AI-102: The Retired Azure AI Engineer Exam

AI-102, Designing and Implementing a Microsoft Azure AI Solution, retired on June 30, 2026. It should no longer be presented as an available examination or as the current route to the Azure AI Engineer Associate credential. The original syllabus still has educational value: it provides a snapshot of how Azure AI engineers planned secure AI resources, used speech and language services, processed documents, built search experiences and operated generative applications before Microsoft's certification transition. Candidates now considering a comparable development path should review AI-103, which assesses developing AI apps and agents using Microsoft Foundry, while recognizing that the newer scope is not an item-by-item continuation of AI-102.

Historical AI-102 objectives can still explain capabilities encountered in older courses, résumés and architecture discussions. The important distinction is between learning a technology that remains useful and registering for an examination Microsoft has withdrawn. A feature mentioned in the old objectives may still exist, may have a newer product name or may now be presented differently in Microsoft Foundry. The AI-102 exam retired on June 30, 2026; new candidates should use the applicable current Microsoft AI development certification objectives.

What the former certification assessed

AI-102 targeted developers who could translate a business AI requirement into a working Azure solution. Its objectives ranged from resource provisioning and identity controls to computer vision, language understanding, search, generative AI and document processing. A candidate was expected to know how services interact rather than simply list product names. For example, a support assistant may combine document ingestion, search indexing, prompt orchestration, a model endpoint and monitoring. Understanding those dependencies was more important than treating the assistant as one mysterious AI resource.

The historical exam also spanned development and operations. Engineers selected service tiers, considered quotas, configured access and handled resource secrets. Azure's managed identities and role assignments reduce reliance on hard-coded credentials; virtual-network restrictions and private endpoints may reduce exposure of supported services. Costs follow the services and traffic patterns actually used. A proof of concept that works using broad permissions and uncontrolled model calls may be a poor production design, even if its demo prompt produces fluent text.

The transition from AI-102 to current development credentials is relevant to anyone studying from old materials. Microsoft distinguishes the retired Azure AI Engineer Associate path from the newer Azure AI Apps and Agents Developer Associate credential. There is no safe basis for claiming an AI-102 voucher or course guarantees success on AI-103, nor for pretending that every service in a 2025 objective list is still named the same way. Verify the new credential directly rather than editing only the exam code in an old study plan.

Design search and knowledge-grounded applications

One durable AI-102 skill was choosing the right approach to find evidence before generating an answer. A conventional keyword index serves exact term matching well, while vector similarity can retrieve semantically related passages. Hybrid retrieval combines signals, and reranking may improve ordering for difficult questions. The design begins with source authority, refresh frequency, document access and desired citation quality. A search index containing stale documents or incomplete metadata can undermine an otherwise competent language model.

Document ingestion is a pipeline rather than an upload button. Files can contain scanned images, tables, uneven reading order, embedded metadata and mixed-language content. Extraction should preserve useful structure and record the source, section and permissions associated with each chunk. Oversized chunks may dilute relevance; excessively small fragments can lose context. Retrieval evaluation should check whether the correct passages were actually returned before blaming a model for an unsupported answer. Answers that sound confident without a matching source need special scrutiny.

The grounding of enterprise agents with trusted data extends this historical concern into modern workloads. Indexing sensitive material does not remove the need for document-level authorization and safe connector scopes. Model prompting cannot substitute for retrieval permissions. Test with differently privileged users and record whether references come from the right document version. A successful question-answering demo with an administrator identity proves little about production access isolation.

Apply vision, language and speech to real workflows

Computer vision decisions start with the task. Extracting printed text from a form, recognizing objects in a photograph and describing a scene require different output guarantees. A production system may need bounding regions, confidence indications or machine-readable fields, not a narrative caption. Images can be blurry, rotated, poorly lit or dominated by irrelevant background. Evaluate representative failures and decide what requires human correction before extracted values enter an authoritative business record.

Language services similarly differ by purpose. Key-phrase extraction, sentiment analysis, named-entity recognition, translation and summarization are not substitutes for one another. A legal disclosure needs faithful translation; customer-service triage might value intent and entity extraction; a financial document might require exact currency values and provenance. Treat generated structured text as untrusted until validated against a schema and source evidence. Model improvements do not eliminate the need for domain-specific test sets.

Speech workflows add time, audio quality and privacy considerations. Transcription errors may cluster around proper nouns, accents or noisy environments. A call-center assistant should know which text comes from transcribed audio, which was inferred and which was retrieved from a knowledge base. Delayed or inaccurate translations can create operational errors when the transcript feeds another system. Audio retention, consent and access policies must be considered separately from whether a speech model correctly recognizes a demonstration phrase.

Operate generative AI with safety and reproducibility

Generative model selection depends on context-window behavior, modality, latency, throughput, cost and evaluation results. The largest model is not automatically the most appropriate. A small classifier may be better for deterministic routing; a larger model may be justified for complex language generation; a tool-augmented application may need both. Engineers should benchmark representative prompts and worst-case traffic instead of basing a production decision on a single polished example.

Prompt changes are software changes. Altering the system instruction, retrieval settings or generation parameters can shift answer style, citation behavior and failure rates. Keep versioned configurations, reference datasets and release criteria. The evaluation of AI agents before production is relevant when assessing grounding, accuracy, safety and whether tool actions remain within policy. A prompt that performs well on average can still fail catastrophically on a small set of high-impact requests.

Security includes threats that do not look like traditional malware. A retrieved document could contain an instruction to ignore system rules, exfiltrate data or invoke a tool. Treat that content as task data, not as an authority above the application's instructions and server-side controls. The security implications of prompt injection are particularly important for document-grounded agents. Enforce authorization at service boundaries, validate tool arguments and avoid exposing powerful integrations solely because the model promises to act responsibly.

Connect old AI-102 knowledge to current roles

The current AI-103 examination explicitly includes developing generative and agentic solutions, computer vision, text analysis and information extraction using Microsoft Foundry. That overlaps the underlying engineering disciplines once explored under AI-102, but the modern scope puts stronger emphasis on agent tools, orchestration, evaluation, grounding and operational safeguards. Candidates should compare the current study guide directly with their existing knowledge rather than assume a retired practice test predicts the new exam.

A second adjacent path is AI-300, which emphasizes production operations for machine learning and generative AI. Development and operations meet at deployment boundaries but evaluate different abilities. Designing a text-extraction pipeline does not establish competence in repeatable release workflows, drift monitoring or incident rollback. Conversely, an operations engineer may automate model deployment without being responsible for every vision or translation feature. Choose a certification according to actual responsibilities.

The retired AI-900 fundamentals exam is also often mentioned alongside AI-102 in older roadmaps. Microsoft now uses AI-901 for its Azure AI Fundamentals credential. Fundamentals study helps with vocabulary, not with replacing advanced development experience. For a new candidate, a sensible sequence is to understand AI workloads and responsible use, build a narrowly scoped Foundry solution, then learn agent execution, monitoring and security with a current exam guide.

Consider a pre-retirement vision pipeline that identified text in scanned delivery receipts and sent extracted identifiers to a logistics system. A migration should test more than whether the new service returns JSON: compare page orientation handling, the association of labels to values, confidence behavior, and the treatment of unreadable handwriting. The downstream service should reject malformed account numbers even if the upstream model calls them plausible. Preserve representative input documents with no real customer information and record what human review was required at each stage. This turns an older exam lab into a reproducible engineering reference and can expose a risk that a casual feature-for-feature comparison misses.

A second case involves search permissions. Rebuilding an index under a new service account may accidentally make documents accessible beyond their previous audience. Test access using multiple identities and compare original source ACLs with filters applied during retrieval. Modernizing inference while weakening data isolation would be a regression, not a successful migration.

Reuse historic material without repeating historic mistakes

A useful migration project starts by cataloging an old AI-102 lab: the source documents, services, permissions, indexing logic, model configuration, evaluation data and monitoring. Mark each dependency as still supported, renamed, altered or requiring replacement. Verify the current service documentation before changing the architecture. A training lab may contain obsolete setup screens or API details even when its underlying retrieval and data-protection lessons remain correct.

Build a second version of the same solution using current supported tooling, and compare behavior under identical scenarios. Check whether the retrieval finds the same evidence, whether a tool invocation is correctly authorized, how latency changes, and whether safety tests reveal different failures. Keep both configurations documented. This makes the transition a technical learning exercise rather than an exercise in swapping branding terms inside paragraphs.

Above all, do not treat archival evidence as booking guidance. Microsoft's retirement list and the AI-102 study guide confirm June 30, 2026 as the retirement date. An AI-102 page can still educate readers about enduring architecture, but any call to prepare for or schedule AI-102 as if it were current would mislead them. The correct endpoint for a new exam plan is the active official certification listing and its currently effective study guide.