AI Wisdom

Microsoft associate level

Microsoft Certified: Azure AI Apps and Agents Developer Associate

Code
AI-103
Length
120 minutes
Pass mark
700 out of 1,000
Price
Set by country, shown when you book
Valid for
12 months, renewed by a free online assessment
Study time, our estimate
60 to 80 hours

This replaced AI-102 on June 19, 2026, and it is the one to get if you build on Azure. It is an agents and Foundry exam now. Generative and agentic work is the largest domain, and the classic vision and language services have shrunk to around 10 to 15% each.

Take it if

You build AI apps in Python on Azure. It also counts toward the AB-100 expert certification.

Skip it if

Your production stack is not on Azure. The questions are Foundry-specific and will not carry over.

Before you start

Check yourself against this list first. If something is missing, close that gap before you pay for an exam.

  • You develop apps in PythonMicrosoft expects this for the exam.
  • You know core Azure servicesIdentity, networking and resource basics are assumed.
  • You understand generative AI basicsAI-901 covers this if you are new.

Helpful first: AI-901 Next: AB-100AI-300

All exam details
Format
Usually 40 to 60 questions, per Microsoft
Length
120 minutes
Pass mark
700 out of 1,000
Price
Set by country, shown when you book
Valid for
12 months, renewed by a free online assessment
Delivery
Pearson VUE, online or test center
Prerequisites
None formally. Python and Azure experience expected.
If you fail
Wait 24 hours, then 14 days

What the exam covers

  • Plan and manage an Azure AI solution25–30%
  • Implement generative AI and agentic solutions30–35%
  • Implement computer vision solutions10–15%
  • Implement text analysis solutions10–15%
  • Implement information extraction solutions10–15%

What to focus on

  • Choosing the right Foundry service, model and retrieval method for a task. It opens the planning domain, which is 25 to 30% of the exam.
  • Building RAG and agents: tool schemas, memory, function calling, and multi-agent orchestration with approval steps.
  • Keyless security: managed identity, private networking and role policies.
  • Quotas, rate limits, cost, and monitoring drift, safety events and grounding quality.
  • Responsible AI as configuration: safety filters, evaluators, trace logging, and indirect prompt injection through text in images.
  • Azure Content Understanding and hybrid search for grounding pipelines.
Everything on one mapEvery topic and key idea in this certification.
mindmap
  root((AI-103))
    Plan and manage
      Keyless access
      Quotas and limits
      Agent governance
    Generative AI and agents
      RAG
      Agent design
      Orchestration
      Observability
    Vision, text and speech
      Multimodal
      Indirect injection
      Custom speech
    Information extraction
      Search modes
      Content Understanding
      Enrichment
How a Foundry RAG request flowsA map of the pieces AI-103 tests.
sequenceDiagram
  participant U as User
  participant A as Your app
  participant S as Azure AI Search
  participant M as Foundry model
  U->>A: Question
  A->>S: Hybrid query, keyword plus vector
  S-->>A: Top chunks
  A->>M: Prompt with chunks
  M-->>A: Grounded answer
  A-->>U: Answer with citations

Lesson 1 of 4, 25–30% of the exam

Plan and manage

A quarter of the exam is choosing services and keeping them safe and affordable.

Keyless access
Managed identity with role assignments removes the secret entirely.
Quotas and limits
Rate limits apply to bursts, so compare peak traffic with deployment capacity.
Agent governance
Use oversight modes, constraints and tool-access controls for agents that can act.
Keyless access to FoundryNo secret is stored anywhere in your app.
sequenceDiagram
  participant A as Your app
  participant E as Microsoft Entra ID
  participant F as Foundry
  A->>E: Request a token as the managed identity
  E-->>A: Short-lived token
  A->>F: Call with the token
  F->>F: Check the role assignment
  F-->>A: Response

Common trap

Rotating API keys when the requirement is no secrets at all.

Try it yourself

Move a Foundry app from an API key to managed identity.

Official docs: What is Microsoft Foundry?Provisioned throughput

Lesson 2 of 4, 30–35% of the exam

Generative AI and agents

The largest domain, and the reason this exam replaced AI-102.

RAG
Retrieve first, then generate from what you found. Evaluate for fabrication, relevance and safety.
Agent design
Define the role, goals, conversation tracking and tool schemas.
Orchestration
Multi-agent flows need safeguards and approval steps where actions matter.
Observability
Trace requests, watch token use, safety signals and latency.
Anatomy of a Foundry agentWhat you define when you build one.
classDiagram
  class Agent {
    role
    goals
    conversation tracking
  }
  class Tools {
    APIs
    search
    custom functions
  }
  class Memory
  class Safeguards {
    approval steps
    tool access
  }
  Agent --> Tools
  Agent --> Memory
  Agent --> Safeguards

Common trap

Treating a fluent answer as a correct one. Measure groundedness.

Try it yourself

Build an agent with one tool and memory, then add a human approval step.

Official docs: Foundry Agent ServiceRAG in Azure AI Search

Lesson 3 of 4, 10–15% + 10–15% of the exam

Vision, text and speech

Each looks small, but together they are worth up to 30%.

Multimodal
A model can take images in the prompt for captions and grounded answers.
Indirect injection
Text hidden in an image can carry instructions, so filter it.
Custom speech
Train on your vocabulary when domain terms are misheard.
How text in an image attacks a modelFilter visual input as well as typed input.
flowchart LR
  A["Uploaded image"] --> B["Hidden text instruction"]
  B --> C["Multimodal model reads it"]
  C --> D{"Filter in place?"}
  D -->|No| E["Model follows it"]
  D -->|Yes| F["Blocked and logged"]

Common trap

Skipping these because your job is agents. Weak smaller domains fail people.

Try it yourself

Add image input and speech to one agent.

Official docs: Custom SpeechWhat is Microsoft Foundry?

Lesson 4 of 4, 10–15% of the exam

Information extraction

Good retrieval depends on what you put into the index.

Search modes
Hybrid search combines keyword and vector, which handles exact codes and vague questions.
Content Understanding
OCR, layout and field extraction produce clean output for RAG.
Enrichment
Skills add structure during indexing.
Which search mode?Choose by what users actually type.
flowchart TD
  A{"What do users search with?"} -->|"Exact IDs and codes"| B["Keyword"]
  A -->|"Natural questions"| C["Vector"]
  A -->|"Both"| D["Hybrid"]

Common trap

Using vector search alone when users search by part number.

Try it yourself

Index scanned PDFs and compare keyword, vector and hybrid results on 10 queries.

Official docs: Hybrid searchContent Understanding

Study plan

  1. Weeks 1 and 2: Foundry and RAG

    Stand up a Foundry project, deploy a model, and build RAG on Azure AI Search with hybrid and vector queries.

  2. Week 3: agents

    Build an agent with tools and memory, then a multi-agent flow with a human approval step. Connect both to a Foundry project from Python.

  3. Week 4: secure and monitor

    Switch everything to managed identity, add content filters, turn on tracing, and run an evaluation for groundedness.

  4. Week 5: the smaller domains

    Cover image generation, multimodal understanding, speech and Content Understanding. Each is small, but together they are 30 to 45% of the exam.

  5. Exam week

    Take the free practice assessment, then the mock below, and revisit any domain under 70%.

Exam day and after

  • You can open Microsoft Learn during associate and expert exams, but the clock keeps running. It helps with one lookup and not with twenty, so use it for exact names and limits.
  • Most questions cover generally available features, though preview features can appear if they are widely used.
  • Try the free exam sandbox first so the interface is not a surprise on the day.
  • Study materials written for AI-102 are out of date. Check that anything you buy names AI-103.
Test yourself9 practice questions, scored by exam topic.
Start the mock exam

Keep going

Exam facts checked on October 31, 2026 against AI-103 study guide, skills as of April 16, 2026. Lessons are our own explanations, so use the official docs as the final word, and confirm price and dates before booking.