AWS Certified AI Practitioner

Customizing or fine-tuning a foundation model on proprietary datasets typically provides better domain alignment and can improve task-specific performance, but it comes with tradeoffs. Two primary tradeoffs are higher cost and higher implementation complexity. Higher cost arises from the compute resources required for training/fine-tuning, data storage, and ongoing maintenance, as well as potential licensing and infrastructure expenses. Higher implementation complexity comes from the need to prepare and curate training data, set up training pipelines, handle versioning, evaluate and validate models, and integrate the customized model into production.

Other options are less accurate as primary tradeoffs: customization usually aims to improve accuracy (so "Reduced accuracy" is not generally correct unless done poorly). "Increased risk of hallucination" is not guaranteed — customizing with high-quality domain data often reduces hallucination for domain-specific queries, though it can introduce new failure modes if data are noisy. "Higher latency" is possible if a customized model is larger or served differently, but it is not an inherent or guaranteed tradeoff of customization and is not as central as cost and implementation complexity.