3.2 Google Cloud's AI and ML Solutions

Selecting a Google Cloud AI/ML Solution: The Three Tiers

Google Cloud organizes its AI/ML portfolio into three tiers, ordered from lowest to highest customization and effort:

1. Pre-trained APIs — Ready-to-use models Google has already trained on massive datasets. You call an endpoint; no training, no data labeling. Examples: Vision API, Natural Language API, Translation API, Speech-to-Text, Text-to-Speech.
2. AutoML — You upload your own labeled data and Google trains a custom model with AutoML's automated model-selection and hyperparameter tuning. No ML coding required; you get a model tailored to your domain.
3. Custom models on Vertex AI — Your data-science team writes the model code, chooses the architecture, and controls training, tuning, and deployment. Maximum control and differentiation, maximum expertise required.

The exam expects you to recognize a use case's tier and justify it along four axes.

The Four Selection Axes

Domain 3 names four decision dimensions explicitly. Memorize them.

A simple mnemonic: speed and low effort trade against differentiation and expertise. Pre-trained APIs win on speed/effort and lose on differentiation; custom models are the opposite.

Matching Use Cases to Tiers

The decision logic the exam tests:

• Use pre-trained APIs when the task is generic and off-the-shelf intelligence is enough. Examples: extracting text from receipts (OCR via Vision API), translating a website (Translation API), detecting sentiment in support tickets (Natural Language API), transcribing call center audio (Speech-to-Text). No unique data advantage; you need results now.
• Use AutoML when you have your own labeled data and a generic model underperforms, but you do not have a deep ML team. Examples: classifying your company's specific product defects from photos, tagging your specialized document types, classifying custom support categories. You differentiate on data without needing ML engineers.
• Use custom models on Vertex AI when the problem is core to your competitive advantage, your data and features are unique, and you have a data-science team that can architect and tune models. Examples: a fraud model with proprietary feature engineering, a recommendation system tuned to your catalog and user behavior, a forecasting model with custom time-series logic.

Tradeoffs in Practice

A common exam pattern asks which product to choose given a scenario. The reasoning pattern:

1. Identify the differentiation need. Is this a commodity task (translation, OCR) or a proprietary one (your fraud signature)? Commodity → pre-trained. Proprietary → AutoML or custom.
2. Check the team. No ML expertise → pre-trained API or AutoML. Dedicated data scientists → custom on Vertex AI.
3. Check the data. Generic, no labels → pre-trained. Labeled, domain-specific → AutoML. Unique, large, with engineered features → custom.
4. Check the timeline. Needed this week → pre-trained. Acceptable in days/weeks → AutoML. Months available → custom.

Worked Examples

• Scenario A: A bank wants to transcribe customer call audio to text for compliance. Generic speech recognition is sufficient and no in-house NLP team exists. → Speech-to-Text API (pre-trained).
• Scenario B: An agricultural company has millions of labeled photos of its crop varieties and wants a classifier that distinguishes its proprietary hybrids. No ML engineers on staff. → AutoML Vision (custom model on your data, no coding).
• Scenario C: A payments company has a proprietary feature pipeline, a team of ML engineers, and wants a fraud model that outperforms off-the-shelf offerings to gain competitive advantage. → Custom model on Vertex AI.

Why the Tiers Exist Together

The tiers are not a quality ranking — they are a fit ranking. A pre-trained API is the right answer when it solves the problem; building a custom model when an API would do is wasted effort. Conversely, relying on a generic API for a core differentiator cedes competitive ground. The exam rewards matching the tier to the business context, not assuming "custom is best."

Key Takeaways

• Three tiers: pre-trained APIs, AutoML, custom models on Vertex AI — ordered by customization and effort.
• Four axes: speed, effort, differentiation, required expertise.
• Pre-trained APIs win on speed and effort, lose on differentiation.
• AutoML wins on the balance — custom model on your data without ML coding.
• Custom on Vertex AI wins on differentiation, demands the most expertise.
• Match the tier to the use case: commodity task → API; proprietary data, no ML team → AutoML; core differentiator with data-science team → custom.