Choosing the Right Service & Building Responsibly
In Lessons 1-7, you learned how AI and ML work: from classic ML vs deep learning, to neural networks, LLM training, prompt engineering, RAG, fine-tuning, and responsible AI principles.
Now the question is: how do you actually build and deploy these solutions? AWS provides a layered set of AI/ML services so you can pick the right tool for the job, whether you need a quick pre-built API or full control over custom model training.
AWS organizes its AI/ML services into three tiers. The higher you go, the more control you get, but also more complexity and cost. Most teams start at the top (pre-built) and move down only when needed.
No ML expertise needed. Call an API, get results.
Access leading foundation models via API. Customize with your data.
Full control over data, training, and deployment. For ML teams.
Start at Tier 1 (pre-built APIs) whenever possible. Move to Tier 2 (Bedrock) for generative AI use cases. Only go to Tier 3 (SageMaker) when you need full custom model training or have unique requirements that the managed services can't handle.
These services are pre-trained by AWS. You don't need any ML expertise. Just send data to an API and get results back. They cover the most common AI tasks: vision, language, speech, and recommendations.
Analyzes images and videos using ML. Detects objects, people, scenes, and text in images. Also offers facial analysis, content moderation, and custom label detection.
Use when: You need image/video analysis, content moderation, or facial analysis without building custom computer vision models.
Automatically extracts printed text, handwriting, and structured data (tables, forms) from scanned documents and images. Goes beyond simple OCR by understanding document layout.
Use when: You need to extract text and data from invoices, receipts, ID documents, or any scanned paperwork.
Natural language processing (NLP) service that finds insights in text. Detects sentiment, entities (people, places, brands), key phrases, and language. Also supports topic modeling and custom classification.
Use when: You need sentiment analysis, entity extraction, or text classification on customer reviews, support tickets, or documents.
Neural machine translation service that delivers fast, high-quality language translation. Supports thousands of language pairs and allows custom terminology for domain-specific translations.
Use when: You need to translate text between languages at scale, such as localizing content or translating customer communications.
Converts speech to text using automatic speech recognition (ASR). Supports real-time and batch transcription with speaker identification and custom vocabulary.
Use when: You need to transcribe call center recordings, meetings, or media content.
Turns text into lifelike speech using deep learning. Offers dozens of voices across many languages, including neural text-to-speech voices for natural-sounding output.
Use when: You need text-to-speech for accessibility, voice assistants, or audio content generation.
Build conversational interfaces (chatbots and voice bots) using the same technology that powers Alexa. Provides natural language understanding and automatic speech recognition.
Use when: You need to build a chatbot or voice bot for customer service, order tracking, or FAQ automation.
Creates real-time personalized recommendations using the same ML technology used on Amazon.com. Supports product recommendations, personalized search, and customized marketing.
Use when: You need product recommendations, personalized content feeds, or targeted marketing campaigns.
Intelligent enterprise search service powered by ML. Understands natural language queries and returns precise answers from across your document repositories, not just keyword matches.
Use when: You need intelligent search across internal documents, wikis, FAQs, and knowledge bases.
Amazon Bedrock is a fully managed service that gives you API access to leading foundation models (FMs) from multiple providers. You can build generative AI applications without managing infrastructure or training models from scratch.
Key benefit: Choose from multiple model providers through a single, unified API. Your data is not used to train the base models, and it stays within your AWS account.
Bedrock offers foundation models from multiple leading AI companies, including:
The list of available providers and models continues to grow. Check the AWS documentation for the latest list.
Fully managed RAG (Retrieval Augmented Generation) capability. Connect your data sources — documents, databases, web pages — and Bedrock handles chunking, embedding, storage, and retrieval automatically.
Remember Lesson 6? RAG gives models access to your private data. Bedrock Knowledge Bases implements the entire RAG pipeline as a managed service.
Build AI agents that can plan and execute multi-step tasks. Agents can call APIs, query databases, and take actions on behalf of users. Bedrock also supports multi-agent collaboration for complex workflows.
Example: An agent that handles a customer return by looking up the order, checking the return policy, and initiating the refund — all autonomously.
Configurable safeguards for your GenAI applications. Define policies to filter harmful content, block sensitive information (PII), prevent hallucinations, and enforce topic boundaries. Works across any FM on Bedrock and even models hosted outside Bedrock.
Key feature: Includes Automated Reasoning to help prevent factual errors from hallucinations in RAG and summarization use cases.
Fine-tune foundation models with your own data directly in Bedrock. Supports continued pre-training and fine-tuning while keeping your data private and secure within your AWS account.
Remember Lesson 6? Fine-tuning teaches models specialized behaviors. Bedrock lets you do this without managing training infrastructure.
Amazon SageMaker AI is a comprehensive platform for building, training, and deploying custom machine learning models. It covers the entire ML lifecycle: data preparation, model training, tuning, deployment, and monitoring.
Use SageMaker when pre-built services and foundation models don't meet your needs — for example, when you have proprietary data and need a custom model architecture, or when you need to train or fine-tune models at massive scale.
Train custom models using built-in algorithms, your own code, or popular frameworks (PyTorch, TensorFlow). SageMaker manages the infrastructure, scaling, and distributed training.
Deploy models with real-time endpoints, batch inference, async inference, or serverless inference. SageMaker handles auto-scaling and infrastructure management.
Purpose-built infrastructure for large-scale model training. Scales across thousands of AI accelerators and can reduce foundation model training and fine-tuning costs by up to 40%.
SageMaker Clarify helps you detect potential bias in your training data and model predictions, and explains how models make decisions. Key capabilities:
Clarify supports guidelines such as ISO 42001 for AI management systems.
Teams often ask: "Should I use Bedrock or SageMaker?" The answer depends on your use case, team expertise, and how much control you need.
| Aspect | 🔵 Amazon Bedrock | 🔴 Amazon SageMaker AI |
|---|---|---|
| Primary Use | Generative AI apps using foundation models | Custom ML model training and deployment |
| ML Expertise Needed | Low to medium | Medium to high |
| Model Choice | Select from leading FMs (Claude, Llama, Nova, etc.) | Bring your own model or use built-in algorithms |
| Infrastructure | Fully managed, serverless | Managed but you configure instances and scaling |
| Customization | Fine-tuning, RAG, prompt engineering | Full control: architecture, training, hyperparameters |
| Best For | Chatbots, content generation, summarization, RAG | Fraud detection, forecasting, custom NLP, recommendation engines |
| Time to Production | Hours to days | Weeks to months |
| Pricing | Pay per token/request | Pay for compute instances used |
Many organizations use both. For example: Bedrock for a customer-facing chatbot with RAG, and SageMaker for a custom fraud detection model trained on proprietary transaction data. The right choice depends on the specific problem you're solving.
Click on a business scenario below to see which AWS service is the best fit and why.
Walk through these questions to determine which tier of AWS AI/ML services fits your use case.
Is your task a common AI pattern?
(image analysis, text extraction, speech-to-text, translation, sentiment analysis)
Do you need generative AI capabilities?
(text generation, summarization, chatbot, code generation, RAG)
Do you need a custom-trained model?
(proprietary data, custom architecture, specialized predictions, full ML pipeline control)
In Lesson 7, you learned about ethical AI principles: transparency, fairness, accountability, privacy, and safety. AWS provides concrete tools and frameworks to put these principles into practice.
AWS defines eight dimensions of responsible AI to guide organizations throughout the AI lifecycle:
AI systems should not discriminate against individuals or groups
Understand and explain how AI systems make decisions
Protect data and ensure secure AI operations
Prevent AI systems from causing harm
Maintain human oversight and control over AI
Ensure AI outputs are accurate and reliable
Establish policies and processes for AI management
Be open about AI capabilities and limitations
What it does: Configurable safeguards that evaluate user inputs and model responses based on your policies. Provides content filtering, PII redaction, topic restrictions, and hallucination prevention.
Responsible AI dimensions: Safety, controllability, veracity
Works across all FMs on Bedrock and can also be applied to models hosted outside Bedrock via the ApplyGuardrail API.
What it does: Detects bias in training data and model predictions. Explains model decisions using feature attribution. Monitors deployed models for bias drift over time.
Responsible AI dimensions: Fairness, explainability, governance
Helps build less biased and more understandable ML models throughout the entire lifecycle.
What it does: A structured framework for integrating ethics, transparency, and risk management into AI systems. Provides best practices, design principles, and review questions for responsible AI workloads.
Responsible AI dimensions: All eight dimensions
Emphasizes identifying and resolving potential issues as early as possible in the AI lifecycle.
What it does: AWS supports human review workflows where AI predictions are reviewed by humans before taking action. Services like Amazon Augmented AI (A2I) enable this pattern.
Responsible AI dimensions: Controllability, accountability, safety
Critical for high-stakes decisions in healthcare, finance, and legal domains.
Here's how everything you've learned fits together on AWS:
| Concept (Lesson) | What You Learned | AWS Service |
|---|---|---|
| Classic ML vs Deep Learning (L1) | When to use each approach | SageMaker AI (custom models), Pre-built APIs (common tasks) |
| Feature Engineering (L2) | Preparing data for models | SageMaker Data Wrangler, SageMaker Processing |
| Neural Networks (L3) | How networks learn | SageMaker Training, SageMaker HyperPod |
| Language Models (L4) | Tokenization, embeddings, transformers | Amazon Bedrock (access to leading FMs) |
| LLM Training (L5) | Training process and mechanics | Bedrock Model Customization, SageMaker HyperPod |
| Working with LLMs (L6) | Prompting, RAG, fine-tuning | Bedrock (prompting), Bedrock Knowledge Bases (RAG), Bedrock fine-tuning |
| Ethics & Responsibility (L7) | Responsible AI principles | Bedrock Guardrails, SageMaker Clarify, Well-Architected AI Lens |
| AI on AWS (L8 — this lesson) | Choosing the right service | Pre-built APIs → Bedrock → SageMaker (pick your tier) |
You've now covered the full journey: from understanding how AI/ML works (Lessons 1-5), to practical techniques for using it (Lessons 6-7), to choosing the right AWS services and building responsibly (Lesson 8).
You can describe AI, ML, and GenAI concepts. You can identify the right technology for a business problem. You can determine which AWS service fits a specific use case. And you know how to use these technologies responsibly.