StartAzure AI engineering, from exam to production
Azure AI Engineering
Learn every Azure service you need to build AI apps and agents, and pass AI-103 on the way.
The examAI-103: Developing AI Apps and Agents on Azure
Azure AI Apps and Agents Developer Associate. Replaced AI-102, which retired on 30 June 2026. Pass mark 700. Python. Renew yearly with a free online assessment.
Track AThe exam, domain by domain
Five domains, weighted as the official study guide (skills measured as of 16 April 2026).
Practice30 questions, explained
Clear stems, hard choices, a deliberate runner-up. Each explained: premise, mind map, the stem's evidence, the rule.
Scope: this guide follows the AI-103 exam exactly: models, agents, RAG, evaluation, observability, security, responsible AI, vision, text, speech and document extraction. Running agents in production (gateway, hosting, state, release) is outside the exam and outside this guide.
Facts checked against Microsoft Learn on 8 October 2026. Azure AI moves fast: names, versions and preview/GA status change. Each lesson links its source; re-check before relying on a detail.
RoadmapWhat to learn, in what order
Exam weight, then an eight-week plan
More than half the exam is planning, managing and building generative and agentic solutions. Spend your time accordingly.
Week 1Foundry foundations
- Resource vs project, deployment types
- Keyless auth and Foundry roles
- Lessons 1b, 1c · Lab: first keyless call
Week 2Models and the Responses API
- Choosing models, model router
- Reasoning parameters
- Lessons 1a, 2a
Week 3Agents
- Prompt vs hosted agents, tools, toolbox
- A2A, MCP approvals
- Lessons 2c, 2d
Week 4RAG and Search
- Indexes, skills, hybrid + semantic
- Knowledge bases, Foundry IQ
- Lessons 2b, 5a, 5b
Week 5Quality and safety
- Guardrails, evaluators, red team
- Tracing and monitoring
- Lessons 1d, 2e
Week 6Vision, text, speech, documents
- Image/video generation, Content Understanding
- Structured outputs, Translator, Speech, DI
- Lessons 3a–3d, 4a–4c, 5c
Week 7Practice and gaps
- All 30 questions; read every explanation
- Redo lessons behind wrong answers
- Microsoft Learn practice assessment
Week 8Mock exams
- Timed mock exams in exam mode
- Review every miss on Microsoft Learn
- Then book the exam
ServicesEvery Azure service in AI engineering
The services map
How deeply the exam tests each service, and where it sits in our agent harness.
Must know deeply9 services · about 60% of the exam
Foundry, models, Agent Service, guardrails, evaluation, Search, Foundry IQ, Entra, App Insights.
Know well7 services · one domain each
Content Understanding, image and video models, vision models, Language, Translator, Speech, Document Intelligence.
Recognise7 services · supporting questions
Agent Framework, Storage and Cosmos DB, networking, Key Vault, Functions, Monitor alerts, Video Indexer.
Not tested7 · production only
Gateway, Container Apps, PostgreSQL, Prometheus and Grafana, release tooling, durable waits, governance.
| Exam depth | Service | What the exam asks | In our harness |
|---|---|---|---|
| Must know deeply | Microsoft Foundry (resource + projects) | Endpoints, deployment types, quotas and 429s, CI/CD with azd and the SDK | Model hosting; evaluation and tracing views |
| Must know deeply | Foundry Models (incl. Azure OpenAI) | Model choice, model router, Responses API, v1 routes, reasoning parameters | GPT and Claude behind the gateway |
| Must know deeply | Foundry Agent Service | Prompt vs hosted agents, conversations, tools, toolbox, MCP approvals, A2A | Optional for simple agents; ours run their own loop |
| Must know deeply | Foundry guardrails / Content Safety | Harm categories, Prompt Shields, groundedness, protected material, PII; inheritance | Detectors inside our positioned rules |
| Must know deeply | Foundry evaluation + AI Red Teaming Agent | Built-in evaluators, cloud evaluation, Groundedness Pro, attack success rate | Next to AgentTwin in offline evaluation |
| Must know deeply | Azure AI Search | Indexes, indexers, skillsets, integrated vectorization, hybrid + semantic ranker | Knowledge port option when the corpus grows |
| Must know deeply | Foundry IQ (knowledge bases) | Agentic retrieval for agents over MCP; extractive GA, synthesis preview | Same option, for agents |
| Must know deeply | Microsoft Entra ID + Azure RBAC | Keyless auth, managed identity, Foundry roles | Every identity |
| Must know deeply | Application Insights | Tracing models and agents, token analytics, monitoring dashboard | Spans store |
| Know well | Azure Content Understanding | Analyzers for documents, images, audio, video; bring-your-own deployments | Claim documents and photos |
| Know well | Image and video models (gpt-image-2/2.5, Sora 2) | Generate, mask edits, video remix, restrictions, C2PA | Rarely needed for claims |
| Know well | Vision-capable chat models | Captions, visual Q&A, alt text through prompts | Reading claim photos |
| Know well | Azure Language | PII and NER (core); sentiment and summarisation are legacy | PII check beside our redaction |
| Know well | Azure Translator | NMT vs LLM per request, tone, document translation | Hindi and regional languages |
| Know well | Azure Speech + Voice Live | Speech to text and back, custom speech, translation, voice agents | Voice channel, call transcription |
| Know well | Azure Document Intelligence | Prebuilt and custom form models, containers | Standard forms; disconnected sites |
| Recognise | Microsoft Agent Framework | Successor to Semantic Kernel and AutoGen; answer for new orchestration | Alternative loop for .NET teams |
| Recognise | Azure Storage · Azure Cosmos DB | Agent Standard setup: your own storage, threads, vectors | — |
| Recognise | Private networking (private endpoints, VNet) | Foundry inbound privacy; agent outbound isolation | Private endpoints everywhere |
| Recognise | Azure Key Vault | Secrets and customer-managed keys | No keys in code |
| Recognise | Azure Functions | The Functions agent tool; custom Search skills | — |
| Recognise | Azure Monitor alerts | Alerts on evaluation and safety signals | Action groups: page, ticket, trend |
| Recognise | Azure Video Indexer | Listed in the study guide's docs; Content Understanding is the current answer | — |
| Not tested | API Management (AI gateway) | Token limits, metrics, failover in front of models | The one door for model calls |
| Not tested | Azure Container Apps | Containers, revisions, jobs | Agents, MCP servers, jobs |
| Not tested | Azure Database for PostgreSQL | State, ledger, approvals, findings, pgvector | All agent state |
| Not tested | Managed Prometheus · Managed Grafana | Counters, alert rules, dashboards | Two signals, two stores |
| Not tested | App Configuration · Bicep · GitHub Actions | Config and flags; infrastructure as code; OIDC pipelines | Every environment |
| Not tested | DBOS / Temporal / Durable Task Scheduler | Durable approvals and long waits | Payout approvals |
| Not tested | Defender for Cloud · Purview · Azure Policy | Threat protection, data governance, resource guardrails | Governance rail |
LessonsThe AI-103 exam
Five domains, one lesson per skill group
Each lesson: the idea in plain words, key facts, code, exam traps, where it shows up in our harness, and a hands-on lab.
Domain 1 · 25–30%Plan and manage an Azure AI solution
Domain 2 · 30–35%Implement generative AI and agentic solutions
Domain 3 · 10–15%Implement computer vision solutions
Domain 4 · 10–15%Implement text analysis solutions
Domain 5 · 10–15%Implement information extraction solutions
Practice30 questions
Clear stem, hard choice
Every question has one deliberate runner-up: an answer that wins if one fact in the stem changes. Find the sentence that decides it.
These are original practice items in the exam's style, not real exam questions. After the 30 here, take Microsoft Learn's free practice assessment for AI-103.
GlossaryOld names, new names, and terms
Renames you'll meet in old material
- Azure AI Studio → Azure AI Foundry → Microsoft Foundry
- The platform. The old portal is now "Foundry (classic)".
- Azure AI services / Cognitive Services → Foundry Tools
- e.g. "Azure Speech in Foundry Tools". The resource is now a Foundry resource.
- Hub-based project → Foundry project
- Hub projects are classic; new work uses Foundry projects.
- Threads / runs / assistants → conversations / responses / agent versions
- Agents v2, on the Responses API.
- Azure AI User → Foundry User
- Same role ID. Azure AI Developer belongs to Azure ML hubs.
- Content filters → guardrails
- A named set of controls: risk, intervention point, action.
- Knowledge agents → knowledge bases
- Azure AI Search agentic retrieval; Foundry IQ for agents.
- Connected Agents → A2A tool
- Connected Agents isn't in the new Agent Service.
- Prompt flow, Foundry Workflows → Agent Framework
- Prompt flow retires 20 Apr 2027; Workflows 1 Dec 2026.
- Monthly api-version → v1 routes
- /openai/v1/ with the standard OpenAI client.
- DALL-E 3 → gpt-image-2 / 2.5
- DALL-E 3 retired March 2026.
- LUIS → (retired) · CLU, sentiment, summarisation → legacy
- Use LLMs with structured outputs for new text analysis.
- RRF
- Reciprocal Rank Fusion: how hybrid search merges keyword and vector results.
- PTU
- Provisioned throughput units: reserved capacity for predictable latency.
- ASR
- Attack Success Rate, the AI Red Teaming Agent's metric.
- C2PA / Content Credentials
- Signed provenance metadata on generated images.