SageMaker Autopilot

Managed AutoML services are usually a black box: data in, model out, nothing visible in between. SageMaker Autopilot is AWS's tabular AutoML service, and its distinguishing move is the opposite — it hands you the notebooks.

What a job does
You point Autopilot at a tabular dataset in Amazon S3 and name the target column. It infers the problem type and runs the pipeline itself:

• profiles the data and writes a data exploration notebook reporting missing values, cardinality and target balance;
• generates candidate pipelines — each a feature-engineering recipe paired with an algorithm such as XGBoost, a linear learner or a neural network — and records them in a candidate definition notebook you can read and edit;
• trains and tunes those candidates, either in hyperparameter-optimisation mode or in an ensembling mode built on AutoGluon;
• ranks everything on a leaderboard and, on request, deploys the winner to a real-time endpoint or a batch transform job.

Why it matters in AutoML
Autopilot's value is that the automation is inspectable and reproducible. Because every candidate is emitted as code, a team can take the generated pipeline, change one step, and run it as an ordinary training job — the search becomes a starting point rather than a verdict. It also inherits the surrounding platform: experiment tracking, the model registry, and SageMaker Clarify for feature-attribution and bias reports. The trade-offs are the familiar managed-service ones — you pay per training job, the search space is AWS's, and the result travels only as far as the generated artifacts take it.