What changed

Amazon SageMaker AI added APIs and an AWS AI-ML agent skill for synthetic data generation, auto-labeling, human review, quality evaluation and transformation. Source data, generated datasets and labels remain in customer-controlled S3 buckets.

Why it matters

Teams can compose a training or evaluation dataset workflow without stitching together separate generation, labeling and review systems, but model agreement and synthetic volume do not establish correctness.

What to do

Start from a held-out labeled sample, define leakage and diversity checks, route low-confidence rows to named reviewers, and compare final labels with a manually curated baseline.

Caveat

Availability is limited to two US Regions. Confidence is model agreement, human review uses Label Studio Community Edition, and AWS customer results are vendor evidence.