AWS revised the AWS Certified AI Practitioner AIF-C01 guide in 2026 to explicitly include agentic AI alongside artificial intelligence, machine learning, deep learning, large language models, and generative AI. These categories overlap, but they are not interchangeable. A useful mental model starts broad and then asks what kind of capability is being built, how it learns or reasons, what data it uses, and whether it only produces output or can also choose and execute actions.
Understanding the hierarchy prevents two common mistakes: treating every AI system as a generative model, and treating every application that calls an LLM as an autonomous agent. AWS services evolve quickly, but the conceptual distinctions remain useful across vendors and business use cases.
Artificial intelligence is the broadest category
Artificial intelligence describes systems designed to perform tasks associated with intelligent behavior, such as perception, language understanding, planning, prediction, search, recommendation, or decision support. AI does not require machine learning. Rule systems, search algorithms, optimization, and knowledge-based logic can all be part of AI solutions.
Start by defining the business task rather than selecting a model. Separate automation from intelligence: a deterministic workflow can call APIs, move files, or approve requests without using AI. Adding an AI component is justified when prediction, language understanding, perception, generation, or adaptive decision-making is actually required. If deterministic business rules solve the problem reliably, adding a probabilistic model can increase cost and complexity without adding value. AI is an umbrella concept, not a requirement to use the newest model class. Simpler designs remain valid when they meet the requirement with lower risk and cost.
Machine learning learns patterns from data
Machine learning is a subset of AI in which models learn statistical patterns from data instead of having every decision rule written explicitly. Supervised learning uses labeled examples, unsupervised learning looks for structure in unlabeled data, and reinforcement learning optimizes behavior from rewards and feedback.
Examples include fraud scoring, demand forecasting, image classification, anomaly detection, and recommendations. Model quality is measured against task-appropriate metrics and representative data. A model can be statistically strong and still fail the business if predictions arrive too late, cost too much, or perform poorly for an important subgroup. The overview of unsupervised machine learning shows one branch of this family. ML output is usually a prediction, score, class, embedding, or decision signal rather than newly generated human-like content.
Deep learning uses multi-layer neural networks
Deep learning is a subset of machine learning built around neural networks with many layers. It is especially important for image, speech, text, and other high-dimensional data. Modern foundation models and large language models depend heavily on deep-learning architectures.
Deep learning often needs substantial data and compute during training, but inference can be delivered through managed endpoints or foundation-model APIs. The distinction also affects governance: training may involve large datasets and model artifacts, while inference may expose prompts, user inputs, embeddings, and predictions that require separate privacy and retention controls. The AI Practitioner role does not require candidates to implement neural networks; it requires a correct understanding of where deep learning sits in the AI hierarchy and what types of workloads it enables. Training a model, adapting a foundation model, and consuming a managed model API are different operational choices with different data, cost, and governance responsibilities.
Generative AI creates new content from learned patterns
Generative AI produces new text, images, audio, code, video, or other content based on patterns learned during training. Large language models are one important class of generative model, but generative AI is broader than text. The model predicts or constructs an output conditioned on prompts and context rather than simply classifying an input.
Generation is probabilistic. Temperature, decoding strategy, prompt wording, retrieved context, and model version can all change output. Evaluation therefore needs test sets and quality criteria instead of one impressive demo. For higher-risk use cases, the application should have fallback or human-review behavior when output is uncertain or consequential. The same prompt may produce different valid outputs, and fluent output can still be factually wrong. Application design therefore needs evaluation, grounding, guardrails, and user expectations that match the consequences of an incorrect answer.
Foundation models are reusable models trained for broad adaptation
Foundation models are trained on broad data and can be adapted to many downstream tasks through prompting, retrieval, fine-tuning, or other techniques. They can be proprietary or openly available, and they may differ substantially in supported languages, modalities, licensing, safety behavior, and deployment options. Those differences matter to business fit even when benchmark scores look similar. This differs from a narrow model trained only for one prediction target. Foundation models can support text generation, embeddings, summarization, classification, multimodal tasks, and tool-use patterns.
Model choice should depend on modality, quality, context capacity, latency, cost, governance, and deployment needs. Context window describes how much input a model can consider in one request, but a larger context window does not guarantee better reasoning. Irrelevant or conflicting context can reduce quality and increase cost. A larger model is not automatically better. Simpler or smaller models can be faster, cheaper, easier to govern, and accurate enough for a well-bounded task.
Agentic AI combines models with goals, tools, and action loops
Agentic AI describes systems that can pursue a goal through multiple steps, decide what action to take next, use tools or external systems, observe results, and continue until a stopping condition is reached. An agent may use an LLM for reasoning or planning, but the defining feature is the action loop rather than text generation itself.
An LLM chatbot that answers one prompt is not automatically an agent. Agent design introduces state, planning, tool permissions, memory, stopping criteria, and error recovery. The security boundary shifts from “what can the model say?” to “what can the system do?” because a tool-enabled agent may create tickets, modify data, send messages, or operate cloud services. A workflow that selects a tool, calls an API, evaluates the response, updates a plan, and performs another action is more agentic. Greater autonomy creates new governance needs because the system can cause external effects rather than only produce information. Tool permissions should be scoped to the task, high-impact actions should require confirmation where appropriate, and the agent should have bounded retries and stopping conditions so an error does not become a loop of repeated external changes.
Inference is the use of a trained model
Training builds or adapts a model; inference uses the trained model to produce predictions or generated output. AWS AI Practitioner objectives include several inference patterns, such as batch, real-time, asynchronous, and serverless. The right pattern depends on latency, throughput, payload size, cost, and whether the user waits for the result.
Do not confuse inference with learning. Batch inference is suitable when large datasets can be processed without immediate response, real-time inference supports interactive or low-latency decisions, asynchronous inference handles longer jobs without holding a client connection, and serverless patterns trade persistent capacity for managed scaling. Many deployed systems perform inference repeatedly without changing model weights. Feedback may be collected for later evaluation or retraining, but that is a separate lifecycle step.
AI categories can be combined in one business solution
A customer-support system might use conventional ML to route tickets, a generative model to draft responses, retrieval to supply current policy content, and an agent to update a case through approved tools. Each component should have its own success metric: routing accuracy, grounded-answer quality, response latency, tool-action correctness, and business outcome should not be collapsed into one vague “AI quality” score. The components solve different problems and should be evaluated separately.
Data governance also crosses the categories. The guidance on AI data privacy applies whether the workload is predictive, generative, or agentic. Training data, prompts, retrieved documents, model output, tool inputs, and logs can all contain sensitive information.
Choose the simplest AI approach that meets the requirement
Business discussions often jump from a problem directly to generative or agentic AI. Include an operational ownership question before deployment: who monitors quality, who responds to harmful or incorrect output, who approves model changes, and who can disable tool access? Systems with more autonomy require clearer human accountability, not less. A better sequence is to define the outcome, identify the data and risk, determine whether rules or traditional ML are sufficient, and then consider generative or agentic capabilities where they add measurable value. The risks and challenges of AI become more important as systems become less deterministic and more autonomous.
AI, ML, deep learning, generative AI, foundation models, and agentic AI form a related family, not a maturity ladder that every application must climb. A well-designed system may deliberately use a simple classifier for one decision and a generative model for another. The right architecture combines methods according to evidence, risk, and economics rather than novelty. Accurate terminology helps architects select appropriate technology, set realistic expectations, and apply controls that match what the system can actually do.
Keep the model separate from the system around it. An agentic application may use a generative model for reasoning while an orchestrator manages memory, tool permissions, retries, and stopping conditions. That does not make every model call “agentic AI.” The useful classification is based on what the system can observe and do: prediction, generation, retrieval, or goal-directed action. Clear boundaries make evaluation and risk ownership much easier.