tools
• Trace, evaluate, and debug agent behavior in production

Ops

• Prompt engineering and generation parameters
• Observability: tracing, token analytics, safety signals, latency
• Hybrid setups (multiple models or rules engines)

Study tip: On scenario items, prefer “execute the tool in application code and return the real result” over “let the model invent the tool output.” Keep system instructions separate from untrusted user or retrieved content.

Practice: Generative AI & agents questions

2. Plan and manage an Azure AI solution (25–30%)

Heavier than the old AI-102 planning slice, and Foundry-specific.

• Choose models (LLM vs SLM vs multimodal vs Foundry Tools) for task, cost, and latency
• Choose services for generation, grounding, vector search, agent workflows, or multimodal work
• Pick retrieval/indexing methods and memory/tool/knowledge integrations
• Design Azure infrastructure; pick deployment options; configure model and agent deployments
• Integrate Foundry projects with CI/CD (version prompts and agent configs; evaluation gates before prod)
• Manage quotas, scaling, rate limits, and cost
• Monitor model performance, drift, safety events, grounding quality, and index health
• Security: managed identity, private networking, keyless credentials, least-privilege RBAC
• Responsible AI: safety filters, guardrails, evaluators, trace logging, provenance, oversight modes

Study tip: Separate invoke roles from deployment-admin roles. Know when provisioned throughput beats pay-as-you-go for a latency-sensitive agent.

Practice: Plan & manage questions

3. Implement computer vision solutions (10–15%)

Not a Custom Vision trivia dump. Expect generation and multimodal understanding.

• Text-to-image and text-to-video generation; reference media; inpainting / mask-based edits
• Multimodal models for visual Q&A, captions, alt text, and accessibility
• Content Understanding pipelines for visual characteristics and video segments
• Responsible AI for media: unsafe-content filters, indirect prompt injection via text in images, watermarks and policy rules

Practice: Computer vision questions

4. Implement text analysis solutions (10–15%)

Language skills in service of apps and agents — including speech as a modality.

• Entities, topics, summaries, and structured JSON for downstream tools
• Sentiment, tone, safety, and sensitive-content detection (PII before logging)
• Translation via Azure Translator or LLM translation flows
• Speech-to-text / text-to-speech in the agent loop; custom speech; audio reasoning

Practice: Text analysis questions

5. Implement information extraction solutions (10–15%)

This domain punches above its weight because RAG and agents depend on it.

• Ingest documents, images, audio, and video; index for grounding
• Semantic, hybrid, and vector search
• Enrichment (built-in and custom skills); OCR in the RAG ingestion path
• Document / Content Understanding analyzers for structured or markdown output agents can reason over

Study tip: Know when you need deterministic field extraction versus generative, schema-driven understanding for RAG.

Practice: Information extraction questions

6-Week AI-103 Study Plan

Total time: 60–90 hours (about 10–15 hours/week). Microsoft’s own guidance is hands-on before you sit the exam.

Weeks 1–2: Platform and RAG

• Work through Microsoft Learn modules linked from the AI-103 study guide
• Stand up a Foundry project, deploy a chat model, and call it from Python
• Build one RAG path: ingest a document set, retrieve, generate, then evaluate a handful of questions for grounding failures
• Review AI-103 study notes for planning and gen-AI tabs

Weeks 3–4: Agents and multimodal

• Define an agent with at least one real tool (search or a custom function) executed in app code
• Add a risky-tool approval path (even a simple confirm step)
• Practice vision and speech items so they are not surprises on exam day
• Drill flashcards for Foundry, RAG, and agent terms

Weeks 5–6: Extraction and exam simulation

• Index mixed content; compare vector vs hybrid retrieval on the same corpus
• Take three timed practice exams under exam-like conditions
• Log every miss by domain. Re-read explanations for guesses you got right

Hands-On Skills That Show Up on Exam Day

Know these patterns cold:

• RAG quality: chunk size, overlap, metadata filters, hybrid search, and what to do when the model fabricates
• Agent tools: JSON schemas, timeouts, retries, and never trusting the model as the source of tool results
• Identity: managed identity to Foundry and search; no keys in code
• Safety: content filters on generate and on retrieved/tool content; image prompt-injection
• Ops: traces for a failed agent turn; token/latency cost of a verbose prompt
• CI/CD: prompts and agent configs versioned; evaluation gate before production

Pair labs with official Azure AI documentation and Foundry docs — the exam can include commonly used preview features, but most items target GA.

How Many Practice Questions Do You Need?

Most candidates who pass on the first try complete 300–500 practice questions, including at least two or three full mock exams, with a stable score at or above the 700 scaled target (treat ~75%+ on high-quality mocks as a booking signal).

CertStud offers:

• 300+ AI-103 practice questions with explanations
• 3 timed practice exams
• Flashcards and study notes
• Free AI-103 starter set (10 questions, no signup)

Start AI-103 practice →

Archived AI-102 practice remains available for related Azure AI skills only. Do not use it as your primary bank.

Exam Cost and What the Cert Signals

Azure AI app/agent roles in the US commonly land in the $120,000–$160,000 range depending on location and seniority. AI-103 will not replace a portfolio of shipped RAG/agent work — it helps you get past resume screens for Microsoft-stack AI engineering jobs.

Ask about employer reimbursement before you pay out of pocket.

Common Mistakes (and How to Avoid Them)

1. Finishing an AI-102 study plan — the weights and platform story changed; switch to AI-103 materials
2. Memorizing old Cognitive Services names without building a Foundry app or agent
3. Skipping information extraction because it is “only 10–15%” — grounding questions leak into the heavy domains
4. Treating agents as chatbots with extra prompts — tools, memory, and approval gates are in the skills outline
5. Ignoring responsible AI — filters, evaluators, traces, and image prompt-injection are tested
6. Never timing a full exam — case-style items consume minutes fast

Related Certifications

Also see our best AI certifications in 2026 roundup (AI-103 is the current Azure associate path).

Final Checklist Before You Book

• Completed 300+ AI-103 practice questions across all five domains
• Scored consistently in the pass range on at least 2 timed mocks
• Built at least one RAG app and one tool-using agent in Foundry
• Reviewed security (managed identity, private networking, RBAC) and responsible-AI controls
• Read the current official AI-103 skills outline

AI-103 rewards people who have shipped grounded apps and agents, not people who only watched videos. Follow the weights, practice under time, and book when mock scores are stable.

Ready? Start free AI-103 practice questions →