Not every company needs a data science team to have predictive models in production. AutoML automates architecture search and hyperparameter tuning behind a declarative interface.
AutoML on Vertex AI automates two steps that consume most of a data scientist's time: model architecture search (neural architecture search) and hyperparameter tuning. The user provides labeled data and defines the objective (classification, regression, forecasting, object detection); Vertex AI trains, evaluates, and automatically selects the best combination within a compute budget (node-hours) that you define.
This doesn't eliminate the need for judgment: you still need to decide which columns are relevant, how to split train/test, and above all, whether the business problem is well framed as an ML problem.
Vertex AI covers four data domains with AutoML:
- **Tabular**: classification, regression, and forecasting on structured data (CSV or BigQuery tables), including probabilistic forecasting with hierarchical time series. - **Image**: single- or multi-label classification, and object detection with bounding boxes. - **Text**: classification, entity extraction, and sentiment analysis on your own corpora. - **Video**: action classification, per-frame object detection, and temporal event recognition.
Each domain exposes specific evaluation metrics (AUC-ROC, log loss, MAPE for forecasting) directly in the console, with no need for the user to calculate them manually.
`budget_milli_node_hours` is the parameter with the biggest impact on final cost: more budget means more architectures explored, with diminishing returns past a certain point.
The question isn't "which one is better?" but "where's the ceiling on marginal value?" AutoML wins when the dataset is moderately sized (thousands to hundreds of thousands of rows), the problem is standard (churn, fraud detection, demand forecasting), and there's no dedicated team iterating on architectures. Custom Training wins when the problem has specific architectural constraints (a time-series model with known domain components), when you need deeper explainability than Vertex Explainable AI offers, or when data volume justifies custom architectures that outperform what AutoML explores.
In practice, many teams use AutoML as a quick baseline: if a custom model doesn't significantly beat an AutoML model trained in a day, the additional engineering investment isn't worth it.
AutoML's tabular models include Shapley-value-based feature attribution with no additional configuration, showing which variables most influenced each individual prediction. This matters in regulated sectors (credit, insurance) where an automated decision must be explainable to an auditor or the affected customer.
An AutoML model deploys just like any other Vertex AI model: to an autoscaling endpoint, or via batch prediction for periodic mass scoring. Cost has two separate components: training (billed per node-hour once, or every time you retrain) and serving (billed per hour of active node in the endpoint, regardless of traffic volume). For models with sporadic traffic, batch prediction avoids paying for a 24/7 active endpoint.
AutoML eliminates the need to write neural network architectures, but it doesn't eliminate the need for good data engineering. A dataset with data leakage, untreated class imbalance, or poorly defined features will produce a model with excellent validation metrics and poor production behavior, regardless of how much compute budget you throw at it.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel