Core AI Concepts: AI, ML, GenAI, Training, and Inference

A funding request that labels every capability “AI” leaves finance unable to judge what work is being proposed. For the complaint pilot, we need precise language before we can decide whether the complaint backlog needs a system that classifies cases, drafts responses, or does something else entirely.

An algorithm is a method for finding patterns or making decisions. A model is the learned result produced when an algorithm uses relevant historical data. These terms matter because a business buys or adopts a model's capability, not an abstract promise that AI will solve a workflow.

AI, ML, and GenAI boundaries

Artificial intelligence (AI) is the broad field of systems that perform tasks associated with human intelligence, such as interpreting language, recognizing images, making recommendations, or supporting decisions. Common business uses include recommendation engines, natural language processing, computer vision, document extraction, and customer operations support.

Machine learning (ML) is a subset of AI in which systems learn patterns from data rather than relying only on rules written in advance. Deep learning is a subset of ML that uses layered neural networks to handle complex patterns, especially in areas such as language and images. Not every AI solution uses ML, and not every ML solution creates content.

Generative AI (GenAI) is a subset of deep learning that creates new content based on patterns it learned. It can generate text, images, code, summaries, or drafts. A complaint-response assistant that prepares a customer email is a GenAI use case. A model that assigns each complaint to a category is usually predictive ML, even though both capabilities fall under the broad AI label.

Inspect the nested relationship before attaching a solution label to a business requirement.

Training, inference, and predictions

Training is the process in which a model learns patterns from historical examples. In the complaint pilot, the organization could use past complaints that already have known categories such as billing, delivery, product quality, or account access. The model examines the relationship between complaint details and those historical labels.

Inference is the use of a trained model on a new input. When a new complaint arrives, inference applies the learned patterns to that complaint. A prediction is the resulting output, such as a billing category, a likelihood that the complaint will escalate, or an estimated resolution time.

This distinction prevents a common business mistake. Training happens before the model is used for new work, while inference happens as the model processes new work. A model does not independently acquire reliable knowledge simply because time passes. It reflects the historical information and patterns used in training, subject to later updates and oversight.

For the complaint pilot, predictive ML could classify an incoming complaint, forecast likely backlog volume, or score the risk that a case will be reopened. GenAI could draft a reply, summarize a long complaint thread, or answer a service representative's question using provided context. The outputs differ, so the evidence needed to judge them differs as well.

Historical data also creates a boundary. If the organization launches a new product line or changes its refund policy, older complaints may no longer represent current conditions. The categorization model can then make weaker predictions because the business reality changed beyond its learned history. The change in real-world data compared with the training data is called data drift, and the resulting loss of accuracy is called model drift. That risk does not mean ML is unsuitable. It means the requirement ledger must identify the decision, relevant evidence, accountable owner, and acceptable limits before the organization relies on the output.

Compare the business output required from the complaint workflow before selecting a model type.

Labeled history and held-back evidence

Training data comes in two kinds, and the difference decides which problems a team can tackle.

• Labeled data: past examples that already carry the correct answer, such as complaints tagged billing, delivery, or account access. A model that assigns categories learns from this kind of history.

• Unlabeled data: past examples with no recorded answer, such as complaints nobody categorized. It can still be used to look for natural groupings, but it cannot teach the model what the right category is.

A model can score well on cases it has already seen and still fail on new ones. For this reason, teams keep part of the historical complaints out of training as test data (also called a holdout set). Once training ends, they run the model on those held-back cases, where the right answers are known but the model has never learned from them, and compare its output with the recorded labels. This shows how the model may behave on future complaints before the business relies on it.

The same logic explains the limit of historical data from earlier in this lesson. If a new product line produces complaint types that appear nowhere in the labeled history, neither training nor held-back checks can vouch for them. The business response is to add current labeled examples and keep reviewing results, not to assume the model knows the new cases.

ISO/IEC AI frameworks

ISO/IEC 23053 is an international framework that describes AI systems using machine learning concepts. It gives business, technical, legal, and vendor teams a shared way to discuss the parts and lifecycle of an ML-based AI system. It is a vocabulary and conceptual framework, not a claim that a particular model is effective or appropriate for a business use.

ISO/IEC 42001 is an international standard for an AI management system in an organization. It addresses how an organization establishes, implements, maintains, and improves a management system for AI. Leaders can use its language to align accountability and organizational management around AI activities without treating the standard as a substitute for a specific business decision.

The two standards serve different purposes. ISO/IEC 23053 helps teams describe ML-based AI systems consistently. ISO/IEC 42001 addresses the organizational management system around AI. Neither standard tells us whether complaint categorization should use predictive ML or whether customer correspondence should use GenAI. The stated workflow outcome and evidence still determine that choice.

AIB-C01 decision signals

For AIB-C01 scenarios, start with the required output. If the organization needs new text, images, code, summaries, or drafted responses, GenAI is the likely category. If it needs a label, score, ranking, or forecast drawn from historical patterns, predictive ML is the likely category.

Reject options that confuse training with inference. Training uses historical examples to build or update a model. Inference applies the trained model to a new case. Also reject claims that a model reliably knows facts absent from its training data or the information supplied at the time of use.

Use ISO/IEC 23053 when the scenario asks for a shared conceptual framework for ML-based AI systems. Use ISO/IEC 42001 when it asks for an organizational AI management system. The reusable selection rule is simple: identify the requested output first, then choose GenAI for content creation or predictive ML for labels, scores, and forecasts, while recognizing that both remain bounded by the information on which they rely.