AI-901 FDP · Day 3 · All decks
DAY 3 · 07-Jul

Foundations, Responsible AI,
Foundry & Computer Vision

Microsoft Certified: Azure AI Fundamentals — Exam AI-901 (replaced AI-900, Apr 2026)

Domain 1 · Identify AI concepts (40–45%)

Domain 2 · Implement in Foundry (55–60%)

AI-901 is NOT AI-900

Old (AI-900, retired)

  • Individual services in isolation
  • No coding
  • Form Recognizer, LUIS, QnA Maker…

New (AI-901)

  • Build in Microsoft Foundry
  • Assumes basic Python + SDK/REST
  • Agents + Content Understanding
Teach lab-first. 55–60% of the exam is implementing in Foundry.

Exam facts

CodeAI-901Pass700 / 1000 (scaled)
Questions~40–60Time~45–60 min
Cost~$99 USDExpires?No (fundamentals)

Domain 1

Identify AI concepts & capabilities

40–45%

Domain 2

Implement with Microsoft Foundry

55–60%

The whole course, one map

mindmap
  root((AI-901))
    Identify (40-45%)
      Responsible AI
      How GenAI works
      AI workloads
    Implement in Foundry (55-60%)
      Deploy & prompt models
      Foundry SDK
      Agents
      Text & Speech tools
      Vision & image/video gen
      Content Understanding

AI workloads — scenario → capability

Create

text · images · video · code → generative models

Act

tools + multi-step → agent

Understand text

sentiment · entities · keywords · summary

Speech

STT (recognition) · TTS (synthesis)

Vision

classify · detect · OCR · describe

Extract

structured fields → Content Understanding

Responsible AI — F.R.I.T.A.P.

flowchart TD
  RAI[Responsible AI] --> F["Fairness: no bias"]
  RAI --> R["Reliability & Safety: system safe"]
  RAI --> I["Inclusiveness: everyone"]
  RAI --> T["Transparency: disclose + limits"]
  RAI --> A["Accountability: people answerable"]
  RAI --> P["Privacy & Security: protect data"]

The trap examiners love

Reliability & Safety

The system behaves correctly/safely under unexpected conditions. Rigorous testing.

Accountability

People / orgs are answerable. Governance, oversight board, compliance.

"An oversight board reviews the AI's decisions and answers for them" → Accountability, not reliability.

Check yourself

A loan model approves fewer applicants of one ethnicity. Which principle is violated?
Fairness — unequal outcomes across demographic groups = bias. Transparency = disclosing it's AI; Accountability = who answers for it.

How generative AI models work

  • Token — unit of text; limits & billing are token-based
  • Embedding — numeric vector of meaning → search / RAG
  • Transformer — architecture using attention
  • LLMs predict the next token (not a DB lookup) → can hallucinate
  • Fix hallucination with grounding / RAG (give it your data)
  • Multimodal = also handles images/audio

Microsoft Foundry — where you build

flowchart LR
  A["Foundry (ai.azure.com)"] --> B[Hub] --> C[Project]
  C --> D[Model catalog] -->|deploy| E["Deployment
name + endpoint + filter"]
  C --> G["Agents"]
  C --> F["Tools: Speech, Content Understanding"]
  E --> H[Endpoint] --> I["Your app (SDK/REST)"]

Hub = top workspace · Project = a solution · Deployment = a callable model.

Deploy & call a model

  1. Model catalog → Deploy (choose type: Standard / Global Standard / PTU / Batch / Serverless)
  2. Note the deployment name — this is what code references
  3. Open the playground → set system message → send user prompt
  4. Tune temperature max_tokens top_p
temperature/top_p = randomness. max_tokens = length. Never mix them up.

Check yourself

In the SDK, what do you pass as model=?
Deployment name. You reference the deployment you created, reached via its endpoint (auth = key or Microsoft Entra ID).

Computer vision — capabilities

CapabilityOutputLocation?
Image classificationwhole-image labelNo
Object detectionlabels + bounding boxesYes
OCR (read)text in imagetext region
Image descriptioncaption / tagsmixed
Just a label → classification. Label + where (and multiple objects) → detection.

How vision models work

Pixels

image = grid; convolutions find edges/shapes

CNN

classic deep net for classify/detect

ViT / Multimodal

attention over image patches; reason over image + text

In AI-901 you mostly send an image to a multimodal model and prompt it.

Interpret an image (multimodal)

resp = client.chat.completions.create(
  model="gpt-4o",   # multimodal deployment
  messages=[
    {"role":"system","content":"Describe images for accessibility."},
    {"role":"user","content":[
      {"type":"text","text":"Describe this and read any text."},
      {"type":"image_url","image_url":{"url": IMAGE_URL}}
    ]}
  ])
print(resp.choices[0].message.content)

The image rides inside the user message content.

Generate images & video

Image-gen

DALL·E 3 / GPT-image → still images from text

Video-gen

text → short video clips new

Multimodal

understands images/video as input

img = client.images.generate(model="dall-e-3",
     prompt="A watercolor fox in snow", size="1024x1024")
print(img.data[0].url)

Check yourself

You must create a short promo video from a text brief. Which model type?
Video-generation. Multimodal understands visuals; image-gen makes stills; video-gen makes clips.

Day 3 lab & homework

Lab 1 — Deploy your first model

Create a project → deploy GPT-4o-mini → chat in the playground → move temperature/max_tokens.

Vision playground

Upload a photo to GPT-4o (describe + read text); deploy an image model and generate 2 images.

Then: Practice Test 1, Q1–20.

Recap

  • AI-901 = Foundry + agents + Content Understanding + basic code
  • F.R.I.T.A.P. — reliability(system) ≠ accountability(people)
  • Tokens, embeddings, grounding fixes hallucination
  • Deployment name + endpoint; temperature≠length
  • Multimodal understands · image-gen/video-gen create

Next → Day 4: Text, Speech & Content Understanding