Prepare and test your skills

Prepare and test your skills

The core categories of Google Cloud AI and ML solutions are pre-trained APIs, Vertex AI (opens in a new tab) platform services, and custom model development tools. The three categories differ in who trains the model and how much work the user does. Pre-trained APIs require no training, Vertex AI provides a managed platform for the full model lifecycle, and custom tools train models on organization-specific data.
Pre-trained APIs are Google-managed services that perform a specific AI function without any training by the user. Google has already trained the underlying models, so data flows from the application to the API and back as a finished result. The application depends on the service for inference but does not manage training data or infrastructure.
Common pre-trained APIs include:
Pre-trained APIs require no training data and no machine learning expertise, so teams can add AI quickly. Their limitation is that they only recognize the categories, entities, and languages in their original training and cannot learn company-specific labels.
Vertex AI platform services provide a unified API, client library, and platform for building, training, deploying, and running inference. Vertex AI connects pre-trained capabilities with custom model development in one managed environment. Models deployed on Vertex AI receive requests for inference and return outputs to applications.
Key Vertex AI capabilities include:
Vertex AI is the appropriate choice when a business needs to manage training, deployment, and inference together without assembling separate tools for each stage.
Custom model development tools train models on a customer's own data and labels so the models recognize organization-specific patterns. These tools sit between pre-trained APIs with no training and fully custom code with maximum control. The tradeoff for higher accuracy on specialized data is that the team must supply labeled examples and perform setup and evaluation.
BigQuery ML brings machine learning to the data instead of moving data to a separate platform, which speeds development for SQL practitioners. AutoML is preferred when a general-purpose model is not accurate enough but the team does not want to write all training code.
Choosing between the three categories is a decision about who trains the model and how much customization is needed. Pre-trained APIs fit general tasks such as OCR and translation, Vertex AI fits lifecycle management and inference at scale, and custom tools fit company-specific labels and patterns.
A decision tree that branches on task standardness, warehouse data, and custom labels to select pre-trained APIs, BigQuery ML, AutoML, or Vertex AI.
| Category | Who trains the model | Best for |
|---|---|---|
| Pre-trained APIs | General tasks that fit an existing model | |
| Vertex AI platform services | Google manages the platform; customer configures and deploys | Inference, image generation, corpus search, and full lifecycle management |
| Custom model development tools | Customer, using AutoML, BigQuery ML, or deep learning images | Company-specific labels, entities, and patterns |
The categories also compose as layers. A custom model trained with AutoML can be deployed and served through Vertex AI, and output from a pre-trained API can feed a BigQuery ML pipeline as a use case grows.
Assessing factors influencing solution selection means matching a business problem to pre-built AI services or custom machine learning workflows. Pre-built services solve common tasks through an API call, while custom workflows build unique models with dedicated environments and hardware. Data volume, data type, expertise, customization, cost, and integration determine which path fits.
Pre-built AI services are ready-to-use products that accept a standard data type and return a result without model training. Data flows from the application to the service API and back to the application, so the application depends on the service but does not manage training. These services fit images, audio, text, video, and documents when the task is common.
Choose pre-built services when machine learning expertise is limited and speed matters, because the team calls an API instead of preparing large training datasets and managing infrastructure.
Custom machine learning workflows are developer-led environments and infrastructure for creating, training, and scaling unique models. Data flows from storage into a development environment, training runs on accelerated compute, and the resulting model is deployed for use. The workflow depends on both the software environment and the underlying hardware.
TensorFlow Enterprise provides a scalable environment for building TensorFlow models. Colab Enterprise provides collaborative notebooks for writing and running machine learning code. Deep Learning VM provides preconfigured virtual machines, while Deep Learning Containers provides preconfigured containers that run consistently across environments. Cloud TPU provides hardware acceleration that speeds up large training and inference jobs.
Choose custom workflows when no pre-built API matches the problem, when model behavior must be highly customized, or when large volumes of specialized training data justify the extra effort.
How data, expertise, cost, and integration guide the choice is a combined evaluation of six factors for each use case. Data type decides which pre-built API could apply, while data volume decides whether custom training is feasible. Expertise and customization decide whether the team can and must build its own model, and cost and integration decide whether an API call or a longer project makes business sense.
| Option | Data and customization | Expertise and cost | Integration |
|---|---|---|---|
| Pre-built AI services | Standard data types and small to moderate volumes; limited customization | Application skills, not training skills; lower upfront effort and faster value | Application sends data to the service API and receives results |
| Custom ML workflows | Large volumes of relevant training data; high customization | Machine learning expertise; higher effort for setup, training, and tuning | Team owns data preparation, training infrastructure, and model lifecycle |
Agent platforms add a third pattern for connected outcomes. Gemini Enterprise Agent Platform (opens in a new tab) builds and manages autonomous agents, Agent Platform Vision ingests and stores video for agents and analytics, and Enterprise Knowledge Graph consolidates organizational knowledge to support those agents. Select these platforms when the outcome requires connected agents and knowledge rather than a single API call or a single custom model.
Exam tip: Speech-to-Text converts audio to text, while Text-to-Speech converts text to audio — the direction of the flow determines the service.
Mapping a business use case to a Google Cloud AI product means matching data type, customization need, and desired outcome. Document AI (opens in a new tab) handles end-to-end document understanding, specialized APIs handle single tasks like vision or language, Vertex AI (opens in a new tab) unifies custom training and inference, and BigQuery ML brings prediction directly to warehouse data.
Document processing converts dark data trapped in emails, PDFs, scanned forms, receipts, and images into usable fields. Document AI automates data capture at scale with processors such as Enterprise Document OCR and Form Parser, which convert pages to text and extract names, addresses, and phone numbers. A typical pipeline runs in fixed order:
A four-step process flow showing documents from sources like Drive, scanners, and Cloud Storage flowing into Document AI Enterprise Document OCR, then to classification and entity extraction with the Natural Language API or AutoML Natural Language, and finally to storage or downstream systems for validation and analysis, with a footer noting complementary products Cloud Translation API, Workflows, and Vertex AI AutoML.
Complementary products complete the solution. The Cloud Translation API translates extracted text and whole PDF and DOCX documents, Workflows connects Google Cloud services into reliable pipelines, and Vertex AI provides AutoML training when pretrained processors are not enough.
Converting images to text creates machine-readable text before classification, extraction, or search. Enterprise Document OCR is built for document-heavy recognition at scale, while the Cloud Vision API uses TEXT_DETECTION or DOCUMENT_TEXT_DETECTION to detect printed text, handwriting, and text in PDF or TIFF files. Data flows from the image file to the chosen service, which returns text for later steps. When the need is to locate visual regions rather than read all words, AutoML Vision object detection converts sections of images into locatable objects that can feed a document workflow.
Document classification assigns a category label after text is available, so optical character recognition must run first and classification runs second on the resulting plain text. The Cloud Natural Language API classifies content with a generalized list of categories and requires no training. AutoML Natural Language classification trains a custom model so a business can use its own category labels.
| Product | Model approach | Choose when |
|---|---|---|
| Cloud Natural Language API | Pretrained with generalized categories | Standard categories fit and no training data is available |
| AutoML Natural Language classification | Custom-trained on business labels | Company-specific categories are needed and labeled examples exist |
| AutoML Vision classification | Image-based classification | The task is classifying image content itself, not words in documents |
Exam tip: For document classification, text-based methods are generally more accurate than AutoML Vision classification, which looks at pixels rather than words.
Entity analysis identifies specific information inside documents, such as proper nouns, company branding, phone numbers, and addresses. Document AI Form Parser and pretrained processors extract known entities and common patterns during document processing. The Cloud Natural Language API identifies common, public entities in plain text, while AutoML Natural Language entity extraction creates a custom model for company-specific terms. Choose pretrained extraction when targets are public figures, brands, or standard patterns, and choose custom extraction when targets are internal product names, codes, or domain terms that a general model would not know.
Predictive analytics answers forward-looking questions such as forecasting demand, detecting anomalies, and segmenting customers. BigQuery ML trains, evaluates, and runs inference with GoogleSQL queries inside the warehouse, so machine learning comes to the data instead of requiring data movement. It supports forecasting with TimesFM, ARIMA_PLUS, and ARIMA_PLUS_XREG models, plus classification, regression, clustering, dimensionality reduction, recommendations, anomaly detection, and contribution analysis. Teams work through the console, bq command-line tool, REST API, or Colab Enterprise notebooks where SQL and Python can be used together, and explainable AI features help explain predictive results.
Customer interaction use cases need lifelike conversation and fast understanding of requests, reviews, and messages. Dialogflow and Conversational Agents provide virtual agents that carry multi-step conversations across chat and voice channels. For analysis where data already lives in tables, BigQuery offers AI functions that call Gemini models and Cloud AI APIs from SQL. General-purpose generative functions give full control over model, prompt, and parameters for generation, summarization, extraction, and enrichment, while managed functions use streamlined syntax where BigQuery chooses the model to optimize cost and quality for routine filtering, rating, and classification. Task-specific functions extend the same SQL pattern to natural language processing, translation, document processing, audio transcription, and computer vision.
Pre-trained APIs are Google-managed services that require no training by the user, Vertex AI provides a unified managed platform for building, training, deploying, and running inference, and custom model development tools train models on organization-specific data and labels. Pre-trained APIs fit general tasks such as OCR and translation, Vertex AI fits lifecycle management and inference at scale, and custom tools fit company-specific labels and patterns.
Use Cloud Natural Language API when standard generalized categories fit and no training data is available because it requires no training. Use AutoML Natural Language classification when company-specific category labels are needed and labeled examples exist, and run optical character recognition first so classification runs second on the resulting plain text.
BigQuery ML trains, evaluates, and runs inference with GoogleSQL inside the warehouse so machine learning comes to the data instead of requiring data movement. It supports forecasting with TimesFM, ARIMA_PLUS, and ARIMA_PLUS_XREG models plus classification, regression, clustering, dimensionality reduction, recommendations, anomaly detection, and contribution analysis.
An enterprise development team is building an expense management application that needs to extract text and recognize standard objects from user-uploaded receipt photos.
The team has the following requirements and constraints:
Which category of Google Cloud AI offerings should the team select?
Last updated