AWS Certified AI Practitioner — Task Statement 1.2 Study Notes

Identify Practical Use Cases for AI

Task Statement 1.2 focuses on recognizing where AI/ML adds business value, selecting appropriate ML approaches, knowing when AI is not suitable, and matching use cases to AWS managed AI/ML services.

1. When AI/ML Can Provide Value

AI and ML are especially useful when a problem involves patterns, predictions, language, images, speech, or large volumes of data.

Common sources of value

1.1 Automating repetitive work

AI can perform repetitive, tedious, or time-consuming tasks continuously.

Examples:

Classifying incoming customer support requests
Extracting information from invoices
Transcribing meetings and call-center recordings
Screening uploaded images for inappropriate content
Routing customer calls
Detecting duplicate or suspicious transactions

Benefits include:

Lower employee workload
Faster processing
Fewer manual errors
More consistent operations
Ability to operate 24/7

1.2 Supporting human decision-making

AI can analyze data and provide recommendations, scores, or predictions to help people make decisions.

Examples:

Fraud probability scores for financial transactions
Loan-risk assessments
Product recommendations
Predictive maintenance alerts
Medical decision support
Customer sentiment analysis

Important: AI often assists humans rather than completely replacing them. High-impact decisions may require human review.

1.3 Analyzing large amounts of data at high speed

AI is valuable when the data volume or processing speed exceeds human capability.

Examples:

Scanning millions of transactions for fraud
Searching large enterprise document repositories
Analyzing sensor streams from oil wells
Detecting objects in real-time video
Finding trends in customer feedback
Processing large collections of text, images, or audio

1.4 Recognizing patterns and deviations

ML models are effective at finding patterns and identifying unusual behavior.

Examples:

Fraud detection
Network intrusion detection
Failed sensor detection
Manufacturing defect detection
Medical error detection
Account takeover detection

1.5 Forecasting and reducing waste

AI can predict future demand, events, or resource requirements.

Examples:

Product demand forecasting
Predictive maintenance
Energy consumption forecasting
Inventory planning
Predicting delivery times
Forecasting staffing requirements

The business benefits may include:

Less inventory waste
Reduced downtime
Lower operational costs
Better resource allocation
Improved customer service

2. When AI/ML Is Not Appropriate

AI is not automatically the best answer to every problem. The exam may test whether you can recognize when a simpler, deterministic, or non-AI solution is more appropriate.

2.1 When the cost exceeds the benefit

Training and operating ML systems can require significant resources:

Data collection and preparation
Data labeling
Computing power
Model training
Model deployment
Monitoring
Retraining
Security and governance
Human review of uncertain predictions

Before building an AI solution, perform a cost-benefit analysis.

Example

A company wants to reduce fraud losses by $50,000 per year.

If the total cost of building, operating, monitoring, and retraining the model is $100,000 per year, the AI solution is probably not economically justified.

Exam approach

Look for questions asking whether AI should be used when:

The problem is small and easily solved with rules
The expected savings are minimal
There is not enough data to justify model development
A managed service or simple automation would be cheaper
The operational cost of inference is greater than the business value

Exam tip: Do not assume that a more advanced AI model is automatically better. The correct solution is often the one that satisfies the business requirement at the lowest reasonable cost and complexity.

2.2 When a specific outcome is required

ML models generally make predictions based on learned patterns. They do not guarantee a specific result.

If the application must always produce an exact, predictable result, use:

Business rules
Traditional software logic
A deterministic workflow
A rule-based system
A state machine

Deterministic systems

A system is deterministic when the same input always produces the same output, assuming the rules and system state have not changed.

Example:

If credit score >= 750
and loan amount <= $10,000
then approve automatically.

This is a rule-based decision and is deterministic.

ML systems

ML systems estimate probabilities or predictions based on data. Their outputs may be affected by:

Model parameters
Input data
Model version
Thresholds
Randomness in some model processes
Changes in the underlying data

Examples:

Probability that a transaction is fraudulent
Probability that a customer will buy a product
Predicted house price
Predicted demand next month

Important distinction

Requirement	Better approach
Always apply the same business rule	Rule-based system
Predict whether an event is likely	ML classification
Estimate a numerical value	Regression
Discover groups without labels	Clustering
Find unusual observations	Anomaly detection
Generate or summarize flexible language	Foundation model

Exam trap: “AI is more intelligent, so it should be used for all decisions” is incorrect. If the required behavior is fully specified and deterministic, traditional software may be more appropriate.

2.3 When explainability or interpretability is mandatory

Some AI decisions affect people’s finances, employment, healthcare, insurance, or access to services. In these scenarios, the organization may need to explain why a decision was made.

Interpretability

Interpretability is the ability to understand how a model arrived at a prediction or decision.

Complex models, especially deep neural networks and foundation models, can be difficult to interpret. This is often referred to as the black-box problem.

Tradeoff

There may be a tradeoff between:

Model complexity and predictive performance
Explainability and accuracy
Flexibility and transparency

A less complex model may be easier to explain, even if it is less accurate.

Possible alternatives include:

Linear or logistic regression
Decision trees
Rule-based systems
Simpler scoring models
Human review workflows

Exam tip: If complete transparency is a business, legal, or compliance requirement, consider a simpler traditional model or rules-based approach instead of a highly complex model or FM.

2.4 When there is insufficient or unsuitable data

ML typically needs relevant, representative data.

Potential problems include:

Too few training examples
Poor-quality data
Biased data
Missing labels
Data that does not represent production conditions
Privacy or regulatory restrictions
Data that is too old or irrelevant

In such cases, alternatives may include:

A rule-based approach
A managed pre-trained AI service
Human review
Collecting more data
Using a foundation model with appropriate grounding
Redesigning the problem

2.5 When the problem is not predictive

ML is generally appropriate for finding patterns, classifying information, or predicting likely outcomes.

It may not be appropriate when:

The answer is already known through a simple lookup
A database query solves the problem
A fixed workflow is sufficient
A mathematical formula gives the required answer
A deterministic policy must be enforced
No useful historical data exists

3. Supervised and Unsupervised Learning

Understanding the data is the first step in selecting an ML technique.

3.1 Supervised learning

In supervised learning, the training data contains:

Input features or attributes
Known target values or labels

The model learns the relationship between inputs and known outputs.

Examples

Inputs	Known target
Transaction details	Fraud or not fraud
House features	House price
Medical test results	Disease or no disease
Email contents	Spam or not spam
Customer information	Customer churn outcome

Supervised learning is used for:

Classification
Regression

3.2 Unsupervised learning

In unsupervised learning, the data contains inputs but no target labels.

The model attempts to discover patterns or structures in the data.

Examples:

Grouping customers by behavior
Finding unusual sensor readings
Discovering topics in documents
Identifying natural segments in a population

Unsupervised learning is used for:

Clustering
Anomaly detection
Some dimensionality-reduction and pattern-discovery tasks

4. Selecting the Correct ML Technique

4.1 Classification

Classification predicts a categorical or discrete value.

The output belongs to one or more predefined classes.

Binary classification

Binary classification has two mutually exclusive classes.

Examples:

Fraud or not fraud
Spam or not spam
Disease or no disease
Customer will churn or will not churn
Fish or not fish

Multiclass classification

Multiclass classification has more than two possible classes.

Examples:

Classifying a document as politics, finance, religion, or another topic
Identifying an animal species
Categorizing a support ticket as billing, technical, account, or shipping
Identifying the type of product in an image

Classification output

A model may produce:

A predicted class
A probability or confidence score for each class

Example:

Fraud: 0.92
Not fraud: 0.08

The organization may select a threshold for taking action.

Exam clues for classification

Look for words such as:

Category
Class
Label
Yes/no
Fraud/not fraud
Type
Topic
Sentiment
Approval status

4.2 Regression

Regression predicts a continuous numerical value.

Examples:

House price
Product demand
Temperature
Revenue
Delivery time
Energy consumption
Number of units likely to be sold

Linear regression

Linear regression models a relationship between input variables and a continuous output.

Simple linear regression

Uses one independent variable.

Example:

Predicting height from weight

Multiple linear regression

Uses multiple independent variables.

Example:

Predicting house price using:
Number of bedrooms
Number of bathrooms
Square footage
Lot size
Location-related features

Logistic regression

Despite its name, logistic regression is normally used for classification, particularly binary classification.

It estimates the probability of an event occurring, usually between 0 and 1.

Examples:

Probability that a person has heart disease
Probability that a transaction is fraudulent
Probability that a customer will default on a loan

Example:

Fraud probability = 0.87

The business can then apply a threshold:

If fraud probability >= 0.80, send for review.

Exam trap: Logistic regression is a classification technique, not a continuous-value regression technique.

Exam clues for regression

Look for:

Predict a price
Estimate a quantity
Forecast demand
Predict a temperature
Calculate expected revenue
Predict a numerical score or measurement

4.3 Clustering

Clustering groups unlabeled data into similar groups, called clusters.

The goal is to make:

Items within the same cluster as similar as possible
Items in different clusters as different as possible

The algorithm uses selected features and a similarity or distance function.

Example: customer segmentation

A company may group customers based on:

Purchase frequency
Average order value
Product preferences
Website activity
Clickstream behavior

The resulting clusters could represent:

High-value frequent buyers
Occasional discount shoppers
New customers
Customers likely to stop purchasing

Exam clues

Look for:

No labels
Discover natural groups
Segment customers
Group similar users
Organize data into categories that were not predefined

Exam trap: If the classes are already labeled, the problem is classification. If the groups must be discovered from unlabeled data, it is clustering.

4.4 Anomaly detection

Anomaly detection identifies rare or unusual observations that differ significantly from normal behavior.

Examples:

Fraudulent transactions
Failed sensors
Network intrusions
Medical errors
Manufacturing defects
Unusual account activity
Leaks in pipelines or storage tanks

Anomalies may not be known in advance, so anomaly detection is often associated with unsupervised or semi-supervised learning.

Exam clues

Look for:

Rare events
Outliers
Deviations from normal behavior
Unusual activity
Failed equipment
Suspicious observations

Classification versus anomaly detection

These can appear similar, but the data is different:

Scenario	Appropriate technique
Historical transactions are labeled fraud/not fraud	Classification
The system must discover unusual transactions without labels	Anomaly detection
Customers are divided into predefined categories	Classification
The system discovers natural customer groups	Clustering

5. Common Real-World AI Applications

5.1 Computer vision

Computer vision enables systems to understand images and video.

Use cases:

Object detection
Face detection and recognition
Image classification
Video analysis
Content moderation
Optical character recognition
Defect detection
Visual search
Identity verification

AWS services

Amazon Rekognition: Image and video analysis
Amazon Textract: Text, handwriting, forms, and tables in documents

5.2 Natural language processing

NLP allows computers to process and understand human language.

Use cases:

Sentiment analysis
Text classification
Entity extraction
PII detection
Topic detection
Search
Summarization
Chatbots
Question answering
Translation

AWS services

Amazon Comprehend
Amazon Kendra
Amazon Lex
Amazon Translate
Amazon Bedrock

5.3 Speech recognition

Speech recognition converts spoken audio into text.

Use cases:

Call transcription
Meeting transcription
Real-time captions
Voice analytics
Search across recorded calls
Voice-controlled applications

AWS service

Amazon Transcribe

5.4 Text-to-speech

Text-to-speech converts written text into spoken audio.

Use cases:

Reading news articles aloud
Accessibility applications
Voice prompts
Interactive voice response systems
Audio content generation

AWS service

Amazon Polly

5.5 Recommendation systems

Recommendation systems predict products, media, content, or services that users may be interested in.

Examples:

“You might also like”
Recommended movies
Suggested music
Personalized travel options
Product recommendations
Personalized marketing campaigns

AWS service

Amazon Personalize

5.6 Fraud detection

Fraud detection identifies suspicious or potentially fraudulent activity.

Examples:

Payment fraud
Fake account creation
Account takeover
Suspicious product reviews
Unusual transactions

AWS services

Amazon Fraud Detector
Amazon SageMaker AI for custom models
Other AWS services may support the surrounding data and analytics pipeline

5.7 Forecasting

Forecasting predicts future numerical values or events.

Examples:

Demand forecasting
Sales forecasting
Resource planning
Inventory requirements
Energy usage
Equipment failure risk

Typically, forecasting is a regression or time-series ML problem.

5.8 Knowledge bases and intelligent search

Knowledge systems help users find relevant information in large collections of enterprise content.

Examples:

Searching internal documents
Answering employee questions
Finding technical procedures
Searching policies and manuals
Providing grounded answers from company data

AWS service

Amazon Kendra: Intelligent enterprise search using natural language understanding

5.9 Agentic AI

Agentic AI applications can:

Understand a user’s goal
Plan or decide what actions are needed
Call tools, APIs, or external systems
Retrieve information
Take actions
Return a response

Examples:

A travel assistant that searches inventory and books a trip
A customer service agent that looks up an order and initiates a refund
An IT assistant that checks system status and creates a ticket

Foundation models are often used as the reasoning and language component, while tools and APIs perform controlled actions.

Exam consideration: Agentic AI should use appropriate permissions, guardrails, monitoring, and human approval for high-impact actions.

6. AWS Managed AI/ML Services

For common use cases, AWS provides pre-trained, managed services that can be accessed through APIs. These services can eliminate the need to collect large datasets, train models, and manage infrastructure yourself.

6.1 Amazon Rekognition

A managed computer vision service for images and videos.

Capabilities include:

Face detection
Face comparison
Face recognition
Object and scene detection
Image and video labeling
Real-time video analysis
Custom labels for proprietary objects
Text detection in images
Unsafe or inappropriate content moderation

Example use cases

Identity verification
Security monitoring
Searching image and video libraries
Detecting objects in streaming video
Moderating user-uploaded content

Exam tip: Rekognition is for image and video analysis. It is not the primary service for extracting tables and forms from scanned documents; that is Amazon Textract.

6.2 Amazon Textract

A managed document analysis service.

It extracts:

Printed text
Handwriting
Forms
Key-value pairs
Tables
Structured information from scanned documents

Common workflow

Scanned document
        ↓
Amazon Textract extracts text and fields
        ↓
Amazon Comprehend analyzes the extracted text

Example:

Textract extracts text from a customer feedback form
Comprehend determines the sentiment and identifies entities

6.3 Amazon Comprehend

A managed NLP service that discovers insights and relationships in text.

Capabilities include:

Sentiment analysis
Key phrase extraction
Entity recognition
Language detection
Text classification
Topic analysis
PII detection

PII detection

Comprehend can identify information such as:

Names
Addresses
Email addresses
Phone numbers
Credit card numbers

It can provide confidence scores. A business can define a confidence threshold for automatic action, such as redacting information only when confidence exceeds a specified level.

Exam tip: Confidence thresholds can help balance automation with false positives and false negatives. A higher threshold generally means fewer automatic actions but potentially more missed detections.

6.4 Amazon Lex

A managed service for building conversational interfaces using voice and text.

It uses technologies similar to those behind Amazon Alexa.

Common use cases:

Customer service chatbots
Interactive voice response systems
Call routing
Self-service applications
Conversational interfaces for websites and applications

Example

A traditional IVR may require users to press numbers. Lex allows customers to state their request naturally, such as:

“I need to change my delivery address.”

6.5 Amazon Transcribe

A managed automatic speech recognition service.

Capabilities include:

Converting speech to text
Processing live audio
Processing recorded audio and video
Supporting many languages
Providing captions and transcripts
Enabling search and analysis of spoken content

Common use cases:

Real-time captions
Call-center transcription
Meeting transcription
Media subtitling
Voice analytics

Exam trap: Transcribe converts speech to text. Polly converts text to speech.

6.6 Amazon Polly

A managed text-to-speech service.

It converts text into natural-sounding speech using deep learning.

Common use cases:

Reading articles aloud
Accessibility applications
Voice prompts
IVR systems
Audio versions of news content

Exam tip:
- Transcribe: speech → text
- Polly: text → speech

6.7 Amazon Translate

A managed neural machine translation service.

It translates text between supported languages while considering the context of the sentence.

Common use cases:

Multilingual customer support
Real-time chat translation
Translating documents
Localizing applications and websites

6.8 Amazon Kendra

An intelligent enterprise search service.

It uses NLP to understand natural-language questions and find relevant information across enterprise repositories.

Example query:

“How do I connect my Echo Plus to my network?”

Kendra is useful when organizations need search over:

Internal documents
Knowledge repositories
Policies
Manuals
Technical documentation
Enterprise content systems

6.9 Amazon Personalize

A managed recommendation service.

It can create personalized recommendations based on user behavior and preferences.

Common use cases:

Retail product recommendations
Media recommendations
Entertainment personalization
Personalized marketing
Customer segmentation

Example:

“Customers who viewed this item may also like…”

6.10 Amazon Fraud Detector

A managed service for detecting potentially fraudulent online activity.

Common use cases include:

Online payment fraud
Fake account creation
Account takeover
Fraudulent product reviews
Suspicious checkout activity

It provides models and capabilities designed for common online fraud scenarios.

6.11 Amazon Bedrock

Amazon Bedrock is a fully managed service for building generative AI applications using foundation models.

Capabilities include:

Access to foundation models from Amazon and other providers
Text generation
Summarization
Question answering
Image generation, depending on the selected model
Model customization
Knowledge bases
Retrieval Augmented Generation
Integration into generative AI applications

Foundation models

Foundation models are large, broadly trained models that can be adapted to many tasks.

They can support:

Content generation
Conversational assistants
Summarization
Code generation
Classification
Question answering
Image generation

Retrieval Augmented Generation

RAG allows a model to retrieve information from an external knowledge source before generating a response.

User question
      ↓
Retrieve relevant enterprise data
      ↓
Provide retrieved data to the model
      ↓
Generate a grounded response

Benefits include:

More current answers
Answers based on private organizational data
Reduced reliance on the model’s original training data
Better grounding and fewer unsupported responses

Exam trap: RAG does not retrain the foundation model. It retrieves relevant information and provides it to the model at inference time.

6.12 Amazon SageMaker AI

Amazon SageMaker AI is used when an organization needs custom ML models, workflows, or greater control than prebuilt AI services provide.

It supports the ML lifecycle, including:

Data preparation
Data labeling
Feature engineering
Model development
Model training
Distributed training
GPU-based training
Model evaluation
Model deployment
Real-time inference endpoints
Batch inference
Model monitoring
Retraining workflows
Use of pre-trained models as starting points

When to choose SageMaker AI

Use SageMaker AI when:

A pre-trained managed AI service does not solve the problem
The organization has proprietary data
A custom model is required
Specific model architecture or algorithm control is needed
A custom training workflow is required
The model must be deployed with custom inference behavior
The organization needs control over the ML lifecycle

Exam tip: Use managed AI services for common, well-defined tasks. Use SageMaker AI for custom model development and broader ML workflows.

7. Traditional ML Models Versus Foundation Models

The choice between a traditional ML model and a foundation model depends on the task, data, risk, cost, and operational requirements.

7.1 Traditional ML models

Traditional models are usually trained for a specific task using task-specific data.

Examples:

Linear regression
Logistic regression
Decision trees
Random forests
Gradient-boosted models
Custom neural networks
Classification models
Forecasting models

Advantages

Focused on a specific problem
Often less expensive to run
Easier to evaluate for a defined task
May be easier to explain
Lower latency in some situations
More predictable output format
Easier to constrain to a specific business purpose

Appropriate when

The task is narrow and well-defined
Labeled data is available
Low latency is important
Cost must be minimized
Explainability is required
Regulatory controls favor simpler models
The output must follow a strict format
A numerical prediction or classification is needed

Example:

Predict whether a transaction is fraudulent using structured transaction data.

A traditional classification model may be more suitable than a general-purpose FM.

7.2 Foundation models

Foundation models are broadly trained models that can be adapted to many tasks using prompts, customization, or additional context.

Advantages

Flexible
Useful for many language and content-generation tasks
Can work with limited task-specific labeled data
Can support natural-language interaction
Can summarize, generate, classify, and answer questions
Can be integrated into agentic workflows

Appropriate when

The task involves natural language
Content generation is needed
The task is open-ended
Users interact conversationally
Multiple related tasks are required
There is limited labeled training data
RAG can provide organization-specific context
Rapid prototyping and flexibility are important

Examples:

Customer service assistant
Document summarization
Natural-language travel planner
Content generation
Conversational question answering
Agent that calls business APIs

7.3 Factors affecting the choice

Factor	Traditional ML	Foundation model
Task	Narrow and specific	Broad or open-ended
Data	Labeled task-specific data	May work with prompts or retrieved context
Explainability	Often easier	Often more difficult
Cost	Often lower for simple tasks	Can be higher depending on usage
Latency	Often lower	May be higher
Output	Predictable class/value	Flexible generated content
Regulation	Often easier to govern	Requires stronger controls
Customization	Retrain or tune a specific model	Prompting, fine-tuning, or RAG
Best for	Classification, regression, forecasting	Generation, summarization, assistants

Exam trap: Foundation models are not always the best choice. A small traditional model may be more appropriate for a simple binary classification task with strict cost, latency, or explainability requirements.

8. Real-World Examples

8.1 Mastercard — fraud detection

Mastercard uses AI and ML to score transactions based on their probability of fraud.

Benefits included:

Increased detection of fraudulent transactions
Reduced false positives
Faster transaction analysis

Generative AI can also use transaction history and contextual information to improve the fraud score.

Exam concept

Fraud detection may use:

Classification when historical fraud labels exist
Anomaly detection when unusual patterns must be discovered
Traditional ML for structured transaction scoring
Generative AI for additional contextual reasoning

8.2 DoorDash — conversational IVR

DoorDash used Amazon Lex to replace an IVR system that required customers to navigate touch-tone menus.

Benefits included:

Natural-language interaction
Reduced hold times
Increased self-service adoption
Improved customer experience

Exam concept

Use Amazon Lex when the requirement is a voice or text conversational interface.

8.3 Laredo Petroleum — predictive maintenance and monitoring

Laredo used streaming sensor data from oil and gas wells, including:

Pressure
Temperature
Flow rate

SageMaker AI models helped identify where maintenance teams should focus attention and detect issues such as:

Potential equipment problems
Leaks
Environmental risks
Flaring or venting events

Exam concept

This is an example of:

Real-time ML inference
Predictive maintenance
Anomaly detection
Streaming data analysis
Custom models built with SageMaker AI

8.4 Booking.com — recommendations and generative AI

Booking.com uses ML recommendations for travel-related offerings.

Its AI Trip Planner uses generative AI to understand a customer’s request and then retrieve relevant booking recommendations and reviews.

This is an example of RAG because the model uses external, current information rather than relying only on its original training data.

Exam concept

A generative AI assistant that retrieves current product or booking data is an example of RAG.

8.5 Pinterest — visual search

Pinterest Lens allows users to take a picture of an object and find visually similar products.

The workflow may involve:

Computer vision
Image embeddings or similarity search
Labeled product images
Human-assisted data labeling
Frequent model retraining
Amazon S3 for image storage
SageMaker Ground Truth or labeling tools for annotation

Exam concept

Visual similarity and identifying objects in images are computer vision use cases.

9. Service Selection Quick Reference

Requirement	AWS service
Analyze images and videos	Amazon Rekognition
Extract text, forms, handwriting, and tables from documents	Amazon Textract
Analyze sentiment, entities, topics, or PII in text	Amazon Comprehend
Build voice or text chatbots	Amazon Lex
Convert speech to text	Amazon Transcribe
Convert text to speech	Amazon Polly
Translate text	Amazon Translate
Intelligent enterprise search	Amazon Kendra
Personalized recommendations	Amazon Personalize
Detect online fraud	Amazon Fraud Detector
Build generative AI applications with FMs	Amazon Bedrock
Build and train custom ML models	Amazon SageMaker AI
Label data for ML	SageMaker Ground Truth
Retrieve enterprise information for generative AI	Bedrock Knowledge Bases / RAG

10. Exam Tips and Common Traps

Tip 1: Start with the output type

Ask what the model needs to produce:

Category → classification
Number → regression
Group → clustering
Unusual event → anomaly detection
Generated text or image → foundation model/generative AI

Tip 2: Check whether labels exist

Labeled inputs and targets → supervised learning
Inputs without target labels → unsupervised learning

Tip 3: Do not confuse classification and regression

Example	Technique
Predict fraud/not fraud	Classification
Predict fraud probability	Classification, often logistic regression
Predict fraud loss amount	Regression
Predict house price	Regression
Predict house price category such as low/medium/high	Classification

Tip 4: Logistic regression is classification

The name is misleading.

Linear regression → continuous numerical output
Logistic regression → probability of a class or event

Tip 5: Distinguish clustering from classification

Predefined labels → classification
Discovering groups → clustering

Tip 6: Distinguish classification from anomaly detection

Known examples of fraud → classification
Unknown unusual behavior → anomaly detection

Tip 7: Remember speech direction

Amazon Transcribe: audio/speech → text
Amazon Polly: text → speech

Tip 8: Remember Rekognition versus Textract

Rekognition: images and videos, objects, faces, unsafe content
Textract: scanned documents, text, handwriting, forms, tables

Tip 9: Do not build a custom model unnecessarily

If a managed, pre-trained service supports the use case, it is often the preferred choice because it reduces:

Development effort
Data requirements
Training effort
Infrastructure management
Operational complexity

Tip 10: Know when SageMaker AI is appropriate

SageMaker AI is generally selected when:

A custom model is required
Custom training is needed
Specialized workflows are needed
The organization needs control over deployment and inference
Prebuilt AI services are insufficient

Tip 11: RAG is not model retraining

RAG retrieves external information and supplies it to the model during a request.

It is useful for:

Current information
Private company data
Enterprise knowledge
Product catalogs
Customer-specific information

Tip 12: Consider explainability and regulation

For a loan approval, medical decision, or other high-impact use case, ask:

Can the decision be explained?
Is human review required?
Is the model biased?
Is the model auditable?
Is a simpler model or rule-based system more appropriate?

Tip 13: More advanced does not always mean more suitable

A foundation model or deep neural network may be unnecessary for:

A simple threshold decision
A basic numerical calculation
A small binary classification task
A deterministic policy
A low-volume process

The best answer may be traditional software, a rule-based system, or a simpler ML model.

11. Final Decision Framework

When given an AI use-case scenario, work through these questions:

What business problem is being solved?
Is the task repetitive, high-volume, pattern-based, or prediction-based?
What type of output is needed?
Category
Number
Group
Anomaly
Generated content
Are labeled examples available?
Is the result required to be deterministic?
Are explainability or regulatory controls important?
Would a rule-based system solve the problem more simply?
Does an AWS managed AI service already support the use case?
Is a custom model required?
Would a traditional ML model or foundation model be more appropriate?
Is external or private information needed?
Would RAG be useful?
Do the expected benefits justify the cost and operational complexity?

The central exam principle is:

Choose the simplest appropriate solution that meets the business, technical, regulatory, cost, and operational requirements.