AI-901 · Microsoft Azure AI Fundamentals

Roadmap · 31 Aug – 13 Sep 2026 · Exam Mon 14 Sep

Fourteen days to Azure AI Fundamentals.

A day-by-day plan for exam AI-901, built against the skills outline published 15 April 2026. Over half of this exam is hands-on implementation in Microsoft Foundry, so more than half of this plan is spent writing code — not reading about it.

Exam code: AI-901
Domains: 2
Passing: 700 / 1000
Skills as of: 15 Apr 26
Languages: 13
Retirement: None

Ignore everything written for AI-900. That exam retired on 30 June 2026, and almost every third-party course, video, and question dump you will find still targets it. AI-901 has two domains instead of five, drops Azure Machine Learning entirely, and is 55–60% hands-on implementation with Python assumed. Microsoft also renamed the product surface in 2026 — Azure AI Foundry is now Microsoft Foundry, Azure AI Services are Foundry Tools, and the Assistants API is the Responses API. Check the publication date on anything you read outside this kit.

§ 01 · Strategy

How the fourteen days break.

Four phases, weighted to match the exam rather than the syllabus's page count. Concepts come first because they anchor everything else, but they get only four days — Domain 1 is 40–45% of the marks and by far the easier half. The Foundry implementation block gets seven days, including both weekends, because you cannot learn an SDK by reading about it. Weekdays run about two and a half hours; the four weekend days carry the labs and the long drills.

Days 01 – 04

Concepts

Responsible AI, how generative models work, model choice and deployment, and the scenario-to-service map.

Domain 1 · 40–45%

Days 05 – 08

Foundry

The SDK, conversations, agents, text and speech. The heaviest material, on the two longest days.

Domain 2 · objectives 2.1–2.2

Days 09 – 11

Multimodal

Images in prompts, image generation, and Content Understanding across four modalities.

Domain 2 · objectives 2.3–2.4

Days 12 – 14

Mastery

Two practice assessments, targeted repair of your five weakest topics, then a calm stop.

Full syllabus · 100%

The one rule that matters: Domain 2 is 55–60% of the exam and it is assessed as implementation. Every day from 05 onward has a lab attached. If you are ever short on time, cut the reading, not the code.

§ 02 · Anatomy

The two domains.

Microsoft publishes a skills-measured document listing every testable topic under two weighted domains. Every objective below is mapped to a specific day in this roadmap, and to a notes file and a quiz in this kit.

Domain 1

40–45%

Identify AI concepts and capabilities

Describe principles of responsible AI — fairness, reliability & safety, privacy & security, inclusiveness, transparency, accountability
Identify AI model components and configurations — how generative models work, choosing a model by capability, deployment options and configuration parameters
Identify AI workloads — generative and agentic AI, text analysis, speech, computer vision, information extraction
Common text analysis techniques — keyword extraction, entity detection, sentiment analysis, summarisation
Speech recognition and synthesis; computer vision and image-generation models
Techniques to extract information from text, images, audio, and video

Domain 2

55–60%

Implement AI solutions by using Microsoft Foundry

Create effective system and user prompts for generative AI models
Deploy a model and interact with it in the Foundry portal
Create a lightweight chat client application by using the Foundry SDK
Create and test a single-agent solution in the portal; build a client for an agent
Build a text-analysis application; respond to spoken prompts with a multimodal model; use Azure Speech in Foundry Tools
Interpret visual input in prompts; create new visual outputs; build an app with vision capabilities
Extract information from documents, forms, images, audio, and video using Azure Content Understanding

§ 03 · The Plan

Fourteen days, fourteen briefings.

Each day names what to learn, what to read, what to build, and four questions to test yourself against before you close the laptop. If you cannot answer the self-check, the day is not finished.

§ 03.a · Calendar

A glance across the fortnight. Tap any day to jump to its full briefing. Weekends carry the labs.

Week 1Concepts & SDK

Concepts

Stand up Foundry, then learn responsible AI

2h 30m

Concepts

How generative models actually work

2h 30m

Concepts

Choosing a model, deployment types, and…

2h 30m

Concepts

AI workloads — the scenario-to-service map

2h 30m

Foundry

Prompt engineering and the Foundry portal

2h 30m

Foundry

The Foundry SDK, conversations, and…

5h+

Foundry

Text analysis and speech, end to end

5h+

Week 2Multimodal & mastery

Consolidation

Consolidation — Domain 1 complete, Domain…

2h 30m

Multimodal

Visual input — images in prompts

2h 30m

Multimodal

Image generation and responsible AI for…

2h 30m

Extraction

Content Understanding — one service, four…

2h 30m

Mastery

Syllabus complete — first full assessment

2h 30m

Mastery

Targeted repair and full-bank drill

Phase I · Concepts

2 h study
30 m setup

Mon 31 Aug

Stand up Foundry, then learn responsible AI

Get a Foundry project and a deployed model working today. 55–60% of this exam is implementation and you cannot cram that in week two — every day from 05 onward assumes a working environment. Once the environment is up, start on responsible AI, the objective most likely to appear as a scenario question.

Notes 1.1 Responsible AI Quiz 1.1 Responsible AI Lab 01 First call

Learn

The six responsible AI principles and the trigger words that map a scenario to each
Fairness versus inclusiveness — outputs that disadvantage a group versus a system nobody can use
Guardrails: four harm categories, four severity levels, filtering both prompt and completion
Prompt shields, protected material detection, groundedness detection
The generative AI process: Identify → Measure → Mitigate → Operate

Study

Hands-on

Build the environment first, before any reading. Sign in to the Foundry portal, create a Foundry resource and a project, and deploy a small chat model such as gpt-5-mini. Copy the project endpoint from the project overview page.

Set up the labs: cd labs && pip install -r requirements.txt, copy .env.example to .env, fill in the endpoint and deployment name, run az login, then python 01_first_call.py.

Self-check

Which principle covers a model that performs worse for one demographic — and which covers an app with no screen-reader support?
How many Guardrails harm categories are there, and does filtering apply to the prompt, the completion, or both?
What is the difference between a jailbreak and a cross-prompt injection?
Did lab 01 print a response? If not, fix it today — everything from Day 5 depends on it.

Phase I · Concepts

2 h study
30 m quiz

Tue 1 Sep

How generative models actually work

The mechanics beneath every other objective: tokens, embeddings, the Transformer, and attention. Then the decision the exam returns to again and again — fine-tuning versus RAG. Get this right and a dozen scenario questions answer themselves.

Notes 1.2 Models and configuration Quiz 1.1 Responsible AI

Learn

Tokens (≈ ¾ of a word), the billing unit, and the context window = prompt + completion
Embeddings as meaning-vectors, compared with cosine similarity
The Transformer and the attention mechanism — why long-range references resolve
Encoder (understand) versus decoder (generate) blocks
Pre-training → fine-tuning → grounding. Facts need RAG; behaviour needs fine-tuning
Prompt techniques: system message, zero-shot, few-shot, chain of thought

Study

Hands-on

In the Foundry chat playground, send the same prompt at temperature 0, then at temperature 1.5, five times each. Watch determinism collapse.

Then set a system message — "You are a terse geography tutor, never exceed 20 words" — and observe how it persists across turns while the user prompt changes.

Finish with quiz 1.1 to lock in yesterday's responsible AI material while it is fresh.

Self-check

Your bot must answer from a handbook updated weekly. Fine-tune or RAG, and why?
Does max tokens limit the prompt, the completion, or both?
Why should you not tune temperature and top P together?
What does attention let a model do that earlier architectures could not?

Phase I · Concepts

2 h study
30 m quiz

Wed 2 Sep

Choosing a model, deployment types, and inference parameters

The commercial and operational half of Domain 1. Which model, deployed how, tuned with which parameters. These questions are highly patterned — learn the mapping table and they become free marks.

Notes 1.2 Models and configuration Quiz 1.2 Models and configuration

Learn

Models sold by Azure versus models from partners and community — hosting, support, billing
Matching capability to task: chat, reasoning, multimodal, embedding, image generation, SLM
Deployment types: Global Standard, Standard, Data Zone, Provisioned, Batch, Developer
Managed compute versus serverless, and instant access
Parameters: temperature, top P, max tokens, frequency and presence penalties, stop sequences
Quota: TPM, PTU, and the 429 response

Study

Hands-on

Open the model catalog in the portal. Filter by inference task, then by deployment options, then by lifecycle. Open three model cards and compare the Benchmarks and License tabs.

On your existing deployment, find the TPM quota setting and note the current value. Then find the Guardrails configuration attached to it and inspect the four category thresholds.

Take quiz 1.2. Anything under 85% goes on tomorrow's re-read list.

Self-check

Predictable throughput with no throttling — which deployment type?
All processing must stay in the EU — which deployment type?
Four million records overnight, cost matters, latency does not — which type?
What does the model argument in your code actually take?

Phase I · Consolidation

2 h study
30 m quiz

Thu 3 Sep

AI workloads — the scenario-to-service map

The single highest-yield objective in Domain 1. Almost every question is a scenario that must be matched to a workload family and a service. Today is about building the reflex, not the theory.

Notes 1.3 AI workloads Quiz 1.3 AI workloads

Learn

The five families: generative & agentic, text analysis, speech, computer vision, information extraction
Generative versus agentic — the line is whether the system takes actions through tools
Text analysis: key phrases versus entities, sentiment's four labels, extractive versus abstractive
Speech: real-time versus batch, diarisation versus speaker recognition, SSML, Custom Neural Voice
Vision: classification versus detection, OCR, Face capabilities and which are Limited Access
Retired services — LUIS, QnA Maker, Anomaly Detector, Personalizer, Content Moderator, Metrics Advisor

Study

Hands-on

No new code today. Instead, drill the map. Cover the right-hand column of the Rapid scenario → service map table in note 1.3 and work down it aloud. Repeat until you can do all 15 rows without hesitation.

Then take quiz 1.3, and re-take quiz 1.2 if yesterday's score was below 85%.

Domain 1 is now complete. Before you close the laptop, write out the five workload families from memory and name the primary service for each. If any is shaky, that objective gets ten minutes tomorrow morning.

Self-check

Count cars in a photo — classification or detection?
Which sentiment label do people forget exists?
A system that reads an email and issues a refund via an API — generative or agentic?
Name four retired services that appear as exam distractors.

Phase II · Foundry

2 h study
30 m lab

Fri 4 Sep

Prompt engineering and the Foundry portal

Domain 2 begins — the 55–60% block. Today is the naming table and the portal. Learn the 2026 renames cold: the exam uses current names and offers retired ones as distractors.

Notes 2.1 GenAI apps and agents Lab 01 First call

Learn

The rename table: Microsoft Foundry, Foundry Tools, Responses API, Conversations/Items/Responses
Resource hierarchy: subscription → resource group → Foundry resource → project → deployments
Foundry projects versus hub-based projects, and when you still need classic
System message anatomy: role → task → tone → format → grounding → guardrails
Techniques: few-shot, chain of thought, delimiters, output format specification
The deployment name trap — it is what your code passes as model=

Study

Hands-on

Re-run python 01_first_call.py and read every line. Then modify it: change the system message to force JSON output, and confirm temperature 0 gives you byte-identical results across runs.

In the portal, walk the full deployment path once more from scratch: catalog → filter → deploy → name it something deliberately different from the model name → test in the playground → call it from code using that deployment name. That one exercise kills the most common exam trap.

Self-check

What replaced Threads, Messages, Runs, and Assistants?
In the Responses API, which parameter carries the system message?
Why is a system message not a security boundary?
What is the format of a Foundry project endpoint?

Phase II · Foundry

4–6 h deep work
1 h quiz

Sat 5 Sep

The Foundry SDK, conversations, and agents — the heaviest day

Your first long session, on the most heavily weighted material in the exam. By tonight you should have written a chat client, built an agent, versioned it, and understood exactly what makes a call stateful. Do not rush this one.

Notes 2.1 GenAI apps and agents Quiz 2.1 GenAI apps and agents Lab 02 Chat client Lab 03 Prompt agent

Learn

azure-ai-projects 2.x, AIProjectClient, get_openai_client(), responses.create()
DefaultAzureCredential — Entra auth, no keys in code
Stateless versus stateful: conversation=<id> is the whole difference
Prompt agents versus hosted agents versus the ephemeral Responses API pattern
Knowledge tools versus action tools; what a toolbox is and why it is MCP-compatible
The agent lifecycle: create, test, trace, evaluate, optimize, publish, monitor
Agent identity, RBAC, versioning, and publishing to Teams and Microsoft 365 Copilot

Study

Hands-on

Lab 02 — the key lesson of the whole exam. Run python 02_chat_client.py. Part A asks a follow-up with no conversation id and the model has no idea what "its" refers to. Part B passes conversation=conversation.id and it resolves. Understand precisely why.

Lab 03 — agents. Run python 03_prompt_agent.py. It creates an agent, chats across three turns, then creates a second version with different instructions. Watch the refusal behaviour on the out-of-scope question — that is your system message working as a guardrail.

Then go to the portal. Build the same agent by hand in Build → Agents. Add a tool. Test it in the agents playground. Open the trace and read every model call and tool invocation. Create a third version and roll back to the first.

Finish with quiz 2.1 — 53 questions, the largest bank in the kit.

Self-check

What single argument makes a Responses API call stateful?
Prompt agent, hosted agent, or Responses API direct — for a LangGraph app you already wrote?
When you bind with get_openai_client(agent_name=...), what can you stop passing?
Which cost applies to hosted agents but not prompt agents?

Phase II · Foundry

4–6 h deep work
1 h quiz

Sun 6 Sep

Text analysis and speech, end to end

Second long session. Two routes to text analysis and when each wins, then the full Speech surface including SSML — which is the answer to more questions than its size suggests. Build the complete voice loop today.

Notes 2.2 Text and speech Quiz 2.2 Text and speech Lab 04 Text analysis Lab 05 Speech

Learn

The central judgement: specialist Foundry Tool versus general multimodal model
Azure AI Language: sentiment with scores, opinion mining, key phrases, entities, PII, summarisation, CLU
Multimodal audio input — content as a list of typed parts
Speech to text: real-time, batch, fast transcription, diarisation, custom speech
Text to speech: neural voices, SSML, Custom Neural Voice and its Limited Access gate
Speech translation versus Azure AI Translator
The three-stage voice pipeline, and what a multimodal model does and does not replace

Study

Hands-on

Lab 04. Run both routes over the same three reviews. Notice that route A gives you a bespoke JSON shape and route B gives you scored, categorised structure. Read the mixed-sentiment result on review one.

Lab 05. Create a Speech resource if you have not already, then run it. Compare out_plain.wav against out_ssml.wav — the SSML version pauses, stresses "four", and slows the final line. Then speak into the microphone and let Part E complete the full loop: your voice → text → model → speech.

Extend it yourself. Edit the SSML to add <say-as interpret-as='telephone'> and a second <voice>. Making SSML muscle memory today is worth several marks.

Finish with quiz 2.2.

Self-check

Which two credentials does the Speech SDK always need?
Real-time transcription with speaker labels — multimodal model or Azure AI Speech?
Which SSML element controls speaking rate, and which inserts a pause?
Translating a 40-page PDF — Speech or Translator?

Phase II · Consolidation

2 h revision
30 m quiz

Mon 7 Sep

Consolidation — Domain 1 complete, Domain 2 half done

A deliberately lighter day after two heavy weekend sessions. No new material. Today is about finding the gaps while there is still time to close them.

Notes Exam cheat sheet Quiz All quizzes

Learn

Re-read the cheat sheet sections for Domain 1 and for objectives 2.1 and 2.2
Rebuild your mental model of the rename table without looking
Revisit anything from quizzes 1.1 through 2.2 that scored below 85%

Study

Hands-on

Re-take every quiz from 1.1 to 2.2 on Hard tier only. Hard questions are the distractor traps, and they surface misconceptions that easy questions hide.

Keep a written list of every question you get wrong. That list — not the notes — is what you revise on Days 13 and 14.

Take the official practice assessment for the first time. Expect a mediocre score; you have not covered vision or information extraction yet. The point is to calibrate the question style, not to pass.

Self-check

Can you recite the six responsible AI principles and one trigger phrase for each?
Can you write the six-line SDK boilerplate from memory?
Which deployment type for each of: cheapest, no throttling, EU-only, bulk overnight?
What is on your wrong-answer list, and which note covers each item?

Phase III · Multimodal

2 h study
30 m lab

Tue 8 Sep

Visual input — images in prompts

Back to new material. The image content-part pattern is identical in shape to the audio pattern from Day 7, so this should feel familiar. The exam's real interest is in when a multimodal model is the wrong tool.

Notes 2.3 Vision and image generation Lab 06 Vision

Learn

Passing images by URL and as base64 data URLs
Multiple input_image parts in one message — how comparison works
Structured extraction from an image with a JSON schema at temperature 0
Azure AI Vision: VisualFeatures — CAPTION, OBJECTS, READ, TAGS
Classification versus detection; Custom Vision for your own classes
Azure AI Face and which capabilities are Limited Access

Study

Hands-on

Run python 06_vision.py Parts A to D. Part C sends two images in one message — read the comparison output carefully and note that both images were considered together.

Drop your own photograph in as sample.jpg and run Part D. Try a receipt, a whiteboard, or a handwritten note, and see where the model is strong and where it guesses.

Then write the failure case yourself: ask the model for exact pixel coordinates of an object. Note how confident and how wrong it is. That is precisely why Azure AI Vision object detection exists.

Self-check

How do you send two images for comparison in one prompt?
Which VisualFeatures flag returns bounding boxes, and which returns handwritten text?
Which Face capabilities need Microsoft approval?
When is a multimodal model the wrong tool for an image task?

Phase III · Multimodal

2 h study
30 m quiz

Wed 9 Sep

Image generation and responsible AI for synthetic media

The shorter half of objective 2.3, and the place where Domain 1 and Domain 2 meet: provenance metadata and disclosure are the transparency principle applied to generated images.

Notes 2.3 Vision and image generation Quiz 2.3 Vision and image generation Lab 06 Vision

Learn

images.generate() and its parameters: size, quality, n, style, response format
The prompt recipe: subject + style + composition + constraints
Generation versus inpainting versus variations
Guardrails on image models; why requests for real people are refused
C2PA provenance metadata and the transparency obligation
The production shape: specialist services give precision, the model gives judgement

Study

Hands-on

Deploy an image model if you have not already, then run Part E of lab 06. Generate the same subject three times: once at quality="low", once at "medium", once at "high". Compare the output and note the latency difference.

Then test the guardrail: ask for an image of a named living public figure. Read the refusal. That is Domain 1 and Domain 2 meeting in a single API call.

Take quiz 2.3.

Self-check

Which two parameters most directly drive image-generation cost?
What is C2PA metadata, and which responsible AI principle does it serve?
Why would you request base64 rather than a URL?
Your own product defects that no prebuilt model knows — which service?

Phase III · Extraction

2 h study
30 m lab

Thu 10 Sep

Content Understanding — one service, four modalities

The last new objective, and a rich source of exam questions because it has several precise, testable rules. Three of them decide most questions: extract is documents-only, analysis is asynchronous, and confidence enables straight-through processing.

Notes 2.4 Information extraction Lab 07 Content Understanding

Learn

The analyzer — content extraction settings, field schema, model deployments
Prebuilt versus custom analyzers, and labelled training examples
The three methods: extract (documents only), classify, generate
Confidence scores (0–1) and grounding; estimateFieldSourceAndConfidence
enableSegment for splitting documents and video
The asynchronous submit → Operation-Location → poll pattern
Markdown output for RAG versus JSON for automation; GA version 2025-11-01

Study

Hands-on

Run python 07_content_understanding.py. Read the printed custom analyzer schema line by line and identify which fields use extract, classify, and generate — and why each was chosen.

In the portal, open the Content Understanding playground. Upload a real invoice or receipt of your own and run the prebuilt analyzer. Inspect the confidence score on every field and find the weakest one.

Then upload something that is not a document — a short audio clip or a photo — and observe which extraction methods are available and which are not.

Self-check

Which extraction method is documents-only, and what do you use for video instead?
Is the Analyze API synchronous or asynchronous, and what header do you poll?
What is the commercial purpose of confidence scores?
What does grounding give you that a bare value does not?

Phase IV · Mastery

2 h drill
1 h practice exam

Fri 11 Sep

Syllabus complete — first full assessment

Every objective is now covered. Today you find out where you actually stand, under something close to exam conditions.

Quiz 2.4 Information extraction Notes Exam cheat sheet

Learn

Take quiz 2.4 to close out the syllabus
Then a full pass of the cheat sheet, top to bottom, without stopping
Update your wrong-answer list from every quiz taken so far

Study

Hands-on

Take the official practice assessment under exam conditions: no notes, no pausing, one sitting, phone in another room.

Score it honestly. Below 700 means you have real gaps and Days 13 and 14 must be targeted, not general. Above 850 means you can spend the weekend on polish rather than panic.

Then run the exam sandbox so the interface holds no surprises on Monday. If you are testing online, run the Pearson VUE system check on the exact machine you will use.

Self-check

What was your practice score, and which domain dragged it down?
List your five weakest topics. Those are the whole of tomorrow's plan.
Is your exam booking confirmed, with the correct code and time zone?

Phase IV · Mastery

4–6 h targeted drill
1 h practice exam

Sat 12 Sep

Targeted repair and full-bank drill

The most valuable day in the plan, if you use it correctly. Do not re-read everything. Attack the five weakest topics from yesterday, then prove the repair held.

Quiz All quizzes Notes Exam cheat sheet Lab All labs

Learn

Re-read only the notes covering your five weakest topics
Re-take every quiz on Hard tier — all seven, 100+ questions
Rebuild your weakest lab from an empty file, without copying

Study

Hands-on

Morning — repair. For each of your five weakest topics: read the note section, do the self-check at the bottom, then re-take that quiz on all tiers. Move on only when you are above 90%.

Afternoon — rebuild. Delete your weakest lab file and write it again from scratch, from memory, in an empty editor. Most people should pick lab 02 or lab 03. If you can write the AIProjectClient → get_openai_client → responses.create chain with a conversation id from memory, the SDK questions become free marks.

Evening — assess. Second full practice assessment. Compare against Friday. The delta tells you whether today worked.

Self-check

Did your five weak topics all clear 90% on re-test?
Could you write the SDK boilerplate from an empty file?
Is your practice score above 800 yet?
Anything still failing goes on one index card for tomorrow. Nothing else.

Phase IV · Mastery

3–4 h light revision
then rest

Sun 13 Sep

Cram, calm, and stop

The day before. The goal is calm confidence, not fresh data. Everything today is revision of material you already know, and it ends early.

Notes Exam cheat sheet Quiz All quizzes

Learn

One unhurried pass of the cheat sheet, end to end
The twelve final traps at the bottom of the cheat sheet — read them three times
Your index card of remaining weak points
The rename table, one last time

Study

Hands-on

Morning. Read the cheat sheet straight through, once, without stopping to look things up. Then re-take the two quizzes you have historically scored worst on — all tiers, one attempt each. Do not chase a perfect score.

Afternoon. Read your index card. Read the twelve final traps. Then stop. Close the laptop by mid-afternoon.

Evening. Confirm the booking, the start time, the time zone, and your ID. If testing online, clear the room and run the system check one final time. Lay out your ID. Sleep at your normal hour — do not attempt an unusually early night your body will fight.

Do not take a full practice exam today. A poor score rattles you; a great score makes you careless.

Self-check

Can you recite the six principles, the four Guardrails categories, and the three extraction methods?
Do you know which of extract, classify, generate is documents-only?
Is your ID ready and your booking confirmed for tomorrow?
Have you actually stopped studying?

§ 04 · Library

Everything you need, in one place.

Official sources first. Microsoft renamed most of this product surface in 2026, so older blogs, videos, and courses are actively misleading — check the publication date on anything you read outside this list.

Official exam pages

Practice & rehearsal

Microsoft Foundry documentation

DocsWhat is Microsoft Foundry
DocsFoundry Agent Service
DocsFoundry Models & the catalog
DocsFoundry Tools — the current service list
DocsAzure AI Content Understanding
QuickFoundry SDK quickstart

This kit

§ 05 · Traps

The twelve most-missed facts.

These are the distractor traps this exam leans on hardest. If you know nothing else on the morning of the exam, know these.

Extract is documents-only. Images, audio, and video use classify or generate.
Content Understanding is asynchronous. Submit, read Operation-Location, poll.
Threads, Messages, Runs, Assistants are retired vocabulary. It is Conversations, Items, Responses, Agent versions.
conversation=<id> is what makes a call stateful. Omit it and there is no memory.
instructions is the system message in the Responses API — there is no system-role messages array.
model= takes the deployment name, which need not match the model name.
Sentiment has four labels. Positive, negative, neutral, and mixed.
Guardrails filter the prompt and the completion — four categories, four severity levels each.
max_tokens caps the completion, not the prompt. The context window covers both.
Fine-tuning does not teach current facts. Facts need RAG; fine-tuning shapes behaviour.
Face identification and Custom Neural Voice are Limited Access and need Microsoft approval.
LUIS, QnA Maker, Anomaly Detector, Personalizer, Content Moderator and Metrics Advisor are retired. Never the right answer.

§ 06 · Playbook

Exam day, hour by hour.

A calm ritual beats last-minute cramming. Follow the timings below to arrive at the desk in the right state — knowledgeable, rested, unhurried.

T – 12 hours

The night before

Skim the cheat sheet once. Do not attempt new material
Confirm the booking, the exam code, ID requirements, and the time zone
Online: run the Pearson VUE system check on the exact device you will use
Lay out your ID. Charge everything
Sleep at your normal time. Do not sabotage yourself with an early night your body will fight

T – 60 minutes

The final hour

Light meal, water, caffeine only if you take it daily
Online proctoring: clear the room, phone away, single monitor, disconnect other displays
Test centre: leave 30 minutes early, and bring two forms of ID
Read the twelve traps in § 05 one last time, then set it aside — no notes in the exam

During the exam

At the desk

Read each question twice. Note qualifiers: always, never, only, most, least, best
Eliminate obviously wrong answers first — retired service names are usually free eliminations
On code questions, look for the conversation id and the instructions parameter before anything else
Mark for review anything you are less than 90% sure of, and return with fresh eyes
Trust your first instinct unless you find concrete evidence to change it
Answer every question. There is no negative marking

One thing not to do: take a full-length practice exam on the morning of the exam. A poor score will rattle you; a great score will make you careless. By exam morning the goal is calm confidence, not fresh data.