FREE Microsoft AI-103 Practice Questions by Domain (2026)

AI-103 Practice in 2026: Start With the Official Blueprint

Exam AI-103: Developing AI Apps and Agents on Azure is the exam for the Microsoft Certified: Azure AI Apps and Agents Developer Associate credential. It is now generally available, after an earlier beta period. Microsoft retired AI-102 on June 30, 2026, so new candidates pursuing the replacement Azure AI associate path should prepare for AI-103 rather than the retired outline.

The current AI-103 study guide lists five skill domains as of April 16, 2026. This guide turns those published objectives into a cautious practice plan without inventing a fixed question count or treating common Microsoft exam patterns as confirmed AI-103 facts.

Start free AI-103 practice questionsPractice questions with detailed explanations

---

Current AI-103 Status and Published Exam Facts

As of July 29, 2026, Microsoft's English credential page no longer carries the beta label, offers Pearson VUE scheduling, and identifies the credential as intermediate level. Microsoft's June 2026 credentials roundup also lists AI-103 as generally available.

Keep the format boundaries straight

Microsoft does not publish a fixed AI-103 question count. A candidate report, training provider, or generic claim that Microsoft exams usually contain a certain range does not establish an official AI-103 count. Likewise, the credential page says interactive components may appear; it does not promise a lab, case study, or any specific item mix.

The 700 passing score is scaled, not a statement that 70% of questions must be correct. Microsoft also notes that some questions may be unscored and that there is no penalty for guessing. Answer every item, but do not try to convert a practice-test percentage directly into Microsoft's scaled score.

AI-103 began as a beta exam. Microsoft's beta-exam policy explains why beta scores are delayed while exam data is analyzed. That policy should not be presented as the current AI-103 experience: the exam is now generally available. The official AI-103 Practice Assessment is still unavailable, so use the exam sandbox for interface familiarity and clearly labeled third-party questions for practice.

---

Official AI-103 Skills Measured: April 16, 2026

The two largest ranges total 55-65%, so they deserve the largest share of study time. That arithmetic does not tell you how many questions will appear in either domain. Microsoft explains that the bullets illustrate how a skill may be assessed, related topics may appear, and most questions cover generally available features while commonly used preview features may also appear.

A sound practice bank should preserve Microsoft's ranges instead of replacing them with invented point estimates such as 27% or 33%. Use the ranges to allocate time, not to predict an exact exam form.

---

Domain 1: Plan and Manage an Azure AI Solution (25-30%)

What Microsoft lists

The official outline has four objective groups:

1. Choose appropriate Foundry services. Select models, services for generative work, grounding, vector search, agents, and multimodal processing, plus retrieval, indexing, memory, tool, and knowledge-integration methods.
2. Set up AI solutions in Foundry. Design Azure infrastructure, choose deployment options, configure model and agent deployments, and integrate Foundry projects with CI/CD pipelines.
3. Manage, monitor, and secure AI systems. Manage quotas, scaling, rate limits, and cost; monitor model performance, drift, safety, grounding, ingestion, index health, and relevance; configure identity, networking, credentials, and roles.
4. Implement responsible AI. Configure filters and guardrails, apply evaluators and safety instrumentation, implement trace and provenance auditing, and govern agents with oversight, constraints, approvals, and tool-access controls.

Practice method

Create comparison questions that make one requirement decisive. Examples include choosing an identity approach for least privilege, selecting a deployment approach for a workload constraint, identifying the monitoring signal that reveals poor grounding, or placing an approval gate before a high-impact tool call. These are practice designs derived from the objective list, not a claim about the wording of live AI-103 items.

Build one Foundry project and record evidence for each objective group: infrastructure diagram, deployment configuration, identity and network controls, quota view, evaluation result, and trace containing an approval decision. If you cannot explain why each control is present, return to the matching Microsoft Learn documentation.

---

Domain 2: Implement Generative AI and Agentic Solutions (30-35%)

This is the largest published domain. It combines three substantial objective groups.

Build generative applications by using Foundry

Microsoft lists deploying and consuming LLMs, small models, code models, and multimodal models; implementing RAG; designing tool-augmented and multistep workflows; evaluating fabrication, relevance, quality, and safety; using Foundry SDKs and connectors; and connecting an application to a Foundry project.

A useful lab is a small RAG application with a documented ingestion path, retrieval configuration, model deployment, evaluation run, and failure case. Practice questions should ask which layer is responsible for a stated problem: source ingestion, retrieval, grounding context, model behavior, evaluation, or application integration.

Build agents by using Foundry

The outline covers agent roles and goals, conversation tracking, tool schemas, retrieval, function calling, memory, APIs, knowledge stores, search, Content Understanding, custom functions, multi-agent orchestration, safeguards, approval controls, monitoring, evaluation, and error analysis.

Build an agent with at least two tools and one explicit approval boundary. Then test malformed tool arguments, an unavailable tool, irrelevant retrieval, and a request that should not be executed automatically. Those exercises make the listed control and error-analysis objectives concrete without pretending to reproduce Microsoft exam questions.

Optimize and operationalize generative AI systems

Microsoft includes prompt engineering, model parameters, reflection, chain-of-thought evaluations, self-critique loops, tracing, token analytics, safety signals, latency breakdowns, and orchestration across models, flows, or hybrid LLM-and-rules systems.

For practice, compare two traces and choose the evidence-supported correction. Keep tuning, evaluation, safety, cost, and latency separate: a change that improves one measure may worsen another.

---

Domain 3: Implement Computer Vision Solutions (10-15%)

Microsoft divides this domain into three groups:

• Image and video generation: generate from prompts and reference media, configure image edits such as inpainting and masks, edit generated video, and select platform controls.
• Multimodal understanding: analyze visual context, produce captions and accessible descriptions, answer questions from visual evidence, use Content Understanding, analyze video segments, configure single-task and pro-mode pipelines, and identify image or video regions.
• Responsible multimodal AI: classify unsafe visual content, mitigate indirect prompt injection through embedded image text, and enforce policies such as watermarks, prohibited-symbol flags, and brand requirements.

A compact lab can cover all three: analyze reference images, generate an accessible description, process a video segment, and test an image containing instruction-like text. Practice should distinguish generation, editing, understanding, and policy enforcement instead of treating every visual task as generic computer vision.

---

Domain 4: Implement Text Analysis Solutions (10-15%)

The official outline covers two groups.

Language-model text analysis includes extracting entities, topics, summaries, and structured JSON; detecting sentiment, tone, safety issues, and sensitive content; translating with Azure Translator in Foundry Tools or an LLM-powered flow; and customizing outputs for domain tasks.

Speech solutions include speech-to-text, text-to-speech, custom speech models, speech as an agent modality, multimodal reasoning from audio, and speech translation through language models and Foundry Tools.

Practice service selection by writing down the actual requirement before comparing options: structured output, domain-specific extraction, deterministic translation, voice interaction, custom speech, or audio reasoning. A small voice-agent loop provides useful experience, but Microsoft does not say every candidate will receive a speech implementation task.

---

Domain 5: Implement Information Extraction Solutions (10-15%)

Build retrieval and grounding pipelines

Microsoft lists ingestion and indexing for documents, images, audio, and video; semantic, hybrid, and vector search; built-in and custom enrichment skills; OCR in a RAG ingestion flow; and direct connections from retrieval pipelines to workflows and agent tools.

Extract content from documents

The outline includes multimodal OCR, layout analysis, field extraction, grounded representations for agents and RAG through Content Understanding, and analyzers that produce structured or markdown output.

Build one end-to-end pipeline from a scanned or layout-rich source to searchable, structured output. Then practice identifying where a defect originates: OCR, layout, field mapping, enrichment, index configuration, retrieval, or the downstream agent. This domain overlaps with RAG, but its emphasis is the extraction and grounding pipeline.

---

A Domain-Weighted Practice Workflow

Pass 1: Blueprint audit

Copy every official objective into a tracker. Mark each one as unfamiliar, conceptual only, lab-verified, or practice-verified. Do not mark an entire domain complete because you recognize its heading.

Pass 2: Weighted questions

Allocate approximately 25-30% of practice to planning, 30-35% to generative AI and agents, and 10-15% to each smaller domain. For a concrete 15-question diagnostic that stays inside all five ranges, use 4 planning, 5 generative-and-agentic, and 2 questions for each smaller domain. This is a study allocation, not a forecast of the live exam.

For every miss or uncertain answer, record the official objective, the decisive requirement in the question, why the closest distractor fails, the Microsoft Learn page that confirms the rule, and one lab step that would verify it. Repeat the weak objective with a new scenario before returning to a mixed set.

Pass 3: Evidence-producing labs

Use one or more labs to produce artifacts tied to the outline: deployment settings, identity configuration, an index, a RAG response with sources, an agent trace, an evaluation result, multimodal output, speech flow, and extracted structured data.

Pass 4: Error classification

Tag every missed practice item:

• Objective gap: you did not know the listed skill.
• Service boundary: you confused two valid services or patterns.
• Configuration gap: you knew the service but not the required setting.
• Evidence gap: you chose an answer without connecting it to the stated requirement.
• Reading gap: you overlooked a constraint.

Repair the specific gap with documentation or a lab before repeating the item.

Pass 5: Timed mixed review

Use timed sets to practice reading and decision speed, but do not label a 40-, 50-, or 60-question mock as a replica of AI-103. Microsoft publishes 120 minutes, not a fixed question count. Use the Microsoft exam sandbox to learn the interface rather than inferring the live item mix from a third-party bank.

Practice across all five AI-103 domainsPractice questions with detailed explanations

---

Moving From AI-102 to AI-103

Microsoft's AI-102 study guide now states that AI-102 retired on June 30, 2026, at 11:59 p.m. Central Standard Time. A current Microsoft Partner Center announcement says the retired Azure AI Engineer Associate credential was replaced by Azure AI Apps and Agents Developer Associate (AI-103) on June 19, 2026.

If you studied AI-102 but did not pass before retirement, retain transferable work in Azure AI Search, responsible AI, extraction, vision, speech, and generative AI, then audit it against every AI-103 objective. Add the current Foundry planning, agent, orchestration, evaluation, multimodal, and operationalization objectives rather than assuming the exam was merely renamed.

If you already earned the retired Azure AI Engineer Associate credential, Microsoft says an earned credential remains valid on your transcript until it expires. Retirement did not automatically convert it into the new credential.

---

Official Sources to Recheck Before Scheduling

• AI-103 study guide and April 16, 2026 skills measured
• Azure AI Apps and Agents Developer Associate credential page
• Microsoft Credentials roundup: June 2026
• Microsoft beta-exam scoring and results policy
• AI-102 retired study guide
• Microsoft Partner Center replacement announcement
• Microsoft credential expiration policy
• Microsoft exam scoring and score reports

Bottom Line

AI-103 is now the current, generally available Microsoft associate exam for Azure AI apps and agents. Build your preparation around the April 16, 2026 five-domain outline: 25-30% planning, 30-35% generative AI and agents, and 10-15% each for computer vision, text analysis, and information extraction. Practice by objective, verify skills with labs, and keep unofficial question counts and assumed item patterns out of your exam-day plan.