Pre-built AI service vs custom model

WHAT THIS TESTS This checks pragmatic engineering judgment about when to consume a managed AI API versus invest in a custom model, weighing real trade-offs rather than defaulting to one.

A GOOD ANSWER COVERS Pre-built services like AWS Rekognition, Google Vision, or Azure AI Vision expose common capabilities such as object and face detection, text extraction, and moderation behind a simple API. They win when the task is a well-solved general problem, your team lacks deep ML expertise or labeled training data, and you need to ship quickly, since there is no training, tuning, or model hosting to manage and you pay per call. Building a custom model wins when your problem is domain-specific and the generic service is not accurate enough, for example recognizing specialized industrial parts or medical features, when you need full control over the model, its updates, and where data is processed for privacy or compliance, or when call volume is so high that per-request pricing becomes more expensive than running your own optimized model. The core trade-offs are time to market and operational simplicity from the managed service versus accuracy fit, control, data governance, and cost at scale from custom. A common middle ground is customization features like Rekognition Custom Labels that fine-tune a managed model on your data.

COMMON WRONG ANSWERS Always building custom for prestige or assumed superiority when an API meets the need faster and cheaper. Ignoring per-call cost crossover at very high volume. Overlooking data residency and privacy constraints that may forbid sending data to a third-party API. Assuming the managed service can be tuned to arbitrary domain accuracy.

LIKELY FOLLOW-UPS At what volume does custom become cheaper. How do you handle accuracy gaps. Where do hybrid customization options fit.

ONE CONCRETE EXAMPLE A startup needs basic content moderation immediately, so it calls Rekognition and ships in days. Later, needing to detect its own niche product categories that the generic model misses, it trains a custom model on labeled images to get the domain accuracy the API cannot provide.