Foundation Models vs Traditional ML — When to Use Each
A recurring AIF-C01 theme: deciding whether a scenario calls for a foundation model or a traditional task-specific ML model. This page walks through a representative question and the decision rule that covers most exam variants.
Short answer
The correct answer is B. Train a traditional supervised text classification model on the labeled dataset and deploy it as an endpoint.
Traditional supervised ML is the right fit here because the problem is narrow, well-defined, and has labeled data. A classifier such as logistic regression, XGBoost, or a fine-tuned small transformer trained on the 4,000 labeled tickets will typically match or beat a general-purpose foundation model on a fixed five-class task, while being far cheaper and faster at inference. Traditional models shine when you have a single task, reasonable labeled data, and stable categories — which is exactly the scenario described. Inference runs on small CPU or GPU instances, costs fractions of a cent per prediction, and latency is in milliseconds. A foundation model, by contrast, is pre-trained on massive unlabeled corpora to be task-agnostic. Its strength is generalizing to new tasks with zero-shot or few-shot prompting, not squeezing the last point of accuracy out of one fixed classification task. Every inference pays for billions of parameters the retailer does not need. Fine-tuning a 70B model on only 4,000 examples is overkill and expensive. Unsupervised clustering throws away the labels entirely and rarely aligns cleanly with business categories. The decisive clues in the scenario are: one task, labeled data, stable categories, and a cost-sensitive requirement — all classic traditional ML signals.
The Question
A retailer has 4,000 labeled customer support tickets tagged with one of five product categories. They want to automatically route new tickets to the right team with the highest possible accuracy for this single classification task and the lowest inference cost. The categories and ticket style rarely change. Which approach best fits this scenario?
Why B is correct
Traditional supervised ML is the right fit here because the problem is narrow, well-defined, and has labeled data. A classifier such as logistic regression, XGBoost, or a fine-tuned small transformer trained on the 4,000 labeled tickets will typically match or beat a general-purpose foundation model on a fixed five-class task, while being far cheaper and faster at inference. Traditional models shine when you have a single task, reasonable labeled data, and stable categories — which is exactly the scenario described. Inference runs on small CPU or GPU instances, costs fractions of a cent per prediction, and latency is in milliseconds. A foundation model, by contrast, is pre-trained on massive unlabeled corpora to be task-agnostic. Its strength is generalizing to new tasks with zero-shot or few-shot prompting, not squeezing the last point of accuracy out of one fixed classification task. Every inference pays for billions of parameters the retailer does not need. Fine-tuning a 70B model on only 4,000 examples is overkill and expensive. Unsupervised clustering throws away the labels entirely and rarely aligns cleanly with business categories. The decisive clues in the scenario are: one task, labeled data, stable categories, and a cost-sensitive requirement — all classic traditional ML signals.
Why the other options are wrong
Zero-shot prompting a foundation model on every ticket works but is inefficient. You pay per-token for every prediction, latency is higher, and accuracy on a narrow five-class task is usually below a purpose-trained classifier. Foundation models earn their cost when you lack labels or need many tasks, not when you have 4,000 labeled examples for one task.
Fine-tuning a 70B foundation model on 4,000 examples is expensive, operationally heavy, and almost always unnecessary for a single classification task. Provisioned throughput adds fixed hourly cost that violates the lowest-inference-cost requirement. Smaller supervised models reach comparable or better accuracy on this workload.
Unsupervised clustering ignores the labeled data you already have, which is the most valuable signal in the problem. Cluster boundaries rarely map cleanly to predefined business categories, so you would need a fragile post-hoc mapping step and accuracy would suffer.
How foundation models differ from traditional ML
Traditional ML models are trained on a labeled dataset for one specific task. A fraud classifier learns to flag fraud. An image model trained on cats and dogs only distinguishes cats and dogs. These models are task-specific, typically small to medium in size, relatively cheap to train and serve, and perform well when you have enough labeled data and the distribution does not drift. They rely on supervised learning (with labels) or unsupervised learning (without labels) applied to a narrow problem. You generally cannot repurpose them — a fraud model will not answer customer questions. Foundation models are a newer class. They are pre-trained on enormous volumes of largely unlabeled data such as web text, code, and images, using self-supervised objectives. Through scale and pre-training they develop broad capabilities and can tackle many downstream tasks without task-specific training. They support zero-shot prompting, where you describe the task in natural language and the model performs it, and few-shot prompting, where you supply a handful of examples. They also exhibit emergent capabilities — reasoning, translation, summarization, code generation — that smaller task-specific models do not. Foundation models are task-agnostic, very large (billions of parameters), expensive per inference, and hosted as APIs on services like Amazon Bedrock or Amazon SageMaker JumpStart. The practical tradeoff is straightforward. If you have one narrow task with labeled data, stable inputs, and cost or latency pressure, a traditional model usually wins. If you have many tasks, unstructured or unlabeled data, natural language interactions, or requirements that change frequently, a foundation model is the better fit. On the AIF-C01 exam, watch for these cues to pick the right camp.
Ready to see how you'd score?
Take the free practice quiz and find out which AWS Certified AI Practitioner domains you need to focus on. No signup required.
Practice 5 similar questions
Same cert, same or adjacent domain. Use these after reviewing the explanation.