The machine learning task of assigning an input to one of a set of predefined categories based on patterns learned from labeled training data. Classification is the structural foundation behind most AI tools that produce yes-or-no, this-or-that decisions in marketing workflows.
Also known as predictive classification, category prediction, supervised classification
Classification learns a mapping from inputs to categories. The training process exposes the model to labeled examples, “this email is spam,” “this lead is high-value,” “this image is brand-safe,” and the model learns the patterns that distinguish each category. When presented with a new, unlabeled input, it applies those learned patterns to assign a category.
Classification tasks come in several varieties. Binary classification assigns inputs to one of two categories. Multiclass classification assigns to one of three or more. Multilabel classification allows a single input to belong to multiple categories simultaneously, which is relevant for content tagging where a piece of content can carry several topic labels at once.
The quality of a classifier depends on the quality and representativeness of its training data, the appropriateness of the model architecture for the task, and the alignment between what the categories mean in the training data and what they mean in the deployment context. A brand safety classifier trained on general content categories may not reflect a specific client’s suitability requirements.
Most of the AI-powered decisions agencies encounter in their toolstack are classification decisions in disguise. Whether to show an ad to a user, whether to route a lead to sales, whether to approve a content piece for a platform, whether to flag an image as risky: each is a classification problem. Understanding what drives classification quality changes how agencies configure, evaluate, and defend these tools.
Label quality determines classifier quality. A classifier is only as good as its training labels. If the historical data used to train a lead scoring classifier was labeled by sales reps with different qualification criteria, the model will learn a noisy and inconsistent definition of “qualified.” Agencies should ask vendors how their training data was labeled and how label consistency was verified.
Category definitions drift over time. What “brand-safe” means, what “high-intent” means, and what “qualified lead” means can shift as client strategy evolves. A classifier trained on last year’s categories will apply last year’s definitions. Agencies should build category review into their AI tool maintenance schedules, not just model performance review.
Confidence scores are more useful than hard predictions. Most classifiers can output a probability rather than just a hard category assignment. Using the probability score rather than the threshold-based prediction enables more nuanced downstream decisions. An agency that uses the raw score for routing rather than a binary yes/no can apply different thresholds for different downstream actions.
An agency is using an AI content classification tool to route incoming client creative briefs to the right studio team. The classifier was trained on the previous year’s brief archive with four team categories. After a studio restructuring that created two new specialist teams, the routing accuracy drops noticeably. The agency investigates and finds the classifier is assigning briefs for the new categories to the closest legacy category, because it has never seen training examples for the new teams. The fix requires generating new training examples for the restructured categories and retraining the classifier, which takes two weeks. The lesson: classifiers are not automatically aware of organizational changes.
The generative AI foundations module of the workshop covers how today’s models work, what they can and can’t do, and how to choose between them.