AWS AI Services: Bedrock, SageMaker AI, and Pre-Trained APIs

AWS offers several kinds of AI capability. The right choice depends on whether you want to call a ready-made AI API, build a generative AI application with a foundation model, train a custom machine-learning model, or provide the compute infrastructure for model workloads.

This is Part 62 of the AWS Cloud Practitioner Study Notes. The broad map is:

Generative AI applications → Amazon Bedrock, Amazon Q
Custom ML lifecycle        → Amazon SageMaker AI
Pre-trained AI APIs        → Rekognition, Textract, Comprehend, and more
AI compute                 → Trainium, Inferentia, GPUs, and ML infrastructure

For the Cloud Practitioner exam, the most important decision is usually: use Bedrock for foundation-model applications, SageMaker AI for building and deploying ML models, and purpose-built AI services for a specific input type or task.

Amazon Bedrock: build with foundation models

Amazon Bedrock is a fully managed, serverless service for building generative AI applications with foundation models from Amazon and third-party providers through AWS APIs.

Application
      ↓ prompt or API request
Amazon Bedrock
      ↓ selects a foundation model
Text, image, embedding, or structured response

Typical use cases include:

• Chatbots and customer assistants
• Summarisation and content generation
• Retrieval-Augmented Generation (RAG)
• Embeddings and semantic search
• Agents that use enterprise systems and APIs
• Model evaluation, customisation, and guardrails

Bedrock is useful when a team wants to add generative AI without operating model-training clusters or hosting foundation models itself. It still requires application engineering: prompt design, retrieval quality, access control, evaluation, cost controls, safety filters, and output validation remain your responsibilities.

Example:

Physiotherapy booking application
        ↓ user asks a question
Amazon Bedrock
        ↓ knowledge base and foundation model
Grounded assistant response

Bedrock can provide access to multiple models through a common AWS experience, but model availability, features, Regions, pricing, and quotas vary. Choose a model based on the task and evaluate its output rather than assuming that one model is always best.

Amazon SageMaker AI: build, train, and deploy ML models

Amazon SageMaker AI is a fully managed machine-learning service for building, training, evaluating, and deploying models. The service was renamed from Amazon SageMaker to Amazon SageMaker AI in December 2024; APIs, CLI namespaces, and many existing resource names still use sagemaker for compatibility.

Training data
      ↓
Prepare and label data
      ↓
SageMaker AI training job
      ↓
Evaluate model
      ↓
Deploy endpoint or batch inference
      ↓
Monitor and retrain

Use SageMaker AI when you have your own dataset, training script, pretrained model, or model-lifecycle requirements. Examples include:

• Fraud or anomaly prediction
• Demand forecasting
• Medical-image classification
• Recommendation models
• Custom regression or classification
• Fine-tuning or deploying a model with specific infrastructure needs

SageMaker AI provides managed training jobs, notebooks and development environments, built-in algorithms, model deployment options, monitoring, pipelines, and tools such as JumpStart and Canvas. It does not mean that AWS automatically solves data quality, feature design, model evaluation, bias, drift, or production governance.

Bedrock versus SageMaker AI

Question	Amazon Bedrock	Amazon SageMaker AI
Main goal	Build generative AI applications	Build and operate ML models
Starting point	Foundation models available through managed APIs	Your dataset, algorithm, pretrained model, or training code
Typical output	Generated text, images, embeddings, or agent actions	Predictions or inferences from a deployed model
Infrastructure management	Less model infrastructure to operate	More control over training and serving infrastructure
Typical user	Application developers and product teams	Data scientists and ML engineers
Exam shortcut	“I want a ChatGPT-like application”	“I want to train or deploy my own model”

The boundary is not absolute. Bedrock supports model customisation and RAG, and SageMaker AI can work with foundation models. The decision depends on how much control the team needs over data, training, runtime, model weights, networking, and deployment.

Amazon Q: AWS and enterprise assistants

Amazon Q is a family of generative AI assistants built on Amazon Bedrock.

Amazon Q Developer

Amazon Q Developer helps developers and AWS users with:

• Code explanation and generation
• Test generation and debugging assistance
• AWS architecture and troubleshooting questions
• Code transformation and modernisation workflows
• Assistance inside IDEs and AWS tooling

Developer in an IDE or AWS console
          ↓
Amazon Q Developer
          ↓
Code explanation, suggestions, tests, or AWS guidance

Q Developer is an AI assistant product for developers and cloud operators; it is not the same as giving an application direct access to a foundation model through Bedrock.

Amazon Q Business and current availability

Amazon Q Business is an enterprise assistant designed to answer questions and help employees use information from organisational data sources. However, AWS documentation states that Q Business is no longer open to new customers starting 31 July 2026 and points new customers towards Amazon Quick for similar capabilities. Existing Q Business customers and exam materials may still reference the service, so treat it as an important concept while checking current availability before choosing it for a new project.

Amazon Quick is an AI-powered workspace that combines business intelligence, AI-assisted chat, workflow automation, and research capabilities. This availability change is a good example of why current AWS service documentation matters more than memorising an old service list.

Pre-trained AI services

AWS also provides specialised APIs for common AI tasks. You send the service an image, document, audio file, text, or user interaction and receive an analysis or transformation without training a model yourself.

Amazon Rekognition: image and video analysis

Rekognition provides computer-vision APIs for images and videos. It can detect objects, scenes, faces, text, unsafe content, celebrities, and other visual attributes, with supported face comparison, liveness, collections, and custom-label capabilities.

Image or video
      ↓
Amazon Rekognition
      ↓
Objects, faces, text, labels, or moderation result

Common uses include media cataloguing, content moderation, identity verification, and visual search. Use appropriate consent, privacy, retention, and human-review controls for biometric or sensitive applications.

Amazon Textract: documents, forms, and tables

Textract extracts text and structure from documents. It can identify forms as key-value pairs, tables and cells, signatures, and specialised information from invoices, receipts, and identity documents.

Invoice or form
      ↓
Amazon Textract
      ↓
Invoice number, fields, table cells, and values

The exam distinction is: ordinary OCR reads characters; Textract analyses document structure such as forms and tables.

Amazon Comprehend: natural-language analysis

Comprehend analyses text using natural-language processing. It can identify sentiment, language, entities, key phrases, topics, and other supported text properties.

Customer review
      ↓
Amazon Comprehend
      ↓
Sentiment, entities, language, and key phrases

It is useful for ticket classification, review analysis, entity extraction, and content triage.

Amazon Lex: conversational interfaces

Lex builds conversational interfaces using text or voice. It identifies user intent and extracts slots or parameters from a conversation:

“I want to book an appointment tomorrow.”
      ↓
Amazon Lex
      ↓
Intent: BookAppointment
Slots: date, time, service

Lex is a managed chatbot and conversational-interface service. Bedrock is the broader foundation-model service for generative responses and agentic application patterns; the two are not interchangeable.

Amazon Transcribe, Polly, and Translate

Service	Input and output	Example
Amazon Transcribe	Speech → text	Convert a meeting recording into a transcript
Amazon Polly	Text → speech	Read an article using a synthetic voice
Amazon Translate	Text in one language → translated text	Translate English support content into Chinese

Transcribe supports batch and streaming transcription patterns. Polly is text-to-speech. Translate is machine translation. These services can be composed into voice assistants, accessibility tools, multilingual support, and media workflows.

Amazon Personalize: recommendations

Amazon Personalize builds personalised recommendations from user interaction data and item metadata:

User activity and catalogue data
      ↓
Amazon Personalize
      ↓
Recommended items or ranked results

It is the exam answer for a managed recommendation engine, such as “customers who viewed this also viewed…” or personalised content suggestions.

Amazon Kendra: enterprise search

Amazon Kendra is an intelligent enterprise-search service that indexes connected organisational content and returns relevant answers or documents for natural-language queries.

Company documents and data sources
      ↓
Amazon Kendra index
      ↓
Employee asks a natural-language question
      ↓
Relevant enterprise answer or documents

For newer generative enterprise assistants, compare Kendra, Amazon Q, Amazon Quick, and a Bedrock Knowledge Base according to the current product capabilities and data-source requirements.

Amazon Augmented AI (A2I): human review

Amazon Augmented AI (A2I), a SageMaker capability, creates human-review workflows for ML predictions when confidence is low or a business rule requires review:

AI prediction
      ↓ confidence below threshold
A2I human review
      ↓
Approved, corrected, or rejected result

Human review is important for sensitive documents, identity decisions, medical workflows, financial decisions, and any use case where an automated error has serious consequences.

AI infrastructure

AI workloads need compute for training and inference. AWS provides purpose-built chips and GPU-backed EC2 or managed-service options:

Infrastructure	Main role
AWS Trainium	Purpose-built accelerator for training machine-learning and foundation models
AWS Inferentia	Purpose-built accelerator for efficient model inference
GPU-based instances	Flexible training and inference for deep-learning workloads
SageMaker AI infrastructure	Managed environments for training, hosting, and inference

The correct accelerator depends on model framework support, performance, memory, cost, availability, and the operational environment. Infrastructure choice is a deeper engineering decision than the Cloud Practitioner service-identification questions.

Putting AI services together

AI services are often composed rather than used in isolation:

Invoice uploaded to S3
      ↓
Textract extracts fields and tables
      ↓
Comprehend classifies text or entities
      ↓
Bedrock summarises the structured result
      ↓
Human review handles low-confidence cases

Another example is a multilingual voice assistant:

User speaks
      ↓
Transcribe
      ↓
Lex or Bedrock understands the request
      ↓
Translate when required
      ↓
Polly returns spoken audio

Every composition needs security and responsible-AI design: IAM permissions, encryption, sensitive-data handling, prompt and output controls, human review, auditability, model evaluation, and cost monitoring.

Choosing the service

Requirement	Service to remember
Build a generative AI application with foundation models	Amazon Bedrock
Train, tune, host, and monitor custom ML models	Amazon SageMaker AI
AI coding and AWS development assistant	Amazon Q Developer
Image or video understanding	Amazon Rekognition
Extract fields and tables from documents	Amazon Textract
Sentiment, entities, topics, or key phrases	Amazon Comprehend
Conversational chatbot with intents and slots	Amazon Lex
Speech to text	Amazon Transcribe
Text to speech	Amazon Polly
Language translation	Amazon Translate
Personalised recommendations	Amazon Personalize
Intelligent enterprise search	Amazon Kendra; compare current Amazon Q or Quick options
Human review of low-confidence predictions	Amazon A2I
Model training accelerator	AWS Trainium
Model inference accelerator	AWS Inferentia

Common exam questions

You want to build a ChatGPT-like application using foundation models.

Use Amazon Bedrock.

You have a dataset and want to train and deploy a custom prediction model.

Use Amazon SageMaker AI.

You need to extract invoice fields and tables.

Use Amazon Textract.

You need to analyse customer-review sentiment.

Use Amazon Comprehend.

You need to recognise objects or faces in images.

Use Amazon Rekognition.

You need a text or voice chatbot based on intents.

Use Amazon Lex.

You need speech-to-text or text-to-speech.

Use Amazon Transcribe or Amazon Polly respectively.

You need product recommendations.

Use Amazon Personalize.

You need a human to review low-confidence ML predictions.

Use Amazon Augmented AI.

Final memory map

Generative AI
→ Bedrock, Amazon Q

Custom ML lifecycle
→ SageMaker AI

Vision and documents
→ Rekognition, Textract

Language and conversation
→ Comprehend, Lex, Translate

Speech
→ Transcribe, Polly

Recommendations and search
→ Personalize, Kendra

Human review and infrastructure
→ A2I, Trainium, Inferentia, GPUs

The one-sentence takeaway is: AWS provides managed AI APIs for specific tasks, Bedrock for foundation-model applications, SageMaker AI for custom ML lifecycles, Amazon Q for assistants, and purpose-built infrastructure for demanding training and inference workloads.

Sources

• Amazon Bedrock documentation overview
• Using foundation models with Amazon Bedrock
• What is Amazon SageMaker AI?
• SageMaker AI machine-learning workflow
• Amazon Q Developer documentation
• Amazon Quick documentation
• Amazon Rekognition
• Amazon Textract
• Amazon Quick and Q Business availability note