Start with data, output, and constraints
Describe the input modality, desired output, available labels, prediction deadline, quality target, privacy requirements, and operating budget. A warehouse classification problem is different from extracting invoice fields or answering questions from internal policies. Establish a representative evaluation set and a simple baseline before comparing tools. A managed service reduces some infrastructure work, but the team still owns data quality, evaluation, access, and monitoring.
A retailer has labeled purchase history in BigQuery and needs a daily churn score. Describe the target and a time-based test. Explain why a warehouse-based classifier is a reasonable baseline to evaluate.
Compare warehouse and managed training options
BigQuery ML supports building and evaluating several model types using SQL over warehouse data. Its supported options include AutoML classification and regression. Managed AutoML can reduce model-development work for supported data and tasks, while custom training gives more control over the training code and environment. Compare achievable quality, data movement, customization, training time, serving requirements, and lifecycle cost; a more flexible approach can add operational responsibility.
The warehouse baseline fails an agreed quality threshold. Before moving to custom training, propose a comparison against a supported managed model using the same split, metric, and serving constraints.
Use specialized APIs for supported tasks
For extracting fields and layout from invoices or forms, evaluate Document AI processors against representative documents. For general image tasks or text detection, evaluate the relevant Vision API capabilities. The overlap does not make them interchangeable: compare required structured fields, file types, languages, quality on difficult samples, throughput, and pricing. Some documents require human review or custom processing when confidence or business impact warrants it.
An accounts-payable team needs vendor name, invoice number, and total from varied scanned invoices. Identify a specialized service to evaluate, build a small labeled test set, and define how uncertain extractions reach a reviewer.
Evaluate foundation models for flexible tasks
For summarization, language reasoning, or multimodal inputs, compare a suitable foundation model with a task-specific solution. Test prompting and, where current external context is needed, retrieval before considering model tuning. Evaluate correctness, evidence support, safety, latency, and cost on a held-out set; include missing or conflicting information. Check model availability, data handling, and regional constraints in current documentation rather than assuming a capability or fixed price.
A policy assistant must answer from current internal documents. Explain why retrieval and access control may be more relevant than tuning on an old policy snapshot. Include an unanswerable question in the evaluation set.
A team needs structured fields from scanned invoices and has limited ML development capacity. Which option is a sensible first evaluation?
Keep exploring
- BigQuery ML model creation ↗
- BigQuery ML AutoML models ↗
- Document AI overview ↗
- Cloud Vision API features ↗
Use the official certification page and its linked resources to validate your coverage. For guided labs, explore the Google Skills learning path. Revisit the practical exercise and explain your decision in your own words before moving on.