← Back to Readiness Assessment

AI-900 Decision Framework

A decision and keyword atlas designed to support the AI-900 Readiness Assessment

Use this framework when unsure during the assessment. Focus on Keywords, Ask Yourself, and Traps.

✨ Generative AI

Azure OpenAI

Chat
Text generation
Summarization
Code generation

Keywords: generate, summarize, write, chat, respond naturally

Ask yourself:
Is the system creating new content?

❌ Trap: If it's extracting facts → Not Generative AI

Embeddings

Numerical representations of meaning
Semantic similarity
Clustering & vector search

Keywords: similarity, semantic, vector, meaning comparison

Ask yourself:
Are we comparing meaning, not keywords?

❌ Trap: Embeddings ≠ Tokens. Embeddings capture meaning; tokens are input units.

Tokens & Tokenization

How a model reads text
Text → Tokens → IDs → Processing
Determines input limits and cost

Keywords: token limit, input size, length, context window

Ask yourself:
Is the question about limits, length, or input size?

Prompt Engineering

Writing better instructions for generative models
AI‑900 scope: concept only (no syntax, no parameters)

Keywords: prompt, instruction, system message, few‑shot

Ask yourself:
Is it about improving how you ask the model?

❌ Trap: Prompt Engineering applies to Generative AI / Azure OpenAI only.
❌ Does NOT apply to AutoML, Vision, or Language.

🧠 Transformers

Evaluate relationships between tokens using attention.

Keywords: attention, context, relationships

Ask yourself:
Is the model determining how words relate?

🎯 Attention Layers

Identify importance of words relative to each other.

❌ Not token splitting
❌ Not prediction alone
✅ It's relationship evaluation

🧩 Multi‑Modal Models

Combine multiple inputs: text, image, audio.

Keywords: multi-input, cross-modal, combined data

Ask yourself:
Are multiple data types used together?

🔍 Search & Knowledge Mining

Azure AI Search

Search documents
Index data
Semantic search

Keywords: search, index, retrieve, documents, knowledge base

Ask yourself:
Are users searching existing content?

❌ Trap: Search does not generate answers

Knowledge Mining (Hybrid)

Search + AI enrichment
Extract insights from large document collections
Combines Azure AI Search with Language / Vision

Keywords: thousands of documents, discovery, insights, enrichment

Ask yourself:
Do we need to discover patterns across many documents?

📥 Indexer vs Index

Indexer — Pulls data from sources, converts to JSON, feeds into index.

Index — Stores searchable content.

Keywords: indexer → ingestion, index → storage/search

Ask yourself:
Are we ingesting data or querying it?

❌ Indexer ≠ search engine
❌ Indexer ≠ replacement for index

🧠 Skillsets

Apply AI enrichment during indexing: OCR, entity extraction, key phrase detection.

Keywords: enrich, AI pipeline, skillset

Ask yourself:
Are we adding AI processing while indexing data?

🔄 Pull Model

Indexers pull data from sources — this is the default ingestion model.

Keywords: pull, retrieve data, ingestion pipeline

❌ Not push-based
✅ Pull-based ingestion model

🔤 Query Syntax

Default: Simple query syntax.

❌ Lucene is NOT default
✅ Only used when explicitly specified

🤖 Conversational AI

Azure Bot Service

Chatbots
FAQ bots

Keywords: chatbot, conversation, virtual agent

Ask yourself:
Is the user talking to a system conversationally?

🛡️ Content Safety

Azure AI Content Safety

Detect harmful content
Text and image moderation

Keywords: harmful, abuse, hate, filtering

Ask yourself:
Is the goal to block or flag unsafe content?

❌ Trap: Sentiment ≠ Safety

🧩 Azure AI Personalizer

Real‑Time Personalization

Chooses the best action per user in real time
Learns from user feedback
Optimizes decisions over time (continuously improves)

Keywords: personalize, rank options, next best action, user behavior, contextual decisions

Ask yourself:
Is the system choosing one best option per user, not just predicting preferences?

❌ Trap: Personalizer ≠ traditional recommendation model
❌ Personalizer ≠ static rules
✅ Personalizer = adaptive decision making

Typical Use Cases

Personalized product banners
News or content ranking
"Next best action" decisions
Adaptive UI choices

👁️ Azure AI Vision

Image Classification

What is in the image (label)
Single category output

Keywords: classify, label, categorize image

Ask yourself:
Do we only need to know what it is?

Object Detection

What + Where (bounding boxes)
Can detect condition/state if visually inferable
Examples: helmet worn, damage, open/closed

Keywords: locate, bounding box, identify defects, position

Ask yourself:
Do we need to know what AND where?

❌ Trap: OCR detects text only, not objects

Face Detection

Detect faces + location
Age, emotion, accessories

Keywords: face, detect person, identify

👤 Azure Face Service Restriction

Requirement: Must submit an intake/approval form (Responsible AI compliance).

Keywords: approval, restricted, compliance

❌ Not pricing tier
❌ Not subscription level
✅ Governance requirement

Task	Output
Image Classification	What is in the image (label)
Object Detection	What + where (bounding box)
Face Detection	Faces + location

📄 OCR vs Document Intelligence

OCR

Extract raw text from images
Printed and handwritten
Photos, screenshots, scanned pages

Keywords: read text, extract text, screenshot, scanned

Ask yourself:
Is the goal to get raw text out of an image?

❌ Trap: OCR does NOT understand forms, tables, or meaning

Document Intelligence

Extract structured information from documents
Invoices, receipts, forms, contracts
Key‑value pairs, tables, fields

Keywords: invoice, receipt, form fields, key‑value pairs, totals, structured

Ask yourself:
Do we need structured fields (not just raw text)?

🚨 Exam trap: "Extract invoice totals and dates" → Document Intelligence, NOT OCR

📊 Document Intelligence Model Types

Two supported models: Prebuilt and Custom.

Keywords: prebuilt, custom training

Ask yourself:
Do we use an existing template or train our own?

Service	Input	Output
OCR	Image with text	Raw unstructured text
Document Intelligence	Forms, invoices, receipts	Structured key‑value pairs, tables, fields

💬 Azure AI Language

Named Entity Recognition (NER)

Extracts named things from text
People, organizations, locations
Dates, money, quantities

Keywords: extract, entity, person, organization, date, structured facts

Ask yourself:
Are we pulling structured facts from text?

❌ Trap: NER extracts facts, it does NOT generate text

Sentiment Analysis

Positive / Negative / Neutral
Opinion mining

Keywords: sentiment, opinion, positive, negative, feedback

Ask yourself:
Are we measuring how people feel?

Key Phrases & Text Classification

Extract important phrases from text
Classify text into categories

Keywords: key phrase, classify text, topic, category

Ask yourself:
Do we need to identify important topics or categorize text?

🧩 Language Detection – NaN Result

NaN appears when the model cannot confidently determine a language due to ambiguity.

Keywords: ambiguous text, unclear language, low confidence

Ask yourself:
Is the input unclear or mixed in a way that prevents confident classification?

❌ Mixed languages ≠ NaN
❌ Invalid score ≠ NaN
✅ NaN = ambiguity

🔗 Entity in Conversational AI

An entity is a specific item or value referenced in an utterance.

Example: "Book a flight to Paris" → Entity = Paris

Keywords: entity, value, parameter, extracted data

Ask yourself:
Are we extracting a specific value from user input?

❌ None Intent

Fallback intent for unrecognized or unmatched input.

Keywords: fallback, unknown input, no match

Ask yourself:
Does this utterance NOT match any defined intent?

❌ Not for training
❌ Not a default prebuilt intent
✅ It's a catch‑all

🎙️ Azure AI Speech

What it CAN do

✅ Speech‑to‑Text (transcription)
✅ Text‑to‑Speech (voice output)
✅ Speech translation
✅ Speaker recognition

Keywords: transcribe, dictate, voice, audio, speaker

Ask yourself:
Is the input or output audio?

What it CANNOT do

❌ Understand image context
❌ Extract entities from text (→ Language)
❌ Perform sentiment analysis (→ Language)
❌ Replace a chatbot alone

❌ Trap: Real‑time transcription ≠ Natural Language Understanding.
Speech = audio processing · Language = text understanding

🔊 Speech Recognition Models

Uses two models: Acoustic (sound → phonetics) and Language (word structure → sentences).

Keywords: audio, phonetics, structure

🔀 Speech vs Language vs Bots – Separator

Service	Handles	Input/Output
Azure AI Speech	Transcription, TTS, translation, speaker ID	Audio
Azure AI Language	NER, sentiment, key phrases, classification	Text
Conversational AI (Bots)	Dialogs, conversation flow, user interaction	Conversation management

🚨 Speech ≠ Language ≠ Chatbot. These are NOT interchangeable.

🌐 Text Translation

Azure AI Translator

Converts text from one language to another
Preserves meaning across languages

Keywords: translate, language conversion, multilingual, cross‑language text

Ask yourself:
Is the goal to convert text between languages, not create or analyze content?

❌ Translation ≠ text generation
❌ Translation ≠ speech‑to‑text (audio)
❌ Translation does NOT summarize or analyze sentiment

Typical Use Cases

Multilingual websites
Customer support translation
Global content distribution

🔗 Entity Linking

Extension of NER

NER identifies: "Microsoft" → Organization
Entity Linking connects it to the known entity in a database
Disambiguates: "Apple" the company vs fruit

Keywords: disambiguate, link entity, knowledge graph, resolve

Ask yourself:
Do we need to connect a mention to a known entity, not just label it?

Why It Matters

Improves search accuracy
Helps build knowledge graphs
Avoids duplicates across mentions

✅ Conceptual understanding only
❌ No implementation details required for AI‑900

📊 Core ML Patterns

Classification

Predict a label / category
Yes/No, type, class

Keywords: classify, predict category, label, yes/no, type

Ask yourself:
Is the output a category or label?

Regression

Predict a number
Price, amount, score, quantity

Keywords: predict number, price, amount, forecast, continuous

Ask yourself:
Is the output a numerical value?

Clustering

Group similar items
No labels (unsupervised)

Keywords: group, similarity, unsupervised, segment

Ask yourself:
Are there no predefined categories?

Anomaly Detection

Detect unusual behavior
Fraud, spikes, failures

Keywords: unusual, abnormal, spike, deviation, outlier, unexpected

Ask yourself:
Is the goal to find what looks different from normal?

Recommendation

Product suggestions
User preferences

Keywords: recommend, suggest, personalize

Ask yourself:
Is the system predicting what someone might like?

📉 Regression Assumptions

Features should be independent to avoid multicollinearity.

Keywords: independent variables, multicollinearity

❌ Dependent features → bad model

🎯 Multiclass Classification

Predicting one label from multiple categories. Example: movie genre.

Binary → Yes/No
Regression → Number
Multiclass → Many categories

🧠 Deep Learning

Advanced ML using multi-layer neural networks.

Keywords: neural network, deep, layers, brain-inspired

❌ Not a specific task
✅ It's a model approach

🖼️ CNN (Computer Vision)

Image classification / pattern recognition.

Keywords: image, vision, convolution

Technique	Output	Use when
Classification	Label / category	Yes/No, type, class
Regression	Number	Price, amount, score
Clustering	Groups	No labels, similarity

🚨 Trap: Same dataset, different objective → different technique

🧪 Azure Machine Learning

Automated ML (AutoML)

Automatically tries multiple algorithms
Optimizes model selection & parameters
Less control, faster experimentation

Keywords: best model, automatic, optimize, quick experiment

Ask yourself:
Do you want the best model quickly without designing pipelines?

ML Designer

Visual drag‑and‑drop pipeline
Explicit control over each step
Full transparency

Keywords: visual, pipeline, drag‑and‑drop, control, steps

Ask yourself:
Do you want to define how the model is built?

Model Versioning

Model registration & versioning
Experiment tracking
Deployment history

Keywords: track versions, model registry, experiment, deploy history

Ask yourself:
Do you need to track multiple versions of a model?

❌ Trap: Azure AI Vision/Language do NOT manage model lifecycle.
❌ Azure OpenAI does NOT version your custom ML models.

Inference

Using a trained model to make predictions
The "production" phase of ML

Keywords: predict, inference, deploy, real‑time, batch scoring

Ask yourself:
Is the model already trained and now being used?

Feature	AutoML	ML Designer
Algorithm selection	Automatic	Manual
Speed	Faster	Slower
Control	Low	High
Visual pipeline	No	Yes

Phase	What happens
Training	Creating a model from data
Versioning	Tracking & comparing multiple model versions
Inference	Using a trained model to make predictions

📏 Evaluation Metrics (Classification)

These are evaluation metrics, not algorithms. They measure how good a classification model is, derived from the confusion matrix (TP, FP, FN, TN).

🔥 Loss Function (Training)

Measures how wrong the model is
Used during training to improve the model
Lower loss = better model

Examples: Regression → MSE · Classification → Cross Entropy

Exam cue: "Error used to TRAIN the model"

❌ Trap: Loss is a training metric, NOT an evaluation metric.
❌ Loss ≠ Accuracy / Precision / Recall / F1

Accuracy

How often the model is correct overall
"Out of all predictions, how many did I get right?"

When useful: Balanced classes, errors have similar cost

Exam cue: "Balanced dataset, generic performance"

❌ Trap: 99% accuracy can be useless on imbalanced data — the model may miss all important cases.

Precision

How correct your positive predictions are
"When the model says YES, how often is it right?"
Formula: TP / (TP + FP)

When important: False positives are costly (spam filters, fraud alerts)

Exam cue: "Avoid false alarms" · "Minimize false positives"

❌ Trap: High precision does NOT mean the model finds all positives.

Recall (Sensitivity)

How many actual positives the model finds
"Out of all real YES cases, how many did we catch?"
Formula: TP / (TP + FN)

When important: False negatives are dangerous (disease, safety, fraud)

Exam cue: "Detect all cases" · "Do not miss positives"

❌ Trap: High recall often increases false positives.

F1 Score

Balance between precision and recall
Combines both into one score
Formula: 2 × (Precision × Recall) / (Precision + Recall)

When best: Imbalanced datasets, both FP and FN matter

Exam cue: "Balance between detecting cases and avoiding false alarms"

❌ Trap: F1 does NOT replace precision or recall — it summarizes them.

Focus	You care about	Exam wording
Precision	Fewer false positives	"Avoid false alarms"
Recall	Fewer false negatives	"Don't miss cases"
F1	Balance of both	"Overall classification quality"
Accuracy	Overall correctness	"Balanced dataset, generic"

🚨 These metrics apply to classification only.
❌ Regression uses MAE, MSE, RMSE.
❌ Clustering uses silhouette, cohesion.

Concept	Used for	Question it answers
Loss	Training	How wrong am I?
Accuracy	Evaluation	How often am I right?
Precision	Evaluation	Can I trust positives?
Recall	Evaluation	Did I miss any positives?
F1 Score	Evaluation	Balance precision + recall

🔥 Ultra‑fast memory trick:

Loss → training → "how wrong"
Accuracy → "overall correct"
Precision → "don't cry wolf"
Recall → "don't miss cases"
F1 → "balance tradeoff"

🧪 Real Example: Disease Detection
Precision	Don't say disease if healthy
Recall	Don't miss a sick patient

👉 In disease detection, Recall is MORE important — missing a sick patient is worse than a false alarm.

Item	What it is
Accuracy / Precision / Recall / F1	Evaluation metrics
Loss (MSE, Cross Entropy)	Training metric
Logistic Regression / Decision Tree / Random Forest / Neural Network	Algorithms

🚨 Loss = training metric. Metrics (Accuracy, Precision, Recall, F1) = evaluation.
🚨 Metrics do NOT train models. Metrics do NOT choose algorithms.

✅ Binary Classification Metric Shortcut:

"Avoid false alarms" → Precision
"Don't miss cases" → Recall
"Balanced classification" → F1 Score
"Overall correct, balanced data" → Accuracy

📐 Evaluation Metrics (Regression)

These metrics evaluate regression models where the output is a number (price, amount, forecast). They are NOT used for classification.

MAE – Mean Absolute Error

Average absolute difference between predicted and actual
Same unit as target value
Treats all errors equally

When preferred: Simple, interpretable metric. Outliers should NOT dominate.

Exam cue: "Average error" · "Simple measure" · "Error in same unit"

MSE – Mean Squared Error

Average of squared differences
Large mistakes punished much more
Unit is squared (less intuitive)

When preferred: Large errors must be heavily penalized

Exam cue: "Penalize large errors" · "Squared error" · "Sensitive to outliers"

❌ Trap: High MSE may look "bad" even when most predictions are reasonable.

RMSE – Root Mean Squared Error

Square root of MSE
Penalizes large errors (like MSE)
Same unit as target value (more interpretable)

When preferred: Penalize large errors AND keep original units

Exam cue: "Error in original units" · "Balance interpretability and severity"

Metric	Penalizes large errors	Same unit as output	Sensitivity to outliers
MAE	❌ No	✅ Yes	Low
MSE	✅ Yes (strong)	❌ No (squared)	High
RMSE	✅ Yes	✅ Yes	High

🚨 These metrics apply to regression only.
❌ Classification uses Accuracy, Precision, Recall, F1.
❌ Clustering uses silhouette, cohesion.

Exam shortcut:

"Predict a number" → Regression
"Average error" → MAE
"Penalize large errors" → MSE
"Same units + penalize large errors" → RMSE

📊 Evaluation Metrics (Clustering)

Clustering is unsupervised — there are no labels. Traditional metrics like accuracy or MAE do not apply.

Silhouette Score

How well each point fits within its cluster (cohesion)
How well it is separated from other clusters

Interpretation: High → well-defined clusters · Near zero → overlap · Negative → wrong cluster

Exam cue: "Evaluate clustering quality" · "How well clusters are separated" · "No ground truth labels"

Cohesion (Intra‑cluster similarity)

How close together points are inside the same cluster
High cohesion → good cluster
Low cohesion → scattered cluster

Key idea: "Do the points in this cluster actually belong together?"

🚨 Cohesion is a concept, not a standalone Azure service.

🚨 Clustering does NOT use: Accuracy, Precision, Recall, F1, MAE, MSE, RMSE.

Exam shortcut:

"Group similar items" / "No labels" → Clustering
"Evaluate cluster quality" → Silhouette / Cohesion
NOT classification or regression metrics

📈 RFM (Recency, Frequency, Monetary)

Customer Segmentation Technique

Recency – How recently they interacted
Frequency – How often they interact
Monetary – How much they spend

Keywords: customer segmentation, marketing analysis, value grouping

Ask yourself:
Is the question about grouping customers by behavior, not prediction?

❌ Trap: RFM = analytics / business logic.
❌ RFM ≠ Machine Learning service.
❌ RFM ≠ Azure AI Vision or Language.

🧠 Convolutional Neural Networks (CNN)

A deep learning model designed for images and visual data. CNNs automatically detect patterns (edges, textures, shapes), combine them into objects, and classify or analyze the image.

Core Idea

CNN = Pattern detector that scans images layer by layer
Early layers → detect edges
Middle layers → detect shapes
Deep layers → detect objects

Keywords: convolution, kernel, filter, feature map, edge detection, pattern recognition, pooling, image classification

Ask yourself:
Is this about images or visual patterns? Are we detecting edges, shapes, or objects? Is the model learning features automatically?

Convolution Layer (Feature Extraction)

Uses filters (kernels) — small matrices (e.g., 3×3)
Slides across image like a scanner
Produces feature maps

Detects: edges, textures, shapes

Feature Maps

Output of convolution
Shows where patterns are detected
Edge detector → highlights edges
Shape detector → highlights shapes

Key idea: Feature maps = "what + where"

Activation (ReLU)

Keeps important signals
Removes weak ones
Adds non-linearity → allows complex learning

Pooling Layer (Downsampling)

Reduces size of feature maps
Max pooling → keep strongest value
Average pooling → smooth values

Benefits: faster computation, less overfitting, robust to position changes

Fully Connected Layer (Decision)

Takes extracted features
Produces final prediction (classification)
Example: "Dog 90%, Cat 5%, Car 5%"

Step	What happens
1. Input	Image (pixels)
2. Convolution	Detect patterns
3. Activation	Keep useful signals
4. Pooling	Reduce size
5. Repeat	Deeper layers = more complex features
6. Fully Connected	Classification

❌ CNN is NOT a traditional ML algorithm
❌ CNN is NOT used for tabular data (usually)
❌ CNN ≠ RNN (sequence data)
❌ CNN ≠ Tokenization (text) / Embeddings / NLP models
✅ CNN = Visual pattern recognition

Exam shortcut:

"Detect objects in image" → CNN
"Classify image" → CNN
"Find edges / shapes" → CNN

🔥 CNN = Learn patterns from images → build features → classify visually

⚖️ Responsible AI Principles (EXAM FAVORITE)

Responsible AI principles often invalidate otherwise "technically correct" options.

Fairness

Bias, equal treatment, demographic impact

Does the system disadvantage a specific group?

Reliability & Safety

Consistent behavior, fail safely, stability

Does the system behave predictably?

Privacy & Security

Personal data, access control, protection, sensitive data

Is sensitive data handled securely?

Transparency

Explainability, user awareness, explanation

Do users know AI is involved?

Accountability

Human oversight, responsibility

Is a human responsible for decisions?

Inclusiveness

Accessibility, different abilities, usable by everyone

Is the system usable by people with different abilities?

🚨 Trap: Accessibility issue → Inclusiveness.
Bias issue → Fairness. These are NOT the same.

⚙️ AI Solution Design Considerations

Beyond ethics — the exam also tests general AI design judgment.

Frequently Tested Considerations

Accuracy vs Explainability
Automation vs Human‑in‑the‑loop
Latency & response time
Cost of inference
Risk of misuse or over‑automation

Ask yourself:
Is full automation appropriate — or should a human remain responsible?

✅ Sometimes the "technically best" option is rejected due to risk or oversight concerns

Generative AI Workloads

Microsoft uses the term "Generative AI workloads"
Includes: text generation, summarization, chat, code generation

Maps to: ✅ Azure OpenAI

❌ Azure AI Language (NER, sentiment) is NOT a GenAI workload
❌ Azure AI Vision is NOT a GenAI workload
❌ AutoML is NOT a GenAI workload
✅ If it creates new content, it's a Generative AI workload

✅ Final Mental Checklist (Use This in the Exam)

What is the input? (text, image, audio)
What is the output? (label, number, text, location)
Is it extracting, comparing, or generating?
Is there structure vs raw content? (Document Intelligence vs OCR)
Is ethics, bias, safety, or accessibility mentioned?

If you answer all 5, the correct option is usually obvious.

🧠 Quick Concept → Trigger Mapping

Concept	Trigger
NaN language detection	Ambiguous text
Entity	Extract value from text
None intent	No match
Indexer	Data ingestion
Skillset	Data enrichment
Pull model	Data retrieval
Embeddings	Semantic meaning
Transformer	Context relationships
Multi-modal	Multiple inputs
CNN	Images
F1	Balanced classification
MAE/MSE/RMSE	Regression
Silhouette	Clustering

🔥 3 Critical Exam Behaviors:

Know the pipeline (Search, ML, NLP)
Don't confuse concepts (Tokens vs Embeddings vs Attention)
Watch for subtle traps (NaN, None intent, Face API approval)