AIF-C01
Deep Dive
SageMaker is the AWS service for building, training, and deploying your own Machine Learning models, covering the entire lifecycle we saw in the previous section: from the notebook to the production endpoint.
Contents
SageMaker is a fully managed platform that brings together all the tools needed for a traditional ML project: data preparation, notebooks, distributed training, hyperparameter tuning, endpoint deployment, and monitoring — without having to manage individual servers.
When to use SageMaker
Two possible sources for a foundation model
When the starting point is a foundation model (not a classic ML model trained from scratch), there are two paths: use an already-published open source pre-trained model (or proprietary one) — such as those available on Amazon Bedrock or SageMaker JumpStart — and optionally fine-tune it with your own data; or train a custom model from scratch, with SageMaker, when no existing model fits the domain or there are requirements for full ownership over the data and architecture. The first option is almost always faster and cheaper; the second gives full control at the cost of much more effort.
It's SageMaker's unified IDE: a visual, web-based environment that integrates Jupyter notebooks, training pipelines, experiments, model registry, and deployment in a single interface. It replaces the manual work of managing individual notebook instances.
From Studio, a data scientist can explore data, write and run training code, compare experiments (different hyperparameters or algorithms), and deploy the winning model, all without leaving the environment or provisioning infrastructure manually.
Create a managed notebook instance to explore data and train machine learning models.
+1 more...
Autopilot automates the process of building an ML model: given a tabular dataset and a target column, it automatically tries different algorithms, performs feature engineering, and tunes hyperparameters to find the best possible model, without the user writing training code.
Typical use case
A team without deep ML specialists needs a customer churn prediction model from a table of historical data. Autopilot automatically generates several candidate models, shows a performance ranking, and lets you deploy the best one with a few clicks.
Ground Truth helps build high-quality labeled datasets for supervised learning, combining human annotators (your own team, an AWS-managed workforce, or Amazon Mechanical Turk) with ML assistance that learns to automatically label the simpler cases, reducing the total labeling cost.
Example: manually labeling 500,000 images with bounding boxes of objects would be extremely costly. Ground Truth manually labels a subset, trains an auxiliary model on those labels, and that auxiliary model automatically labels the cases where it has high confidence — leaving only the hard cases for human review.
Clarify detects potential bias in training data and model predictions, and provides explainability by showing which features influenced each prediction the most (feature importance). It's a central tool in the exam's responsible AI domain — covered in more depth in the D4 module.
Continuously watches models already deployed in production, comparing the characteristics of real-time input data against those of the training set. It automatically detects:
Data drift
The data reaching the model in production changes its distribution relative to the training data (for example, new patterns in user behavior).
Model quality drift
The model's actual accuracy degrades over time, detectable once real labels are available to compare against predictions.
This continuous monitoring connects directly to phase 6 of the ML lifecycle (monitoring) we saw earlier: when Model Monitor detects significant drift, it usually triggers retraining the model with more recent data.
Configure a Model Monitor schedule and see how data drift silently degrades a production model.
+1 more...
This distinction shows up constantly on the AIF-C01: both are AWS ML/AI platforms, but they solve fundamentally different problems.
Amazon SageMaker
Build your own model from scratch (or fine-tune an existing one) with your own data and full control over the architecture.
More effort, more flexibility, requires ML knowledge.
Amazon Bedrock
Use foundation models already trained by Amazon or third parties, accessible via API, without managing training infrastructure.
Less effort, instant start, ideal for generative AI.
Mnemonic rule
SageMaker = factory (you build the model from the data). Bedrock = store (you pick an already-built model and consume it via API).
In more general production-deployment terms, this same decision boils down to two methods:
Managed API service
You consume a model through an API fully managed by the provider (e.g. Amazon Bedrock). You don't provision or scale infrastructure — you pay per use (tokens, requests) and AWS manages the availability of the underlying hardware.
Self-hosted API
You deploy the model yourself as your own endpoint (e.g. a SageMaker endpoint), choosing the instance type, auto-scaling configuration, and update cycle. More control and potential savings at large scale, in exchange for taking on the operation of that infrastructure.
Besides SageMaker AI (training/deploying your own models) and Bedrock (consuming foundation models via API), AWS offers other services that show up at different stages of an AI pipeline — from building the solution to a final user consuming it.
Amazon Q
AWS's family of generative AI assistants (Amazon Q Developer, Amazon Q Business). In 2026 both closed to new customers: Q Developer evolves into Kiro and Q Business into Amazon Quick. Stage: consumption / end product — a user interacts with Q in natural language to write code, resolve questions about corporate data, or automate tasks, without building their own model.
Amazon QuickSight ("Amazon Quick")
Business intelligence service with built-in generative AI capabilities (summaries, natural-language questions about dashboards). Stage: analysis and consumption of results by business users, not ML/training.
Kiro
AWS's agentic IDE: a development environment that uses generative AI and agents to help specify, generate, and maintain code. Stage: build — used during development of the solution itself, not in production.
SageMaker AI and Bedrock
Already covered in detail: SageMaker covers training and deployment of your own models; Bedrock covers access to managed foundation models via API. They're the core of the pipeline's training and inference stages.
AWS managed AI services
For specific, pre-trained AI tasks (without building or training anything), AWS offers ready-to-use APIs: Amazon Transcribe (speech to text), Amazon Polly (text to speech), Amazon Comprehend (NLP — sentiment, entities), Amazon Lex (chatbots with intents), and Amazon Translate (automatic translation between languages). Each one is covered in more depth in the D3 AI services module.
Not every AI problem needs a foundation model. The exam expects you to recognize the criteria that tip the decision toward SageMaker (traditional/custom ML) or toward Bedrock (foundation models).
Regulatory requirements
Heavily regulated sectors (banking, healthcare) may require auditable models with fully controlled data provenance — easier to guarantee with a proprietary model trained on SageMaker.
Need for explainability
If the business needs to justify each individual prediction (e.g. why a loan was denied), an interpretable classic ML model is usually preferable to a "black box" foundation model.
Operational constraints
Budget, maximum acceptable latency, availability of proprietary training data, and time-to-production also determine whether it makes sense to build (SageMaker) or consume (Bedrock).
Did you get this topic?
Test what you just learned
A manufacturing company has a proprietary dataset from its machine sensors and wants to train a failure-prediction model specific to its industrial equipment, with full control over the architecture and hyperparameters. Which AWS service is the best fit?
Put what you learned to the test with original AIF-C01 exam questions.
Start AIF-C01 mock examUnlimited mock exams with the Pro plan
Log in to track your progress.