AI-103: Developing AI Apps and Agents on Azure is the exam for Microsoft Certified: Azure AI Apps and Agents Developer Associate. It replaced AI-102 (Azure AI Engineer Associate), which retired on 30 June 2026. If you were preparing for AI-102, this is the exam to take now. It keeps much of the old content (vision, language, search, document extraction) and adds a large block on generative AI and agents built on Microsoft Foundry, which you build in Python.
This guide follows Microsoft's study guide (skills measured as of 16 April 2026). It covers every skill area, the Foundry code you need to be comfortable with, an 8-week plan, labs, and 20 original practice scenarios.
Quick facts (Microsoft Learn, November 2026)
| Item | Detail |
|---|---|
| Exam | AI-103: Developing AI Apps and Agents on Azure |
| Certification | Microsoft Certified: Azure AI Apps and Agents Developer Associate |
| Time | 120 minutes; proctored; may include interactive components |
| Passing score | 700 |
| Language for code | Python |
| Renewal | Every 12 months, by a free online assessment |
| Retake | 24 hours after the first attempt; longer waits after that |
| Price | Set by the country where you take the exam |
Check the official AI-103 study guide before booking. For the full list of 2026 exam changes, see Microsoft certifications retired in 2026.
Who should take AI-103?
- Python developers moving into GenAI and agent development on Azure.
- Anyone who was preparing for AI-102.
- AI engineers and Forward Deployed Engineers who deliver Azure AI projects.
- Freshers who have passed AI-901 and have built a few projects.
Microsoft expects Python app-development experience and familiarity with general AI, generative AI and Azure services. If Azure itself is new to you, cover AZ-900 basics first.
Skills measured and weights
| Skill area | Weight |
|---|---|
| Plan and manage an Azure AI solution | 25–30% |
| Implement generative AI and agentic solutions | 30–35% |
| Implement computer vision solutions | 10–15% |
| Implement text analysis solutions | 10–15% |
| Implement information extraction solutions | 10–15% |
More than half the exam (55–65%) is planning, managing and building generative AI and agent solutions.
1. Plan and manage an Azure AI solution (25–30%)
Choose the right Foundry services:
- pick the model for each task: LLMs, small language models, multimodal models or Foundry Tools;
- choose services for generation, grounding, vector search, agent workflows and multimodal processing;
- choose a retrieval and indexing method;
- choose memory, tool and knowledge integrations for agents.
Set up AI solutions in Foundry:
- design Azure infrastructure for AI apps and agents;
- choose deployment options and configure model and agent deployments;
- connect Foundry projects to CI/CD pipelines.
Manage, monitor and secure:
- quotas, scaling, rate limits and cost;
- monitoring for model drift, safety events and grounding quality;
- data-ingestion quality and search index health;
- security: managed identity, private networking, keyless credentials and role policies.
Responsible AI:
- safety filters, guardrails and risk detection;
- evaluators and safety evaluations;
- audit trails with trace logs, provenance and approval workflows;
- governing agents through oversight modes, constraints and tool-access controls.
2. Generative AI and agentic solutions (30–35%)
Build generative apps with Foundry:
- deploy and call LLMs, small models, code models and multimodal models;
- implement RAG;
- design multi-step and tool-augmented flows;
- evaluate apps for fabrication, relevance, quality and safety;
- connect an application to a Foundry project through the SDK.
Build agents with Foundry:
- define agent roles, goals, conversation tracking and tool schemas;
- combine retrieval, function calling and memory;
- integrate tools: APIs, knowledge stores, search, Content Understanding and custom functions;
- orchestrate multi-agent solutions;
- run autonomous or semi-autonomous workflows with safeguards and approvals;
- monitor agents and analyse their errors.
Optimise and operationalise:
- prompt engineering and model parameters;
- reflection and self-critique loops;
- tracing, token analytics, safety signals and latency breakdowns;
- orchestrating several models, or mixing LLMs with rules engines.
Background reading: the Microsoft Foundry tutorial, Foundry RAG with Azure AI Search and the Foundry IQ tutorial.
The Foundry code pattern to know
Use keyless authentication (Microsoft Entra ID) with the Foundry project SDK:
# pip install azure-ai-projects azure-identity
from azure.identity import DefaultAzureCredential
from azure.ai.projects import AIProjectClient
project = AIProjectClient(
endpoint="https://<resource>.services.ai.azure.com/api/projects/<project>",
credential=DefaultAzureCredential(), # keyless: managed identity / az login
)
client = project.get_openai_client() # OpenAI-compatible client for this project
resp = client.responses.create(
model="<your-model-deployment-name>",
input="Summarise our travel policy for new joiners in 3 bullets.",
)
print(resp.output_text)From this base, the exam topics extend to RAG (retrieving from Azure AI Search first), agents (an agent with tools such as search, functions or MCP), and evaluation and tracing in the Foundry portal.
3. Computer vision (10–15%)
- Image and video generation: generate from text and reference media; edit images with masks (inpainting) and prompts; edit generated video; use the platform's generation controls.
- Multimodal understanding: captions; visual Q&A; accessible alt text; Azure Content Understanding for visual features and video analysis; single-task vs pro-mode pipelines; finding objects and regions.
- Responsible AI for images: filter unsafe content; detect indirect prompt injection hidden in text inside images; apply watermarks and brand rules.
4. Text analysis (10–15%)
- Language model analysis: extract entities, topics, summaries and structured JSON; detect sentiment, tone, safety issues and sensitive content; translate with Azure Translator or LLM flows; customise output for domain tasks such as compliance summaries.
- Speech: speech-to-text and text-to-speech for agents; speech as an agent modality, including custom speech models; reasoning over audio; speech translation.
5. Information extraction (10–15%)
- Retrieval and grounding pipelines:
- ingest and index documents, images, audio and video;
- configure semantic, hybrid and vector search;
- add enrichment skills;
- set up RAG ingestion with OCR;
- connect retrieval directly to agent tools.
- Document extraction: multimodal pipelines (OCR, layout and field extraction); clean, grounded output from Content Understanding; analysers that produce structured or Markdown output.
8-week study plan (about 2 hours a day)
| Week | Focus | Lab |
|---|---|---|
| 1 | Foundry resources, projects, model deployments, keyless authentication, quotas | Deploy a model and call it from Python (above) |
| 2 | Prompting, parameters, structured output, evaluation basics | Build an evaluation set and compare two prompts |
| 3 | RAG with Azure AI Search: hybrid and semantic search, vectors | Policy-document Q&A with citations |
| 4 | Agents: tools, function calling, memory, MCP | Agent that calls a custom function and a search tool |
| 5 | Multi-agent orchestration, approvals, tracing, safety evaluations | Two-agent workflow with a human approval step |
| 6 | Vision and Content Understanding (images and video) | Invoice and photo extraction pipeline |
| 7 | Language, Translator, Speech | Voice-enabled agent with speech-to-text and text-to-speech |
| 8 | Security (private networking, RBAC), monitoring, CI/CD; practice assessment twice | Revise weak areas and book the exam |
Microsoft offers a free Practice Assessment for AI-103 on AI Skills Navigator. Aim for 80%+ before you book.
20 practice scenarios with answers
These are original scenarios written for this guide, not exam questions.
1. An app must call a Foundry model without storing any keys. What do you use? Microsoft Entra ID authentication (keyless) with a managed identity, plus an RBAC role on the Foundry resource.
2. Answers must be grounded in 5,000 internal PDFs and include citations. What pattern applies? RAG: index the PDFs in Azure AI Search (hybrid with semantic ranking) and pass the retrieved passages to the model with source references.
3. Users report answers that are confident but wrong. Which evaluation helps? Groundedness or fabrication evaluation against the retrieved context, plus relevance checks.
4. Model calls are being throttled at peak times. What do you check first? Quotas and rate limits on the deployment, then scaling or provisioned capacity, and add retry with backoff.
5. The agent must not issue refunds above ₹10,000 without a person. How? Put an approval step (human-in-the-loop) in the workflow and restrict which tools the agent can use.
6. An agent should look up order status in an internal API. What do you configure? A tool: a custom function or an OpenAPI or MCP tool with a defined schema, with the agent authorised to call it.
7. Two agents (research and writer) must cooperate. What is this? Multi-agent orchestration. Define the roles, how work is handed over, and the shared state.
8. A product team needs the latency split between retrieval and generation. What do you implement? Tracing with latency breakdowns and token analytics.
9. An image uploaded by a user contains the text "ignore previous instructions". What risk is this? Indirect prompt injection through text embedded in an image. Filter it and don't let content in images act as instructions.
10. Generate product images and then replace just the background. Which capability? Image generation followed by mask-based editing (inpainting).
11. Write alt text for every catalogue image to meet accessibility rules. How? A multimodal model, or Content Understanding, configured to produce concise or extended descriptions.
12. Extract fields from scanned invoices for an agent to use. Which service? Azure Content Understanding in Foundry Tools, with an analyser that outputs structured fields or Markdown.
13. Search must match exact codes like "HL-2041" as well as meaning. Which search type? Hybrid search: keyword plus vector, ideally with semantic ranking.
14. A call-centre agent must accept voice and reply by voice. Which services? Azure Speech in Foundry Tools for speech-to-text and text-to-speech, integrated as an agent modality.
15. Telugu customer emails must be answered in English and translated back. What do you use? Azure Translator in Foundry Tools, or an LLM translation flow.
16. Outputs must never include customer phone numbers. What do you configure? Sensitive-content detection and filtering, plus guardrails and redaction before the response is returned.
17. The Foundry project must not be reachable from the internet. How? Private networking with private endpoints and public network access disabled.
18. Model and agent changes must go through review before production. How? Integrate the Foundry project with a CI/CD pipeline that runs evaluations as a quality gate.
19. Answer quality dropped after documents were re-indexed. What do you monitor? Data-ingestion quality, search index health and relevance metrics.
20. A simple FAQ classifier is costing too much on a large model. What do you change? Use a small language model, or a rules engine combined with an LLM fallback, and route requests by complexity.
After AI-103
- Prepare for interviews with 130 Azure AI services interview questions and 75 agentic AI interview questions.
- Build portfolio pieces from 25 agentic AI projects.
- For back-end roles, add AI-200. For infrastructure, add AZ-104.
- See live AI jobs and the Azure for AI engineers guide.
Cloud Soft Solutions in Ameerpet, Hyderabad teaches Microsoft Foundry, Azure AI and agentic AI with hands-on projects. See the FDE PRO program and all courses.
📞 Call +91 96660 19191 · WhatsApp · Flat No. 513, 5th Floor, Nilagiri Block, Aditya Enclave, Ameerpet, Hyderabad
Frequently Asked Questions
What replaced AI-102?
AI-103: Developing AI Apps and Agents on Azure, for the Azure AI Apps and Agents Developer Associate certification. AI-102 retired on 30 June 2026.
How long is the AI-103 exam?
Microsoft lists 120 minutes; the exam is proctored and may include interactive components.
What is the AI-103 passing score?
700.
Which language does AI-103 use?
Python, with Microsoft Foundry SDKs.
Which AI-103 area has the most weight?
Implement generative AI and agentic solutions, at 30–35%.
How often must the certification be renewed?
Every 12 months, through a free online renewal assessment on Microsoft Learn.
