Visual overview
SageMaker AI supports end-to-end ML work, while Amazon Bedrock provides managed access to foundation models for generative and agentic applications.
Technical reference
Use the output type, data, and acceptable uncertainty to choose the technique before selecting an AWS service.
Examples include fraud/not-fraud, document category, or defect class; typically learned from labeled examples.
Examples include demand, price, duration, or temperature forecasts when the target is numerical.
Groups similar observations without a provided target label; useful for discovery and segmentation.
Foundation models can generate text or other modalities; output is probabilistic and requires evaluation for the intended use case.
The application supplies tools, identity, state, safety controls, and termination logic around model reasoning.
AI is the umbrella; ML is one way to build it
Artificial intelligence (AI) is the broad idea of building systems that perform tasks associated with intelligent behavior, such as perception, language understanding, prediction, planning, and decision support. Machine learning (ML) is a major approach within AI: instead of encoding every decision as an explicit rule, an algorithm learns patterns from example data and produces a model. Training is the process that fits model parameters from data; inference is the use of the trained model to produce a prediction or other output for new input.
Deep learning is a family of ML techniques based on multilayer neural networks. It is especially important for unstructured data such as text, audio, images, and video, where hand-designing useful features can be difficult. AI, ML, and deep learning are therefore not competing labels: deep learning is part of ML, and ML is part of the broader AI landscape. A deterministic rules engine can still be an AI-adjacent solution even when no model is trained.
Predictive ML and generative AI solve different classes of problems
Traditional predictive ML commonly maps input features to a class, number, ranking, anomaly score, or cluster. Examples include classifying fraudulent transactions, forecasting demand, predicting equipment failure, or grouping customers. Generative AI instead creates new content—such as text, images, audio, video, or code—based on patterns learned during training. Foundation models (FMs) are large, broadly trained models that can support many downstream tasks and can often be adapted through prompting, retrieval, or customization rather than training a model from scratch.
The correct approach follows the business output. If an application needs a stable numeric probability with clear tabular features, a specialized supervised model may be more appropriate and cheaper to operate than a large language model. If it must summarize documents, answer open-ended questions, generate text, reason over natural-language instructions, or work across several tasks, an FM can offer a much better starting point. Hybrid systems are common: a generative application can call a predictive model or conventional service as a tool.
Agentic AI adds a decision-and-action loop
An agentic system does more than produce a single completion. It can use a model to interpret a goal, choose a next step, call tools or APIs, inspect results, maintain relevant context or memory, and repeat until a stopping condition is met. This introduces an orchestration loop around the model. On AWS, current agent-building options include Amazon Bedrock AgentCore as well as frameworks such as Strands Agents, while Bedrock models can also perform direct tool use through supported inference APIs.
More autonomy creates more responsibility. Every tool expands what a model-driven workflow can affect, so agents need explicit authorization boundaries, validation of tool inputs and outputs, observable traces, retry and timeout behavior, and human approval for consequential actions where appropriate. Treat the foundation model as a probabilistic reasoning component inside a larger software system—not as the system's identity, permissions model, source of truth, or audit log.
Key takeaways
- 01
AI is the broad field; ML learns models from data; deep learning is a neural-network-based subset of ML.
- 02
Training fits a model from examples, while inference uses the model on new input.
- 03
Predictive ML returns predictions, scores, or groupings; generative AI creates new content and can support many natural-language tasks.
- 04
Foundation models are broadly trained models that can be adapted rather than rebuilt for every task.
- 05
Agentic AI wraps models in a loop that can plan, use tools, observe outcomes, and continue toward a goal.
Official AWS sources
Use these primary AWS resources for the source material behind this article and for deeper reference.