Practice Exams:

AI-102 Retired: The Current AI-103 Azure AI Developer Roadmap

Azure AI engineering has moved beyond selecting a pretrained service and connecting it to an application. Current work includes grounding generative responses in organizational information, giving agents carefully constrained access to tools, evaluating outputs, and operating the entire system securely. Those skills are relevant to readers of the historical AI-102 Azure AI Engineer Associate curriculum, but the exam itself is no longer available.

Microsoft
On this page
  1. What changed between AI-102 and AI-103?
  2. Start with Python, Foundry and sound solution design
  3. Build generative applications and agents around controlled tools
  4. Use Azure AI Search and document extraction to ground answers
  5. Apply vision, language and speech capabilities to real requirements
  6. Secure, evaluate and operate the solution
  7. A focused learning project for AI-103
  8. How to use AI-102 history without misleading current candidates

Microsoft retired AI-102 on June 30, 2026. It cannot be newly booked or passed after that date. For current certification preparation, Microsoft offers AI-103: Developing AI Apps and Agents on Azure, associated with Microsoft Certified: Azure AI Apps and Agents Developer Associate. This article explains which engineering ideas remain useful, which legacy services need migration, and how to study the active pathway without mistaking an old exam guide for current scheduling advice.

What changed between AI-102 and AI-103?

The old AI-102 exam covered the design and implementation of Azure AI solutions, including vision, language, conversational interfaces, search, and other managed services. Its historical objectives help explain how developers assembled AI applications before agentic and generative patterns became central. They are a reference for technology history, not a valid new-candidate exam blueprint.

AI-103 is a distinct examination path rather than a new label for the same questions. Microsoft expects experience developing Python applications with Microsoft Foundry and knowledge of both generative AI and Azure services. The issuer’s April 16, 2026 skills-measured outline gives the following assessment areas:

  • Plan and manage an Azure AI solution (25–30%): choose models and services, configure projects and deployments, govern costs and access, and apply responsible AI controls.
  • Implement generative AI and agentic solutions (30–35%): build generative applications, tool-enabled agents, grounding workflows, and evaluation and monitoring practices.
  • Implement computer vision solutions (10–15%): handle visual and multimodal input, image or video generation and editing, and safety requirements.
  • Implement text analysis solutions (10–15%): perform structured extraction, classification, translation, and speech-enabled interactions.
  • Implement information extraction solutions (10–15%): ingest and index content, retrieve reliable evidence, and extract document structure for downstream processing.

These percentages refer to Microsoft’s published AI-103 outline, not AI-102. The issuer may revise the outline or exam policies; check the live AI-103 exam page before booking, and use its study guide as the authoritative map of skills.

Start with Python, Foundry and sound solution design

A practical foundation includes Python, JSON, REST APIs, authentication, Azure identities and role assignments, cloud networking, and the ability to reason about data boundaries. Learn how a Foundry project connects an application to its models and tools, where the credentials and permissions are established, and how quota, latency, regional availability, and cost affect deployment choices. A successful demonstration is not merely a notebook that produces plausible text: it must handle failed requests, access errors, response timeouts, and data-handling constraints.

Begin with a small use case and write down the requirements. For example, a support assistant may need to answer questions based on approved product documentation, refuse unsupported account operations, and refer ambiguous cases to an employee. Decide which source system is authoritative, what evidence the assistant must retrieve, whether private data can be used, and which actions require human approval. That specification becomes the basis for choosing services and for subsequent testing.

Use separate development and production configurations. Keep credentials in approved secret or identity mechanisms rather than source code. Capture the model and prompt configuration, API versions, indexes, and application dependencies so changes can be reproduced. Microsoft Foundry and the surrounding Azure tools can be revised over time; code copied from older AI-102 labs should be checked against the currently supported SDK and service versions.

Build generative applications and agents around controlled tools

Generative systems must be evaluated on more than grammatical fluency. A useful application design defines which requests go directly to a model, which require retrieval from a trusted knowledge source, and which are outside the application’s role. Retrieval-augmented generation (RAG) can ground answers in documents, but it only works when content ingestion, permissions, freshness, and relevance are managed. If the index contains stale policies, a well-written answer can still be wrong.

An agent adds another dimension: a model may choose tools, call APIs, retain conversation context, or coordinate several steps. Define explicit tool schemas and input validation, use least-privilege access, and restrict actions that change records or send messages. For high-impact actions, display the intended effect and require an approval step. Treat retrieved documents and model outputs as untrusted instructions so a malicious document cannot silently expand an agent’s authority.

Test the workflow with both ordinary and adversarial examples. Measure whether it chooses the right tool, whether calls are made with valid arguments, whether it can recover from unavailable systems, and whether it distinguishes sourced answers from guesses. A trace should let developers inspect retrieval evidence, tool calls, failures, latency, and safety decisions without exposing unnecessary personal information. The aim is a controlled, auditable business workflow, not an agent that takes the maximum number of autonomous steps.

Use Azure AI Search and document extraction to ground answers

Many enterprise AI applications depend on knowledge that lives in PDFs, web pages, support records, or structured repositories. Design ingestion around ownership and access rules before choosing an index. Identify the document formats, expected update interval, versioning conventions, and what should happen when information is missing or contradictory. For each answer, decide what constitutes acceptable supporting evidence.

Azure AI Search provides indexing and retrieval options, including text search, vector search, and hybrid approaches. Query design must consider recall, relevance, language, filters, and the authorization model. A demo that retrieves an apparently relevant snippet is not enough; evaluate whether the source is permitted for the requesting user, whether citations actually support the answer, and whether indexing updates reflect current information. Search quality may degrade when content structure, vocabulary, or user questions change.

For documents with tables, forms, images, or complex layouts, the extraction stage requires extra care. Microsoft lists Content Understanding and related document-processing techniques under the AI-103 information-extraction objectives. Validate the resulting text, layout or structured fields against representative examples before using them for RAG or downstream automation. Record confidence and exceptions, especially where a wrong invoice figure, policy clause, or name would cause a real-world error.

The resulting learning project could ingest an approved set of knowledge documents, transform them into searchable chunks, run hybrid retrieval, answer a question with cited source passages, and reject an unsupported answer. Add a test corpus of both answerable and unanswerable questions; assess whether the response is factually supported, not merely convincing.

Apply vision, language and speech capabilities to real requirements

Visual tasks include image understanding, extraction of objects or descriptions, and generation or editing when appropriate. A candidate should know how to choose a capability for a concrete task, verify that the model handles representative images, and manage privacy and safety limitations. Where a system generates visual material, consider provenance, inappropriate content, watermarking or other platform controls, and how failures will be reviewed.

Text analysis includes entity and topic extraction, sentiment or tone detection, structured outputs, and translation. In a document workflow, ask whether the system must return a strict schema, support multiple languages, or preserve the distinction between facts and interpretations. Use test cases with ambiguous phrasing, missing fields, specialized terminology, and confidential content. A model that returns syntactically valid JSON can still extract an incorrect value.

Speech-enabled applications introduce audio quality, recognition accuracy, latency, consent, language support, and accessibility requirements. Implement speech-to-text or text-to-speech only where the use case warrants it, and test real background noise and accents rather than ideal sample audio. AI-103 includes speech within the current text-analysis objectives; it is not necessary to invent a separate legacy speech exam domain.

Some old AI-102 tutorials teach Language Understanding (LUIS), QnA Maker, or their original bot-era integrations as new deployments. These are historical architectures, not default starting points. Microsoft confirms that LUIS authoring and runtime endpoints were fully retired on March 31, 2026; organizations must migrate to supported Azure Language or Foundry patterns. Microsoft also publishes guidance for moving historical QnA Maker solutions. Do not build a new current-exam lab around discontinued endpoints.

Secure, evaluate and operate the solution

Security starts with clear resource ownership and identity boundaries. Use appropriate Azure role-based access control, managed identities or other supported keyless access patterns, private connectivity where needed, and logging that avoids unnecessary exposure of secrets or personal data. Document which data a model or agent can read, how long prompts and outputs may be retained, and who can authorize tool actions. Compliance duties depend on the applicable jurisdiction and workload; passing an exam does not establish regulatory compliance.

Responsible AI should translate into engineering checks. Create representative evaluation data, including difficult cases and known failure modes. Examine factual grounding, harmful or disallowed outputs, demographic or language performance differences where relevant, prompt injection attempts, and unsafe tool use. Configure available safety filters and guardrails, but verify their effectiveness against the application’s real risks. When uncertainty or harm is significant, route the case to a human reviewer with the supporting evidence.

Operational monitoring should cover availability, latency, token or service usage, retrieval performance, rate-limit errors, and quality changes. Review traces and alerts to identify regressions after updating prompts, indexes, tools, models, or SDK versions. Keep an appropriate rollback path for deployments and capture the version of each component used in production. Real user feedback is valuable, but it must be interpreted alongside tests and logs rather than becoming a substitute for measurable evaluation.

A focused learning project for AI-103

Build one end-to-end application rather than collecting unrelated demos. For example, create a knowledge-grounded support assistant in Python that uses a Foundry project, an Azure AI Search index, a restricted lookup tool, and a small set of approved documents. Record requirements, access rules, expected answers and unresolved questions before implementation. The project should exhibit decisions across the actual AI-103 objectives: configuration, retrieval, agent behavior, text or document handling, evaluation and operational controls.

When the application works under normal conditions, test what happens when documents are stale, an API call fails, a retrieved page contains a malicious instruction, or a user requests an action beyond the agent’s permission. Track the result and make a specific correction. Assess the impact of changes in source documents, model versions and access permissions. This practice builds transferable skills and makes exam study more meaningful than memorizing names from a retired blueprint.

Consult the current Microsoft Learn study guide for gaps you have not encountered in the project. Use the official exam sandbox to understand the delivery interface and any available Microsoft practice assessment to diagnose understanding. Avoid claiming a fixed preparation period, a guaranteed number of exam questions, or a promised passing rate: those depend on the candidate and on Microsoft’s current published policies.

How to use AI-102 history without misleading current candidates

It remains valuable to understand how Azure AI Vision, language analysis, document extraction, conversational interfaces and search evolved. Older AI-102 materials can help explain architectural tradeoffs or the migration of production applications, especially when a team still maintains historical integrations. But legacy content should identify its status plainly and avoid presenting unavailable exams or retired endpoints as recommended new implementations.

For a new candidate in October 2026, the relevant decision is whether AI-103 matches the intended Azure AI engineering role. Read the live issuer objectives, confirm that the assessment and credential remain available in the candidate’s region, and practice implementing secure and observable AI workflows. Certification is evidence of assessed knowledge; reliable delivery depends on ongoing technical judgment, testing, and experience.

Related Posts

• Mastering the Azure AI Engineer (AI-102) Exam

• Your Ultimate Guide to Crushing the Microsoft AI-102 Exam

• How to Ace the Microsoft Certified Azure AI Engineer Exam (AI-102)

• Is Coding Required for Microsoft Azure AI? A Simple Guide for Beginners

• Azure AI Fundamentals 2025: Smart Career Move or Just Hype?

• Should You Pursue the Microsoft Azure AI Fundamentals Certification?

• Unlocking Career Potential with the Microsoft AI-900 Certification

• Mastering the Basics: Your Guide to Microsoft Azure AI

• The Ultimate Overview of Microsoft AI-900: What You Should Know

• Azure AI Engineering After AI-102: Practical AI-103 Learning Roadmap