Two approaches to different kinds of tasks
Suppose you have a folder of customer messages. You might want to draft a helpful reply to each message, or you might want to identify which messages need urgent attention. The first task calls for creating content; the second calls for making a prediction about an existing item. This is the practical distinction between generative and discriminative AI.
A generative model produces an output such as a sentence, image or audio clip based on a prompt or other input. A discriminative model predicts a category, score or value from an input. For example, it might label a message as a billing question or estimate how likely it is to belong in that category. Neither approach is inherently better. The useful question is what output your workflow needs.
How the approaches learn
Generative systems learn patterns that help them produce plausible outputs. A text model, for instance, can learn to predict what text comes next given the text so far. When you ask it to draft an email, it generates words in sequence, conditioned on your request and the text it has already produced. That ability makes it flexible, but fluency does not establish that every statement in the email is true.
Discriminative systems learn to distinguish among outcomes. A model trained on examples of messages and their assigned categories can learn patterns associated with each category. Given a new message, it can predict a category or return scores that a team uses to sort messages. Its job is not to write a reply; it is to separate or estimate outcomes based on the input.
The distinction describes what a model is trained or used to do, not a strict boundary around every product. A generative language model can be prompted to classify a message, and a workflow can combine a classifier with a generator. Under the hood, model designs and training methods vary. For a user deciding how to build a workflow, the expected output and how it will be checked matter more than the label on the tool.
Choose by the output you need
Use a generative approach when the result needs to be expressed or composed: a first draft, a rewritten paragraph, a set of headline options or a summary. Give the system the intended audience, relevant context, constraints and examples of the desired format. Treat its output as a draft to review, especially if it includes factual claims, instructions or details about people.
Use a discriminative approach when the result belongs to a defined set of labels or is a value you want to estimate. Common examples include tagging support requests by topic, detecting likely duplicates or prioritizing items for review. Define the categories before you evaluate a model. If reviewers cannot consistently tell two categories apart, a model may struggle with that distinction too.
Some tasks need both. A team could categorize incoming questions, then generate a proposed response using approved information for each category. A person can review the category and the draft before sending anything. Separating those steps makes it easier to spot whether a problem came from incorrect routing, missing context or the generated wording.
Check the right kinds of errors
For generated content, ask whether the output follows the request, represents the supplied information accurately and omits unsupported details. A polished summary can still misstate its source. Where accuracy matters, compare the draft with the original material rather than judging it by how confident it sounds.
For predictions, inspect the cases the system gets wrong, not just the cases it handles easily. If a classifier marks urgent messages as routine, that may matter more to your workflow than occasionally marking routine messages as urgent. Decide what should happen when the model is uncertain: request human review, use a broader category or leave the item unclassified. A prediction score is not a guarantee that the prediction is correct.
Start with a small, reviewable workflow
Before adopting either approach, write down one task, the input it will receive and the output you expect. Gather a small set of representative examples, including ambiguous cases, and compare the outputs with your own criteria. This reveals practical problems that a generic demonstration may not show.
Also consider what information the system needs to process. Remove unnecessary personal or confidential details from inputs, and check the data-handling terms of any service you use. Whether an AI system creates content or makes predictions, it works best as part of a workflow with clear boundaries, appropriate review and a way to correct mistakes.
