1.4.2. Managed vs. Custom: When to Build vs. When to Buy
💡 First Principle: AWS offers a spectrum from fully managed AI services (zero ML expertise needed) to fully custom model building (maximum control). The correct position on this spectrum depends on how unique your problem is—and the exam tests your ability to read that signal.
Use AI Services when the task is standard and general: translating text (Amazon Translate), extracting text from documents (Amazon Textract), identifying objects in images (Amazon Rekognition). These require no ML knowledge and no training data. The exam tests whether you know these services exist and when they're sufficient.
Use Foundation Models (Bedrock/JumpStart) when you need capabilities beyond what AI services offer but don't want to train from scratch. Fine-tune a pre-trained model on your domain-specific data. This is the sweet spot for many production use cases.
Use Built-in Algorithms when you have structured data and a well-defined ML problem (classification, regression, clustering). SageMaker's built-in algorithms are optimized for AWS infrastructure and scale automatically.
Use Custom Code (BYOC/Script Mode) when your problem requires a novel architecture, a framework SageMaker doesn't natively support, or when you need complete control over the training loop. This gives maximum flexibility at the cost of more engineering effort.
The exam signal is usually in the problem description: "The company wants to quickly deploy..." (AI service or foundation model), "The data science team has a custom PyTorch model..." (Script Mode or BYOC), "The team needs to classify tabular customer data..." (built-in algorithm).
⚠️ Exam Trap: Don't default to the most complex solution. If a question describes extracting text from scanned invoices, the answer is Amazon Textract—not training a custom OCR model. The exam penalizes over-engineering.
Reflection Question: A healthcare company wants to extract medical entities from clinical notes. Should they train a custom NER model or use an existing AWS service? What information would you need to decide?
