AWS Learning Material Certification Cheatsheets
AIF-C01 Domain 1 20% domain

Task 1.2: Identify practical use cases for AI

Fundamentals of AI and ML · 3,320 words · source: Certified-AI-Practitioner-AIF-C01/domain-1/task-1-2-cheat-sheet.md

Domain 1: Fundamentals of AI and ML

Task Statement 1.2: Identify practical use cases for AI

Domain 1 is 20% of scored exam content. Task 1.2 focuses on deciding when AI/ML provides value, when it is the wrong tool, which technique fits a scenario, and which AWS managed AI/ML service is the best match.

Use this sheet as the bridge between vocabulary from Task 1.1 and scenario questions in AWS Skill Builder practice. Most questions for this task are "given a business problem, choose the best AI approach or AWS service."

Exam Objective Map

You should be able to:

AWS exam objective What to know for the exam
Recognize applications where AI/ML can provide value Human decision support, scalability, automation, pattern recognition, personalization, prediction
Determine when AI/ML is not appropriate Cost-benefit mismatch, insufficient data, deterministic/rules-based outcomes, high explainability needs, operational constraints
Select the appropriate AI/ML technique Regression, classification, clustering, forecasting, anomaly detection, NLP, computer vision, generative AI, agentic AI
Identify real-world AI applications Computer vision, NLP, speech recognition, recommendations, fraud detection, forecasting, knowledge bases, agents
Explain AWS managed AI/ML service capabilities SageMaker AI, Transcribe, Translate, Comprehend, Lex, Polly, plus Bedrock, Rekognition, Personalize as common scenario anchors
Identify when traditional ML or foundation models are appropriate Structured predictive tasks vs generative/language/multimodal tasks; consider regulation, explainability, cost, latency, and operations

The Big Decision

For scenario questions, ask three questions in order:

1. Is the outcome deterministic or predictive?
2. What type of input and output are involved?
3. Is a managed AWS AI service enough, or do you need custom ML/model building?

Quick rule:

Exact rules -> traditional software or rules engine.
Prediction from historical data -> traditional ML.
Language/content generation or Q&A -> foundation model / generative AI.
Conversational task completion with tools -> agentic AI.

When AI/ML Provides Value

AI/ML is a good fit when the problem involves patterns, uncertainty, scale, or human-language/visual data.

Value pattern Use AI/ML when... Example
Assist human decision making Humans need recommendations, scores, summaries, or risk signals. Fraud risk score for an analyst
Solution scalability Manual review does not scale across large data volumes. Classify millions of support tickets
Automation Repetitive perception, prediction, or language tasks can be automated. Transcribe calls and detect sentiment
Pattern recognition Rules are hard to write because patterns are complex. Detect defects in images
Personalization Outputs should adapt to user behavior. Product or content recommendations
Prediction Historical data can estimate future outcomes. Forecast demand or churn
Natural language understanding Text must be classified, summarized, translated, or searched semantically. Extract entities from legal documents
Knowledge access Users need answers grounded in enterprise documents. Knowledge base chatbot with citations

Exam signal phrases:

  • "Large volume of documents, calls, images, or events"
  • "Patterns are too complex for manual rules"
  • "Improve over time using data"
  • "Predict likelihood, risk, demand, or category"
  • "Summarize, classify, translate, or generate natural language"
  • "Personalize recommendations"

When AI/ML Is Not Appropriate

AWS Well-Architected guidance emphasizes checking whether ML is really the right solution before paying for data preparation, training, hosting, and maintenance.

Do not choose AI/ML when... Better option Exam clue
A specific deterministic outcome is required Rules, workflow logic, SQL, validation logic "Must always return the exact same approved result"
Simple rules solve the problem cheaply Rules-based application "If amount > X, route to manager"
There is not enough relevant data Collect data first or use rules/human review "No historical examples are available"
Data quality is poor and cannot be fixed Data engineering/governance first "Missing, inconsistent, or biased data"
Cost exceeds expected business value Simpler baseline or no automation "Low-volume process with small benefit"
Latency/cost constraints make model inference impractical Cached rules, simpler model, process redesign "Strict millisecond budget and low tolerance for variance"
Full explainability is mandatory and model behavior must be transparent Rules or interpretable traditional ML "Regulator requires clear decision logic"
Errors cannot be tolerated without review Human-in-the-loop, deterministic controls "Incorrect output could create unacceptable harm"

Exam trap: "AI can automate this" is not enough. You must compare against simpler alternatives and business value.

Technique Chooser

Use case Best technique Output AWS anchor
Predict a number Regression Numeric value SageMaker AI
Predict one of two categories Binary classification Yes/no, fraud/not fraud SageMaker AI, Comprehend custom classification for text
Predict one of many categories Multi-class classification Category label SageMaker AI, Comprehend
Group similar items without labels Clustering Segments/groups SageMaker AI
Detect unusual behavior Anomaly detection Normal/abnormal or risk score SageMaker AI
Predict future values over time Forecasting Future demand/sales/traffic SageMaker AI
Recommend items to users Recommendation system Ranked items Amazon Personalize or SageMaker AI
Understand or classify text NLP Sentiment, entity, topic, class Amazon Comprehend
Convert speech to text Speech recognition / ASR Transcript Amazon Transcribe
Convert text between languages Machine translation Translated text Amazon Translate
Convert text to speech Text-to-speech Audio Amazon Polly
Build a voice/text chatbot Conversational AI Dialog and fulfilled intent Amazon Lex
Analyze images or video Computer vision Labels, faces, objects, text Amazon Rekognition or SageMaker AI
Generate text, images, code, summaries, or answers Generative AI / foundation model New content Amazon Bedrock
Answer from company documents RAG / knowledge base Grounded answer with retrieved context Amazon Bedrock Knowledge Bases
Complete multi-step tasks with tools Agentic AI Actions, API calls, workflow steps Amazon Bedrock Agents, Amazon Lex with integrations

Regression vs Classification vs Clustering

These three are explicitly named in Task 1.2. Memorize them.

Technique Use when... Example Common wrong answer
Regression The target is numeric and continuous. Predict revenue, price, delivery time, temperature Classification
Binary classification The target is one of two categories. Fraud/not fraud, churn/not churn, approved/denied Regression
Multi-class classification The target is one of more than two categories. Ticket category, product type, document class Clustering
Clustering There are no labels and you need natural groups. Customer segments from behavior data Classification

Memory hook:

Number -> regression.
Known label -> classification.
Unknown group -> clustering.

Real-World Application Map

Application What it does Best fit Exam clue
Computer vision Extracts information from images/video. Rekognition for managed APIs; SageMaker AI for custom models "Images, video, defects, faces, labels, OCR"
NLP Understands or processes human language. Comprehend, Bedrock, SageMaker AI "Sentiment, entities, key phrases, topics, classify text"
Speech recognition Converts audio speech into text. Transcribe "Call recordings, subtitles, voice notes, transcript"
Translation Converts text from one language to another. Translate "Localize content, multilingual chat, cross-language support"
Text-to-speech Converts text into spoken audio. Polly "Read article aloud, voice prompt, audio output"
Conversational interface Captures user intent through voice or text dialog. Lex "Chatbot, appointment booking, intent, slots"
Recommendation system Ranks items for users based on behavior. Personalize or SageMaker AI "Users also bought, personalized ranking, next best item"
Fraud detection Predicts suspicious activity. SageMaker AI, anomaly detection/classification "Risk score, abnormal transaction, fraud/not fraud"
Forecasting Predicts future values over time. SageMaker AI "Demand, inventory, sales, traffic by day/month"
Knowledge base Retrieves relevant enterprise content for answers. Bedrock Knowledge Bases "Answer from internal docs, citations, grounded response"
Agentic AI Plans and uses tools to complete tasks. Bedrock Agents, Lex plus Lambda, Amazon Q patterns "Take action, call APIs, open ticket, schedule, update records"

AWS Managed Service Chooser

Core services named in Task 1.2

Service Choose it for... Do not confuse with...
Amazon SageMaker AI Building, training, customizing, and deploying custom ML or foundation models. Bedrock, which gives managed access to existing foundation models for GenAI apps
Amazon Transcribe Speech-to-text from real-time streams or batch audio files. Polly, which is text-to-speech
Amazon Translate Neural machine translation for text. Comprehend, which analyzes text but does not translate it
Amazon Comprehend NLP insights from text: sentiment, entities, key phrases, language, topics, PII, custom classification. Bedrock, which is better for open-ended generation and summarization
Amazon Lex Voice/text conversational chatbots using ASR and NLU, with intents, slots, and fulfillment. Bedrock Agents, which can orchestrate GenAI-powered multi-step tasks
Amazon Polly Text-to-speech with lifelike voices and audio streams/files. Transcribe, which is speech-to-text

Common adjacent services

Service Choose it for...
Amazon Bedrock Build generative AI apps with managed foundation models, model customization, RAG, and agents.
Amazon Bedrock Knowledge Bases Connect FMs to private data for retrieval-augmented generation and grounded answers.
Amazon Rekognition Managed computer vision APIs for image and video analysis.
Amazon Personalize Real-time personalized recommendations.
Amazon Q AI assistant capabilities for business, development, or AWS expertise depending on product variant.

Exam trap: managed AI services are best when the task matches a prebuilt capability. Choose SageMaker AI when you need to build, train, customize, or deploy your own model workflow.

Traditional ML vs Foundation Models

Factor Traditional ML is usually better when... Foundation models are usually better when...
Data type Structured/tabular historical data Text, code, images, multimodal, or broad unstructured content
Output Specific prediction, score, class, or numeric value Generated content, summaries, answers, transformations
Task shape Narrow task with clear labels and metrics Broad language/content task or many tasks through prompts
Explainability Decision reasoning must be more interpretable Some opacity is acceptable or can be controlled with guardrails/RAG/evaluation
Regulation Strict auditability favors simpler, controlled models GenAI is acceptable with governance, grounding, monitoring, and review
Cost and latency High-volume, low-latency, predictable inference Value justifies FM inference cost/latency
Operations You can train and manage a model lifecycle You want managed FM access through API
Examples Churn, fraud score, pricing prediction, demand forecast, customer segment Chatbot, summarization, document Q&A, code generation, email drafting

Scenario shortcuts:

  • Use traditional ML for "predict a value/category from structured historical records."
  • Use FMs for "generate, summarize, reason over text, answer questions, or transform unstructured content."
  • Use RAG/knowledge bases when "answers must use company documents."
  • Use rules when "the answer must be exact and deterministic."

Foundation Model Use-Case Fit

Use case FM fit? Reason
Summarize long documents Strong fit Language generation and compression
Generate marketing copy Strong fit Creative text generation
Answer questions from internal policies Strong fit with RAG Needs retrieval from trusted documents
Extract a fixed field from a form Maybe Use Textract/rules if the document pattern is stable; use FM if language variation is high
Approve a loan with strict regulatory explainability Be careful Traditional ML or rules may be easier to explain and audit
Predict monthly sales from historical tabular data Usually not first choice Forecasting/traditional ML is more direct
Route support tickets into known categories Maybe Comprehend/custom classification or SageMaker may be cheaper and more controlled
Give a guaranteed exact tax calculation Poor fit Deterministic rules are required

Service Pairing Patterns

AWS scenarios often combine services.

Pattern Services Example
Call center analytics Transcribe + Comprehend Transcribe calls, then detect sentiment and key phrases
Multilingual support Translate + Comprehend + Lex Translate customer text, understand intent/sentiment, route conversation
Voice assistant Lex + Polly + Lambda Understand request, run business logic, respond with speech
Speech-to-speech translation Transcribe + Translate + Polly Convert spoken language to translated spoken output
Document Q&A Bedrock + Knowledge Bases + S3/vector store Answer employee questions from internal policies
Recommendation engine Personalize + application data Recommend products or content in real time
Custom predictive model SageMaker AI + S3 + endpoints Train and deploy a churn prediction model

Cost-Benefit and Baseline Thinking

Before choosing ML, establish a baseline:

Question Why it matters
What business decision will the model improve? Avoids building AI without a measurable outcome
What data is available? ML needs sufficient, relevant, high-quality data
What is the current/rules-based baseline? Proves ML adds value over a simpler approach
What are data prep, training, hosting, and maintenance costs? ML cost continues after launch
What metric defines success? Accuracy alone may not equal business value
What happens when the model is wrong? Determines human review, safeguards, and service choice
What explainability or auditability is required? Influences rules vs traditional ML vs foundation model

Exam trap: a model with slightly better accuracy may still be the wrong answer if its cost, latency, risk, or maintenance is too high.

Human-in-the-Loop Signals

Use human review when:

  • Decisions affect safety, finance, employment, healthcare, legal rights, or customer trust.
  • The model output is probabilistic and consequences of errors are high.
  • Regulations require review or appeal.
  • The model handles edge cases, low-confidence predictions, or sensitive data.
  • You need labeled data for continuous improvement.

Human-in-the-loop does not mean "AI is useless." It means AI assists or triages while humans make final decisions.

Exam Trap Table

Scenario wording Best answer
"Need exact deterministic calculation every time" Rules/application logic, not ML/FM
"Predict future numeric demand" Forecasting/regression with traditional ML
"Predict whether a transaction is fraudulent" Classification/anomaly detection
"Group customers without existing labels" Clustering
"Convert call audio to text" Amazon Transcribe
"Read generated text aloud" Amazon Polly
"Translate support chats between languages" Amazon Translate
"Detect sentiment and entities in reviews" Amazon Comprehend
"Build a chatbot that identifies user intents and slots" Amazon Lex
"Generate answers from company documents with citations" Amazon Bedrock Knowledge Bases
"Build a GenAI app using managed FMs" Amazon Bedrock
"Build, train, and deploy a custom model" Amazon SageMaker AI
"Recommend products based on behavior" Amazon Personalize or custom recommendation model
"Strict explainability for regulated predictive decision" Rules or interpretable traditional ML before FM

One-Page Memorization Version

  • AI/ML is best for prediction, pattern recognition, automation, personalization, language, speech, images, and scale.
  • Do not use AI/ML when deterministic rules are enough, data is insufficient, costs exceed value, or exact explainability is mandatory.
  • Number output means regression.
  • Known category output means classification.
  • Unknown group discovery means clustering.
  • Future time-based values mean forecasting.
  • Audio to text is Transcribe.
  • Text to audio is Polly.
  • Text translation is Translate.
  • Text insights are Comprehend.
  • Voice/text chatbot intents are Lex.
  • Custom model lifecycle is SageMaker AI.
  • Generative AI with foundation models is Bedrock.
  • Grounded document Q&A is Bedrock Knowledge Bases.
  • Recommendations are Personalize or SageMaker AI.
  • Traditional ML fits structured predictive problems; foundation models fit generation, summarization, Q&A, and broad language/content tasks.

Mini Practice Questions

  1. A retailer wants to predict next month's inventory demand from five years of sales data. Which technique fits?

    • Answer: Forecasting, often using regression/time-series ML. SageMaker AI is an AWS anchor.

  2. A company wants to always calculate tax using published tax tables. Should it use an ML model?

    • Answer: No. This is deterministic rules/application logic.

  3. A support team wants to group unlabeled tickets to discover common themes. Which technique fits?

    • Answer: Clustering or topic modeling. If using managed NLP, Amazon Comprehend can help with text insights.

  4. A contact center wants searchable transcripts from recorded calls. Which service fits?

    • Answer: Amazon Transcribe.

  5. An app needs to convert English text into German text for customer chats. Which service fits?

    • Answer: Amazon Translate.

  6. A company wants a chatbot that asks users for missing information, stores slot values, and invokes backend logic. Which service fits?

    • Answer: Amazon Lex, often with AWS Lambda for fulfillment.

  7. A legal team wants answers from internal policy documents with source citations. Which approach fits?

    • Answer: RAG with Amazon Bedrock Knowledge Bases.

  8. A bank needs a highly explainable decision process for loan eligibility. Should it default to a foundation model?

    • Answer: No. Consider rules or interpretable traditional ML first because explainability and regulatory constraints are central.

  9. A streaming service wants personalized movie recommendations. Which service or approach fits?

    • Answer: Amazon Personalize or a custom recommendation model in SageMaker AI.

  10. A marketing team wants to draft multiple variants of product copy. Which AWS service fits?

    • Answer: Amazon Bedrock with foundation models.

Sources

Official AWS sources used: