AI/ML development lifecycle

The primary drawback is a lack of domain specificity (Option A). Pre-trained foundation models are typically trained on vast, generalized datasets. They may not recognize the unique, proprietary transaction patterns specific to the financial firm, resulting in poor fraud detection accuracy compared to a custom model trained specifically on the firm's historical data. Pre-trained models actually save initial training costs (Option B is false). Foundation models can process tabular data if architected to do so, though it's less common than text (Option C is false). Pre-trained models can be hosted securely inside a private VPC using SageMaker, so they do not expose data to the public internet (Option D is false). Reference: https://docs.aws.amazon.com/bedrock/latest/userguide/what-is-bedrock.html