AWS Certified AI Practitioner Training Course

AIF-C01 is a vocabulary exam before it is anything else. A large share of the questions in this domain do nothing more than describe a system and ask you which category it belongs to, or describe a category and ask which system fits. If you carry the everyday habit of using "AI" and "machine learning" as if they meant the same thing, those questions will cost you marks that require no technical skill at all. So start by fixing the containment relationship in your head.

The nesting order

Artificial intelligence is the outermost term. It covers any technique that lets a computer perform a task normally associated with human intelligence — perceiving, reasoning, deciding, or communicating. That definition is broad on purpose. A hand-written rules engine that decides whether to approve a loan is AI. A chess program built from an explicit search tree and no learned parameters at all is AI. The defining trait is the goal, not the method.

Machine learning is a subset of AI in which the system derives its behavior from data rather than from rules a person wrote. You supply examples, an algorithm adjusts internal parameters to fit patterns in those examples, and the resulting model makes predictions about data it has never seen. The critical difference from the rules engine is that nobody states the rule. If your credit-risk logic was learned from ten years of repayment history rather than written by an underwriter, it is machine learning.

Deep learning is a subset of machine learning that uses neural networks with many layers. The practical significance is what those layers buy you: the network learns its own feature representations instead of relying on a human to engineer them. Classical machine learning on tabular data often works better and cheaper when a domain expert has already decided which columns matter. Deep learning earns its cost on unstructured data — images, audio, natural language — where nobody can enumerate the useful features by hand.

Generative AI is a subset of deep learning. Where a traditional model outputs a label, a number, or a ranking, a generative model outputs new content of the same kind it was trained on: text, images, audio, video, code, or structured data. It is built on very large neural networks — foundation models — pre-trained on enormous general corpora and then adapted to specific tasks rather than trained from scratch per task.

Agentic AI sits on top of generative AI. An agent uses a foundation model as a reasoning engine but adds the ability to take actions: to call tools and APIs, to query data sources, to break a goal into steps, to observe the result of each step, and to decide what to do next. The model that answers a question is generative AI. The system that answers the question, then files the resulting expense report, checks whether it was approved, and emails the requester, is agentic. The 2026 refresh of the AIF-C01 exam guide raised the profile of this category, so expect it by name.

Discriminative versus generative

A second cut is worth internalizing because it explains the boundary between traditional ML and generative AI more precisely than the nesting diagram does. A discriminative model learns the boundary between classes: given this email, is it spam or not. A generative model learns the distribution of the data itself, which is what makes it able to produce plausible new samples from that distribution. Both are machine learning. Only one of them can write you a paragraph.

The words that describe the pieces

Three terms recur throughout the exam and are worth pinning down now.

• A model is the artifact produced by training: the learned parameters plus the structure needed to apply them. It is a file, and it is versioned, stored, and deployed like any other artifact.
• Training is the process that produces the model from data. It is expensive, done in bursts, and does not happen while users are waiting.
• Inference is the process of using a trained model to produce an output for new input. It is comparatively cheap per call but happens continuously, which is why inference — not training — dominates the running cost of most deployed systems.

Where the exam applies this

Expect scenario questions that hand you a business situation and ask for the category. A few patterns repeat:

• A system that flags fraudulent transactions from historical labeled examples is machine learning, and specifically a classification problem. It is not generative AI, and answering "use a foundation model" to that scenario is wrong even though a foundation model could technically be coaxed into it.
• A system that drafts a customer service reply is generative AI.
• A system that reads a support ticket, looks up the customer's order in a database, issues a refund through an API, and confirms it is agentic AI.
• A system that applies a fixed set of thresholds written by a compliance team is AI in the broad sense but not machine learning — no parameters were learned.

The exam rewards precision here because the categories carry different cost, risk, and governance profiles. Generative AI brings hallucination risk and nondeterminism that a classification model does not. Agentic AI brings the additional risk that a wrong decision produces a side effect in a real system rather than merely a wrong sentence on a screen. Those consequences are the subject of later modules, and they only make sense once the categories are separate in your mind.