Skip to content
THE EXAMTOPICS EDITORIAL Explore practice exams

MICROSOFT

Microsoft AI-102 (Retired): Skills That Carry Forward to AI-103

Map useful AI-102 language, vision, extraction, retrieval and security skills to modern Microsoft Foundry AI-103 application design.

Microsoft’s Azure AI Engineer Associate examination, AI-102, retired on June 30, 2026. The Microsoft AI-103 exam, which became available in June 2026, is the replacement examination route toward the Azure AI Apps and Agents Developer Associate credential. The migration is more substantial than a change of code: the current guide adds Foundry-backed agent workflows, explicit evaluation and safety controls, secure access to knowledge and tools, and modern deployment practices. An older AI-102 course is therefore a technical reference, not current scheduling or objective guidance.

Keep the durable engineering principles—language understanding, computer vision, document extraction, safe APIs—while reading the current AI-103 blueprint independently. Microsoft groups the new exam into planning and management (25–30%), generative AI and agents (30–35%), vision (10–15%), text analysis (10–15%) and information extraction (10–15%). A migration plan should test the whole application, not map one retired service name to one new tool. This guide follows a maintenance-request workflow to show what changes and what remains sound.

Keep the engineering problem; recheck the service contract

Imagine a facilities team that receives reports about damaged equipment. Users submit a short text description, sometimes a photograph and occasionally a scanned maintenance form. The application must recognize the asset, retrieve the latest repair procedure, summarize the issue and prepare a case for a technician. AI-102-era coursework might teach useful language, vision and document extraction techniques. In a modern Foundry solution, the engineer also needs to decide how that evidence is grounded, which tool may prepare the case and what remains subject to an authorized approval workflow.

Start by writing an output contract. An extracted asset identifier must correspond to a real source; a reported defect must be tied to visible evidence or explicitly marked uncertain; and the case preparation operation must not quietly turn into a repair authorization. Then choose the capabilities that meet those checks. This produces a clearer study strategy than attempting to translate each old service name mechanically into a new product name.

For this workflow, set an explicit authorization boundary before using any model. The employee may submit a photo or retrieve a manual that their account is allowed to read, but only a separate authorized workflow should create or approve a work order. Give the retrieval layer scoped access to source documents, use managed identity and private networking where they solve the relevant threat, and log which source supported each generated field. When the extracted asset ID is absent from the inventory, stop and request human clarification. A confident paragraph is not a substitute for an identified asset and approved maintenance action.

Legacy study skill Durable concept Modern application decision
Text analytics and language detection Classify and extract signals from language Use purpose-specific text capabilities or a generative model, with validated labels and sensitive-data controls
Computer vision Interpret visual inputs under measurable error Distinguish visual understanding, structured extraction, image generation and safety checks
Document Intelligence / OCR Preserve fields and layout from complex sources Choose structured extraction or Content Understanding before putting documents into a RAG/agent pipeline
Question answering / knowledge sources Locate authoritative information Design ingestion, hybrid/vector/semantic retrieval, access filters, citations and freshness
Deployment and security Run capabilities under controlled identities Configure project identity, managed authentication, network paths, evaluation and release checks

The table is a learning bridge, not a claim that the platforms are exact one-to-one replacements. A technology can be supported in a modern service while using different project objects, SDK clients, quotas and lifecycle contracts. Validate the exact supported path before reusing code.

Language processing: preserve the business outcome

An AI-102 text-extraction exercise might identify entities and sentiment in a complaint. That remains a relevant engineering problem. The modern developer must also decide whether a deterministic label is sufficient, whether generation adds value and how the result will be checked before becoming part of a case workflow.

For the facilities team, classify the message into a support category, extract the reported asset identifier and quote the source phrase that supports the classification. If a message mentions two devices and it is unclear which is broken, the accepted output is not a confident guess: it is a flagged ambiguity. When translating a request from another language, preserve measurements, safety instructions and domain terms, then evaluate meaning rather than only grammatical fluency.

Vision and document processing: avoid forcing everything into one model

A photo of a damaged valve and a scan of its inspection form require different evidence. The photo may need multimodal reasoning about visible components. The form may require OCR, layout-aware field extraction and comparison with structured maintenance data. A modern implementation may combine both, but it should not treat a fluent image description as an authenticated equipment ID or maintenance history.

Suppose the scanned form has two tables and a handwritten replacement date. Test whether the extraction preserves row relationships, units and the distinction between inspection date and required repair date. Check the result against the original page. For the photograph, record what the model can actually see and prohibit it from asserting an invisible internal defect. Content Understanding and document extraction depend on preserving layout and provenance before results enter retrieval or an agent workflow.

Retrieval: from a static question-answering demo to controlled evidence

Older knowledge-base demos can still teach that the system must locate relevant source material before producing a response. The modern challenge is broader. The facilities assistant must retrieve the correct revision of a repair bulletin, filter out procedures the technician is not authorized to access and expose which passage supports a recommendation.

Create a deliberate test. Index two policy revisions; grant one test user access to both and another user access only to the public maintenance guide. Ask each user for the same procedure. The correct path must obey access and version rules before the generation step. If the index is stale, a grounded but outdated answer can still be wrong. If retrieval finds no eligible passage, the application should say so rather than instruct the model to invent one.

This is the RAG and grounding skill behind a reliable current application: good retrieval is a maintained evidence pipeline, not a one-time upload.

Agents and state: a new control problem, not a new label

A current application may let an agent choose between a policy search, a document analyzer and a read-only equipment lookup. It could then prepare a structured repair-request draft. That flexibility creates an additional question: does the model have permission to choose its next evidence-gathering step, and where is the irreversible business action controlled?

Model choice also belongs in the migration review. The simplest solution to a narrow extraction problem may be a deterministic parser or a specialist document model, whereas a multimodal model adds value when a photograph and a maintenance narrative must be interpreted together. Revisit model selection for real workloads only after recording required fields, acceptable errors, response time and disclosure limits. Compare evidence on the same small test set; do not infer that a newer general-purpose model is automatically safer or more accurate for every extraction task.

Keep durable case state in the application rather than treating the conversation transcript as an authoritative transaction record. If creating a draft times out, use an operation identifier and the case-management system’s status to check whether it was created. Do not replay a write just because the model did not receive a response. When an agent is asked to approve a repair, the application should enforce the real human approval requirement even if the model thinks the case is obvious.

Build a small migration lab instead of copying old notebooks

Start from a current Microsoft Foundry project endpoint. Use identity-based authentication appropriate to your environment, and configure an approved model deployment separately from source code. For one sample equipment report, retrieve an authorized record, parse the submitted form, request a bounded explanation and produce a draft with source references and a review-needed state.

Then inject a failure at each boundary. Deny the storage identity and inspect the authorization error. Break private DNS and show that network connectivity is a different issue. Supply an outdated bulletin and see whether provenance checks reject it. Interrupt the draft operation and verify that the application does not duplicate the case. A bounded Foundry Python implementation illustrates the project and agent boundaries, but the installed SDK version and resource type still require verification in the target tenant.

A useful migration checkpoint is a document-grounded support assistant for a controlled internal policy library. Define three test documents: a current policy, a superseded version and an access-restricted appendix. The retrieval layer should preserve version and access metadata instead of passing whichever passage happens to rank highest. Ask the assistant a question whose answer differs between policy versions. Record the document identifier it used, whether the user was authorized to see it and what the application does when only obsolete or inaccessible evidence is available. A fluent answer is not a passing test if it cites the wrong revision or crosses the permission boundary.

Repeat the checkpoint with an unavailable search index and with an authorized update arriving during a long conversation. Specify which state belongs to the application, which state belongs to the retrieval service, and which actions require human review. For each failure, write the expected visible outcome before considering a new Foundry SDK client or agent runtime. These are proposed offline test cases, not executed results: actual package compatibility, role configuration, search filtering and rollback behavior require a version-pinned environment and separate technical evidence.

What not to carry forward

Do not use the AI-102 exam fee, retired test outline, exam-day instructions or older practice-score targets as current certification policy. Do not present a legacy API snippet as production-safe merely because its concept remains familiar. Also avoid declaring that every newer generative task belongs in an autonomous agent. Some applications should remain deterministic workflows with a model used only for classification or drafting.

For AI-103 assessment preparation, obtain the dated official objective document and map each exercise to the tasks assessed by the applicable Microsoft AI-103 version. Confirm any issuer practice-assessment availability and access conditions rather than relying on archived links or an unverified domain count. The enduring part of AI-102 preparation is applied engineering judgment: specify an outcome, choose suitable services, prove source and permission boundaries, evaluate edge cases and control the path from recommendation to action.

Continue exploring the technologies and certification concepts that matter most to you.