AI-901 Cheat Sheet: Last-Minute Revision for Azure AI Fundamentals

Everything you need for AI-901 on one page. Read it the day before the exam, then take the practice test one more time.

The exam

Identify AI concepts and capabilities: 40–45%
Implement AI solutions by using Microsoft Foundry: 55–60%
Pass mark: 700 (scaled). Expect scenarios, matching and short Python snippets.

Responsible AI: the six principles

Principle	Think…
Fairness	Equal outcomes for similar people, no bias
Reliability and safety	Works consistently, tested on edge cases, monitored
Privacy and security	Protect personal data, consent, resist attacks
Inclusiveness	Usable by everyone, including people with disabilities
Transparency	People know it's AI and understand its limits and decisions
Accountability	Humans and governance are responsible

Generative AI safety in Foundry: content filters / guardrails, prompt shields, system prompts, grounding, evaluations. Process: identify → measure → mitigate → operate.

How generative AI works

Text → tokens (about 4 characters or ¾ of a word each) → embeddings (meaning as vectors) → transformer with attention → predicts the next token.
Context window = max tokens per request. Pricing is usually per input and output token.
Hallucination = confident but wrong output. Fix with grounding / RAG.
Training: pre-training → fine-tuning → alignment.
Model types: LLM, SLM (Phi), reasoning, multimodal, image generation (diffusion), embedding.

Choosing and deploying models

Choose by capability, then quality, cost, latency, context window, region.
Global Standard: pay per token, any region, the usual default.
Data Zone: processing stays in a geography (EU or US).
Regional Standard: processing stays in one region.
Provisioned (PTU): reserved throughput, predictable latency.
Batch: asynchronous, lower cost, big jobs.
Temperature / top_p: randomness (low = focused). Max tokens: length. Reasoning models may use reasoning effort instead of temperature.

Workloads and keywords

Workload	Keywords
Generative AI	draft, write, chat, create content
Agentic AI	take actions, call APIs, multi-step tasks
Text analysis	sentiment, key phrases, entities, PII, summarise, language
Speech	transcribe, captions (recognition); read aloud (synthesis); translate
Computer vision	classify (whole image), detect (boxes), segment (pixels), OCR
Image generation	create an image from a description
Information extraction	invoices, forms, receipts, fields from audio or video

Prompts

System prompt: role, scope, rules, tone, format, for the whole conversation.
User prompt: the specific request.
Techniques: be specific, add context, delimiters, few-shot examples, ask for a format (JSON), ground in data, say what to do when unsure.

Foundry essentials

Hierarchy: subscription → Foundry resource → project → deployments, agents, tools.
Project endpoint: https://<resource>.services.ai.azure.com/api/projects/<project>.
Portal flow: model catalog → deploy (name + type) → playground → view code.

Code patterns to recognise

Chat with a model

project = AIProjectClient(endpoint=endpoint, credential=DefaultAzureCredential())
openai = project.get_openai_client()
r = openai.responses.create(model="<deployment>", instructions="...", input="...")
print(r.output_text)

Chat with an agent

project.agents.create_version(agent_name="my-agent",
    definition=PromptAgentDefinition(model="<deployment>", instructions="..."))
openai = project.get_openai_client(agent_name="my-agent")
conv = openai.conversations.create()
r = openai.responses.create(conversation=conv.id, input="...")

Text analysis: TextAnalyticsClient(endpoint, AzureKeyCredential(key)) → analyze_sentiment, extract_key_phrases, recognize_entities, recognize_pii_entities, detect_language, begin_abstract_summary.

Speech: SpeechConfig(subscription, region) → SpeechRecognizer(...).recognize_once_async().get() / SpeechSynthesizer(...).speak_text_async(text).get(). SSML controls voice, speed and pauses.

Spoken prompt: content part {"type": "input_audio", "input_audio": {"data": <base64>, "format": "wav"}}.

Image in a prompt: content parts input_text + input_image (URL or base64 data URL).

Image generation: openai.images.generate(model="gpt-image-1", prompt="...").

Content Understanding: client.begin_analyze(analyzer_id="prebuilt-invoice", inputs=[...]) → poller → result() → fields with confidence. REST: 202 + Operation-Location, then poll.

Agents

Agent = model + instructions + tools (+ conversations).
Tools: file search (your documents), code interpreter (run Python, analyse data), web search (current info), functions / OpenAPI (your APIs), MCP.
Responsible agents: least privilege, human approval for big actions, guardrails, evaluation.

Commonly confused pairs

Speech recognition (audio → text) vs speech synthesis (text → audio)
Image classification (what) vs object detection (what and where)
Extractive (reuse sentences) vs abstractive (new sentences) summaries
Fairness (outcomes) vs inclusiveness (access)
Transparency (explain) vs accountability (who's responsible)
Provisioned (reserved capacity) vs Standard (pay per token)
Embedding model (vectors for search) vs chat model (generates text)

Exam-day tips

Read the last line first. Know what's being asked before reading a long scenario.
Find the constraint: cheapest, no code, data residency, real time, guaranteed throughput.
Eliminate obviously wrong answers, then decide between the last two.
Don't overthink fundamentals questions. The simplest service that meets the need is usually right.
Flag and move on if you're stuck; come back at the end.
For online exams, check your room, ID and system (run Microsoft's system test) well before the start time.

Good luck! Once you pass, take a look at the AI courses roadmap for what to learn next.