GCP PMLE · Domain 1 of 6

GCP PMLE Domain 1: Architecting low-code AI solutions

Learn when a problem can be solved without custom training: BigQuery ML and AutoML for predictive models, prebuilt Google Cloud AI APIs for common tasks, and Gemini, Imagen, Veo and Model Garden for generative use cases.

Exam weight
~13%
Official objectives
2
Estimated questions estimate
7–8
Suggested study time estimate
8–12 hrs

About 13 min read Reviewed October 2026

Overview

Architecting low-code AI solutions has two official objectives. Task 1.1 covers building classification, regression, forecasting and clustering models in BigQuery ML or with AutoML on Gemini Enterprise Agent Platform (formerly Vertex AI), including feature work and predictions in SQL and fine-tuning Gemini models from BigQuery. Task 1.2 covers picking a model from Model Garden, using industry and task APIs such as Document AI, Vision and Translation, building with Gemini, Imagen and Veo, and tuning Gemini apps for cost, latency and availability.

At about 13% of a 50-60 question exam, this section is one of the two smallest. It was refreshed in the June 2026 guide, which uses Agent Platform names throughout. As of October 2026 the exam is current with no retirement announced, and Google does not publish a passing score.

How it's tested

Domain 1 scenarios state a business problem, where the data sits and the team's skills, then ask for the least effort option that meets the need. The wrong answers usually work but cost more engineering. Look for clues such as data in BigQuery, SQL-only analysts, document processing or a need to cut Gemini latency.

  • Least-effort model

    Data and skills are described and you choose between BigQuery ML, AutoML, a prebuilt API or custom training.

  • Pick the API or model

    A task such as invoice parsing, image labeling or translation is given and you select the matching API or Model Garden model.

  • Tune the gen AI app

    A Gemini application is too slow, too costly or unavailable at peak, and you choose the setting or design change that fixes it.

Reasoning it through

A retailer's analysts know SQL, sales history sits in BigQuery, and they need weekly demand forecasts per store. Custom training on Agent Platform would work but needs Python and pipelines. AutoML forecasting is possible but requires exporting data. An ARIMA_PLUS model created with CREATE MODEL in BigQuery ML keeps the data in place and fits the team's skills, so BigQuery ML is the expected answer.

Objectives 2

Official objectives from the Google Cloud outline, in outline order.

1.1 Developing ML models using BigQuery ML or AutoML on Gemini Enterprise Agent PlatformModel types in BigQuery ML and AutoML, SQL feature engineering, predictions and Gemini fine-tuning from BigQuery.
What it means
This task asks you to build the right low-code predictive model for a business problem. You choose among classification, regression, forecasting and clustering, engineer or select features inside BigQuery ML, run predictions with SQL, train AutoML models on Agent Platform, and fine-tune Gemini models using BigQuery as the entry point.
What to understand
BigQuery ML trains models with CREATE MODEL statements, so data never leaves the warehouse. Know the common model types: logistic and linear regression, boosted trees, DNNs, k-means for clustering, ARIMA_PLUS for time series and matrix factorization for recommendations. The TRANSFORM clause bakes feature preprocessing into the model so serving applies the same logic. ML.PREDICT, ML.EVALUATE and ML.EXPLAIN_PREDICT cover scoring, evaluation and explanations. AutoML suits teams that want strong accuracy on tabular, image, text or video data without choosing an architecture. BigQuery can also call and tune remote Gemini models.
How it shows up
A team with data in BigQuery and SQL skills needs a standard model type; the answer is usually BigQuery ML, with AutoML when no one wants to pick algorithms.
Common confusion
Picking k-means for a labeled prediction task; clustering is for unlabeled grouping, not for predicting a known target.
1.2 Building AI solutions using Google Cloud AI APIs or foundational modelsModel Garden selection, industry APIs such as Document AI, Vision and Translation, Gemini, Imagen and Veo, and cost, latency and availability tuning.
What it means
This task covers solving problems with models someone else has already trained. You evaluate Model Garden options for a task, build with task and industry APIs such as Document AI, Cloud Vision and Cloud Translation, create and tune solutions with Gemini, Imagen, Veo and models-as-a-service from partners, and optimize Gemini-based applications for cost, latency and availability.
What to understand
Start with the narrowest tool that does the job. Document AI processors extract fields from invoices, receipts and forms; the Vision API labels images and reads text; Translation handles languages. Gemini covers multimodal reasoning and generation, Imagen produces images, and Veo produces video. Model Garden offers Google, open and partner models, some as managed APIs and some you deploy yourself. To control Gemini spend and speed, pick a smaller Flash-class model when quality allows, use context caching for repeated long prompts, use batch prediction for offline work, and use provisioned throughput when steady capacity is required.
How it shows up
A business task or a performance problem with a Gemini app is described, and you pick the API, model or optimization that meets it with least effort.
Common confusion
Building a custom OCR model when Document AI already has a processor for that document type.

High-yield concepts

The terms, rules and formulas this domain turns into questions most often.

  • BigQuery ML

    Create, evaluate and run ML models with SQL inside BigQuery.

    Remember
    No data movement; CREATE MODEL plus ML.PREDICT.
    Tested as
    Default answer when data and skills are in BigQuery.
    Trap
    Exporting data to train elsewhere for no reason.

  • TRANSFORM clause

    Defines preprocessing inside a BigQuery ML model.

    Remember
    The same transformations run at prediction time.
    Tested as
    Training-serving consistency in SQL.
    Trap
    Repeating preprocessing manually in each query.

  • ARIMA_PLUS

    BigQuery ML time-series model with seasonality and holiday handling.

    Remember
    Use for forecasting many series in SQL.
    Tested as
    Demand and sales forecasting scenarios.
    Trap
    Choosing linear regression for a time series.

  • Boosted tree and DNN models

    Higher-accuracy BigQuery ML options for tabular data.

    Remember
    Trade some interpretability for accuracy.
    Tested as
    Model type selection questions.
    Trap
    Assuming BigQuery ML only does linear models.

  • AutoML

    Managed training that searches architectures for you on Agent Platform.

    Remember
    Good accuracy, little ML code, longer training time.
    Tested as
    Low-code image, text, video and tabular tasks.
    Trap
    Using it when a prebuilt API already fits.

  • Model Garden

    Catalog of Google, open and partner models on Agent Platform.

    Remember
    Some are APIs, some you deploy to an endpoint.
    Tested as
    Model selection questions.
    Trap
    Assuming every Model Garden model is serverless.

  • Document AI

    Prebuilt and custom processors for document extraction.

    Remember
    Specialized parsers for invoices, receipts, IDs and forms.
    Tested as
    Document workflow scenarios.
    Trap
    Using generic OCR when structured fields are needed.

  • Gemini model sizes

    Pro-class models for hard reasoning, Flash-class for speed and cost.

    Remember
    Choose the smallest model that meets quality.
    Tested as
    Cost and latency questions.
    Trap
    Defaulting to the largest model.

  • Context caching

    Reuses a large shared prompt prefix across requests.

    Remember
    Cuts cost and latency for repeated long context.
    Tested as
    Gemini optimization questions.
    Trap
    Confusing it with response caching in the app.

  • Provisioned throughput

    Reserved Gemini capacity for predictable traffic.

    Remember
    Protects availability against shared-quota limits.
    Tested as
    Availability questions.
    Trap
    Buying it for small, spiky workloads.

  • Imagen and Veo

    Google's image and video generation models.

    Remember
    Match output modality to model.
    Tested as
    Generative use-case questions.
    Trap
    Picking Gemini for pure image creation.

Side by side

Concepts candidates mix up, compared on the point that decides the answer.

Low-code training options

ConceptMeaningKey differenceCommon trap
BigQuery ML SQL-defined models trained where the data lives. Fastest for SQL teams with warehouse data. Ignoring it because it seems too simple.
AutoML Managed architecture search on Agent Platform. Handles images, text and video, not just tables. Expecting fine control over the model.
Custom training Your own code in a container on Agent Platform. Full control at the highest effort. Choosing it when a low-code tool meets the requirement.

Prebuilt AI options

ConceptMeaningKey differenceCommon trap
Task APIs (Vision, Translation, Speech) Pretrained models behind a simple API. No training data needed. Training a model to do what the API already does.
Document AI Processors that return structured fields from documents. Understands layout and key-value pairs. Treating it as plain OCR.
Gemini via Agent Platform Multimodal foundation model controlled by prompts. Flexible but needs prompt design and evaluation. Using it where a cheaper specialized API fits.

Gemini cost and latency levers

ConceptMeaningKey differenceCommon trap
Smaller model Flash-class instead of Pro-class. Lower cost and latency, slightly less reasoning. Assuming quality always drops too far.
Batch prediction Asynchronous processing of large request sets. Cheaper for work that is not time sensitive. Using it for interactive chat.
Context caching Store a reused prompt prefix once. Helps when many requests share long context. Applying it to prompts that never repeat.

Exam traps

Each trap is an answer that looks right. Learn the clue that should change your mind.

  1. Export to train

    What candidates pick
    Export BigQuery data to Cloud Storage and train a custom model
    Why it looks right
    Custom training feels more professional.
    Why it's wrong
    Moving data adds cost and pipeline work when BigQuery ML trains in place with SQL.

    Clue to watch for Data is already in BigQuery and the team writes SQL.

  2. Clustering for prediction

    What candidates pick
    A k-means model to predict churn
    Why it looks right
    Clustering groups customers, which sounds related.
    Why it's wrong
    Churn is a labeled outcome, so it needs a classification model such as logistic regression or boosted trees.

    Clue to watch for A historical label column exists.

  3. Generic OCR for invoices

    What candidates pick
    Vision API text detection
    Why it looks right
    It reads text from images.
    Why it's wrong
    Vision returns raw text, while a Document AI invoice processor returns structured fields such as totals and dates.

    Clue to watch for The output must be named fields, not text.

  4. Biggest model wins

    What candidates pick
    Switch to the most capable Gemini model
    Why it looks right
    Better reasoning seems safer.
    Why it's wrong
    The scenario asks for lower latency and cost, which a smaller model or caching delivers.

    Clue to watch for Quality is already acceptable; speed or spend is the issue.

  5. Train a new generator

    What candidates pick
    Train a custom image model
    Why it looks right
    Brand-specific images seem to need it.
    Why it's wrong
    Imagen can generate images and can be customized, which is far less work than training from scratch.

    Clue to watch for Marketing visuals are needed quickly.

  6. AutoML for everything

    What candidates pick
    AutoML when a translation API exists
    Why it looks right
    AutoML is low-code too.
    Why it's wrong
    AutoML still needs labeled training data, while the Translation API needs none.

    Clue to watch for The task is a standard, well-served capability.

Answer framework

Use these steps to pick the lowest-effort option that still meets accuracy, cost and latency needs.

  1. Check whether a prebuilt API or Document AI processor already solves the task; if so, use it before considering any training or tuning work.
  2. If the task is generative, start with a Gemini, Imagen or Veo model from Model Garden and improve results with prompting before tuning.
  3. If you need a predictive model and the data sits in BigQuery with SQL skills on the team, build it in BigQuery ML.
  4. If the data is images, text or video, or no one wants to choose algorithms, use AutoML on Agent Platform.
  5. Optimize last: right-size the Gemini model, cache shared context, batch offline work and reserve throughput only for steady, critical traffic.

Practice question

An insurance company receives thousands of scanned claim forms daily and must extract policy number, claim date and amount into BigQuery. The team has little ML experience and needs a solution within two weeks. What should the ML engineer recommend?

Want more like this? The GCP PMLE practice test has questions for every domain.

Checklist

0 of 11

Tick each one you can do without notes. Your ticks stay in this browser.

Study strategy

Start with BigQuery ML hands-on, because it is the most testable tool in this section. In a sandbox project, create a logistic regression, a boosted tree, a k-means and an ARIMA_PLUS model on public datasets, then run ML.EVALUATE and ML.PREDICT on each. Add a TRANSFORM clause to one model and note how prediction queries become simpler. Then train one AutoML model on Agent Platform so you know its workflow and the time it takes.

Next, build a one-page map of prebuilt options: Document AI processors, Vision, Translation, Speech-to-Text, and the Gemini, Imagen and Veo families in Model Garden. For each, write the input, the output and one scenario clue. Read the Agent Platform pages on context caching, batch prediction and provisioned throughput, and summarize when each one pays off.

Finally, drill scenario questions with one rule in mind: choose the least engineering that meets the stated requirement. Because this section is only about 13% of the exam, aim for steady accuracy rather than deep study, and move on to Sections 3 to 5 once you score well here.

FAQ

How much of the GCP PMLE exam is Architecting low-code AI solutions?

Architecting low-code AI solutions is about 13% of the exam, tied with Monitoring AI solutions as the smallest of the six sections, which works out to roughly seven or eight questions.

Will BigQuery ML statements appear in questions?

You will not write code, but questions may show BigQuery ML statements, so you should recognize CREATE MODEL options and the ML functions.

When is AutoML better than BigQuery ML?

When the data is images, text or video, or when the team wants a managed search for a strong architecture and accepts longer training time.

Is Vertex AI still on the exam?

The June 2026 guide uses Gemini Enterprise Agent Platform instead of Vertex AI. The services are the same family, so older Vertex AI material still helps if you map the names.

What gen AI content is in this section?

Choosing models from Model Garden, building with Gemini, Imagen and Veo, fine-tuning Gemini from BigQuery and optimizing Gemini apps for cost, latency and availability.

What is the passing score?

Google does not publish a passing score for the Professional Machine Learning Engineer exam. You receive a pass or fail result.

How should I spot a low-code answer?

Look for clues like limited ML staff, tight deadlines, data already in BigQuery or a common task that a Google API already handles.

Other Artificial Intelligence & Machine Learning certifications

Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.

Test your Domain 1 knowledge

You have reviewed BigQuery ML, AutoML, prebuilt APIs and Gemini optimization. See whether you can pick the lowest-effort solution in timed exam-style scenarios.

Take the GCP PMLE practice test Check your GCP PMLE readiness

Keep going