GCP for AI Companies: An Implementation Guide
Back to Blog

GCP for AI Companies: An Implementation Guide

The short answer

  • Vertex AI is now Gemini Enterprise Agent Platform. Google announced the change on April 22, 2026. Google says Vertex AI services and roadmap items are now delivered through Agent Platform.
  • A common build is simple. BigQuery handles analytics. Cloud SQL or Firestore holds records. Cloud Run or GKE runs the apps. Pub/Sub queues slow jobs. Agent Platform runs matching models.
  • Choose your region first. If candidate data must stay in Canada, Google Cloud has regions in Montréal and Toronto. Privacy compliance is still your job, not Google’s.
  • Turn on the logs you need. Google encrypts stored data by default. Data Access audit logs are off by default for every service except BigQuery.
  • Some older advice is stale. Deployment Manager support ended March 31, 2026. Cloud Source Repositories is closed to new customers. Spot VMs are the newer version of preemptible VMs.

What changed on Google Cloud

Some names and tools in older guides, including earlier versions of this one, are out of date. Google has renamed, retired or replaced several of them.

Use this table as a quick check before you copy an older architecture.

Google Cloud names and tools that changed, as of November 10, 2026
Older name or toolWhat changedWhat to use nowSource
Vertex AIGoogle introduced Gemini Enterprise Agent Platform on April 22, 2026 as its evolution.Gemini Enterprise Agent Platform. Check the docs for SDK and API names.Google Cloud blog
Vertex AI Feature Store (Legacy)Deprecated February 17, 2026. Shutdown is set for February 17, 2027.The current Feature Store, which uses BigQuery as its data source.Deprecation list
Cloud FunctionsRenamed Cloud Run functions on August 21, 2024.Cloud Run functions.Release notes
Cloud Deployment ManagerSupport ended March 31, 2026.Infrastructure Manager, which uses Terraform.Deprecation notice
Cloud Source RepositoriesClosed to new customers on June 17, 2024.Secure Source Manager, or a Git host your team already uses.Release notes
Preemptible VMsSpot VMs are the latest version.Spot VMs for fault-tolerant work.Spot VMs docs

Why teams build on Google Cloud

Google Cloud is one of the three large cloud providers. The right pick depends on your team’s skills, your contracts and where your data must live. This guide does not claim Google Cloud is better than the others.

Here is what Google documents today:

  • Google lists 43 regions and 130 zones on its locations page, last updated October 28, 2026.
  • Google describes Agent Platform as a place to build, scale, govern and optimize AI agents. It brings model selection, model building and agent building together, per the launch post.
  • Google Cloud has regions in Montréal and Toronto, listed on its Compute Engine regions page.

What a talent platform needs from its cloud

Most AI hiring platforms need the same five things:

  • A store for candidate and employer records.
  • A way to search and rank candidate profiles against a role.
  • A queue for slow jobs, such as verification checks.
  • Strict access control and logs for personal data.
  • A clear answer on where data is stored.

Core services for a talent platform

Each service below covers one job in a hiring platform. The mapping is a common pattern, not the only one.

Google Cloud services and where they fit in a talent platform
NeedGoogle Cloud serviceHow a talent platform uses it
Matching and rankingGemini Enterprise Agent Platform (formerly Vertex AI)Train, tune or call models that score candidate profiles against role requirements.
Analytics and reportingBigQueryHiring funnel reports and feature data for models.
Relational recordsCloud SQLAccounts, roles and applications. Supports MySQL, PostgreSQL and SQL Server.
Flexible documentsFirestoreProfile documents with fields that change over time.
App hostingCloud Run or Google Kubernetes Engine (GKE)APIs and web apps. Cloud Run suits simple services. GKE suits teams that need cluster control.
Background jobsPub/SubQueue verification checks that take minutes, so the app stays responsive.
Access controlCloud IAMRoles and permissions for people and services.
Encryption keysCloud KMSCustomer-managed keys where policy requires them.
Finding personal dataSensitive Data ProtectionDiscover and classify personal data stored in your projects.
Web protectionCloud ArmorSecurity policies and DDoS defense for public endpoints.

Cloud Run functions handle small event-driven jobs. Examples are validating a new profile or calling a background check provider.

Set up your environment

Start with structure and permissions

  1. Create separate projects for development, staging and production.
  2. Give people and services only the roles they need.
  3. Set up private networking with a VPC before you deploy services that touch candidate data.
  4. Write infrastructure as Terraform code. Deploy it with Infrastructure Manager or your own CI tool.
  5. Pick regions before you load data.

A test project keeps experiments away from live candidate profiles. It also makes access reviews simpler.

Choose a region for Canadian data

Where candidate data lives is a design choice. Google says you choose which region or zone hosts your resources, and that choice controls where your data is stored and used. See Google’s region guidance.

For Canadian data, look at Montréal and Toronto. Google notes that its Montréal region has three zones in one or two data centers and is expanding to at least three. Ask your Google Cloud contact what that means for your uptime plan.

Google also says services listed on its data residency page can be set to a specific data location. Not every service is on that list. Check each one you plan to use.

Spread workloads across zones. Plan for the loss of a whole region with backups or a second region. Google recommends a disaster recovery plan for that case, as described in its geography and regions guide.

Build matching models on Agent Platform

Vertex AI was Google Cloud’s machine learning platform. On April 22, 2026, Google introduced Gemini Enterprise Agent Platform as its evolution. Google says model selection, model building and agent building carry over, with new features for agent integration, DevOps, orchestration and security.

Google also says Vertex AI services and roadmap items will now be delivered through Agent Platform. If you inherit Vertex AI code, check Google’s current docs for SDK and API names before you plan a migration. Details are in the launch post.

A practical workflow for matching models

  1. Prepare features in BigQuery, such as skills, tenure and role history.
  2. Choose a base model or train your own.
  3. Evaluate on labeled data that your team has reviewed.
  4. Compare results across groups to look for bias.
  5. Deploy to an endpoint and watch for drift over time.

Keep a person in the loop for hiring decisions. Rules in Ontario and New York City may apply to AI used in screening. See the section on privacy and hiring-law checks.

Plan for feature store changes

Google deprecated Vertex AI Feature Store (Legacy) on February 17, 2026. Shutdown is set for February 17, 2027. The current Feature Store keeps feature data in BigQuery. If you built on the legacy version, plan the move now. Check the deprecation list for current dates.

Protect candidate data

Candidate profiles hold names, contact details and employment history. Treat them as personal information from day one.

Encryption

Google says it encrypts customer content at rest by default, with AES-256 used for storage-level data keys. A small number of Persistent Disks created before 2015 use AES-128. See Google’s default encryption page.

Google also encrypts requests in transit with TLS by default. If you need control over keys, use customer-managed encryption keys through Cloud KMS. Google explains both on its encryption keys page.

Audit logs

Audit logs answer who did what, where and when. They are not all on by default.

Which Cloud Audit Logs are on by default
Log typeDefaultWhat it records
Admin ActivityAlways written. You cannot disable it.Actions that change configuration, such as creating VMs or changing IAM permissions.
Data AccessOff by default, except for BigQuery.Reads of configuration and metadata, and user-driven reads and writes of user-provided data.
System EventAlways written. You cannot disable it.System-driven changes, such as autoscaling a managed instance group.
Policy DeniedGenerated by default.Access denied because of a security policy violation.

Google notes that Data Access logs can be large and may add log charges. Turn them on for services that store or process candidate profiles. Reading them takes the Private Logs Viewer role, so limit who holds it. Details are on Google’s Cloud Audit Logs overview.

Compliance reports are not your compliance

Google publishes SOC 2 Type II reports through its Compliance Reports Manager and says it issues only Type II reports. See the SOC 2 page. Google also lists ISO/IEC 27001 and other certifications on its compliance page.

Those reports describe Google’s controls. Google also states that cloud providers cannot formally certify that customers comply with privacy laws. You still own the review of how your platform collects, uses and shares candidate data.

Privacy and hiring-law checks

Not tax or legal advice. This section is general information. Rules change and depend on where you hire. Confirm your duties with qualified counsel.

A talent platform handles personal data and often supports hiring decisions. Both bring legal duties. The table lists rules that come up often for Canadian and US teams.

Rules to check when a platform handles candidate data
RuleWhere it appliesWhat it asksWhere to read it
PIPEDACanada, federalCovers private-sector organizations that handle personal information in commercial activity. People can ask why data is collected, how it is used and who sees it. They can also ask for access or correction.Google Cloud PIPEDA page
Ontario job posting rulesOntario, employers with 25 or more employeesSince January 1, 2026, public postings must say if AI is used to screen, assess or select applicants. The duty also applies when a recruiting firm does the screening for the employer.Ontario ESA guide
NYC Local Law 144New York CityAutomated employment decision tools need a bias audit within one year before use. Audit information must be public. Candidates must get notice.NYC DCWP page

Quebec also has its own privacy law, known as Law 25. Google lists resources for it on its compliance page. Other provinces and US states may add rules.

The Ontario rules also touch fraud. Operators of job posting platforms must give users a way to report fraudulent postings. They must also publish a written policy on how those postings are handled. Ontario describes a fraudulent posting as one made with intent to defraud, such as advertising a job that does not exist to collect personal information.

Control cloud costs

Cloud bills grow when nobody owns them. Assign an owner for each project and review spend on a schedule.

  • Spot VMs: Use them for batch work that can restart, such as re-scoring candidate profiles overnight. Google says it can reclaim Spot VMs at any time and that they are excluded from the Compute Engine SLA. See the Spot VMs docs.
  • Committed use discounts: Buy them only for steady baseline load. Google lists discounts of up to 55% for most machine series and up to 70% for memory-optimized series on resource-based commitments. Read the terms before you commit. See the CUD overview.
  • BigQuery: Storage and compute are billed separately. Queries use either on-demand pricing, based on data processed, or capacity pricing, based on slots over time. Partitioning and clustering tables can reduce the data a query reads. Test this on your own tables. See BigQuery pricing.
  • Budgets and alerts: Set budgets in Cloud Billing and alert the project owner before limits are hit.
  • Data you no longer need: Delete or archive candidate data that has no purpose. Retention duties vary, so confirm them with counsel.

Use Google’s pricing calculator to model growth. This guide does not quote prices, because they change.

Deploy and monitor

Delivery pipeline

  • Use Cloud Build for builds and tests, Cloud Deploy for delivery to GKE and Cloud Run, and Artifact Registry for container images.
  • Host source in a Git service your team already uses. Cloud Source Repositories is closed to new customers.
  • Define infrastructure in Terraform and deploy it with Infrastructure Manager. Do not start new work on Deployment Manager.
  • Use Cloud Run for stateless services. Use GKE when you need cluster-level control.

Monitoring and alerts

  • Cloud Monitoring, Cloud Logging and Cloud Trace cover metrics, logs and request traces.
  • Track API latency for candidate search, error rates, job queue depth and model score drift.
  • Alert only on conditions a person must act on. Noisy alerts get ignored.
  • Run load tests before launch with a tool such as Locust or JMeter.
  • Test restores from backups, not only the backups themselves.

Verify Google Cloud skills before you hire

Building this stack takes people who have done it. Resumes in this field often list tools and cloud names without proof. Verify before you shortlist.

Certifications that map to the work

Google Cloud certifications that map to talent platform work
Work areaGoogle Cloud certificationLevel
ArchitectureProfessional Cloud ArchitectProfessional
Data pipelines and BigQueryProfessional Data EngineerProfessional
Machine learning systemsProfessional Machine Learning EngineerProfessional
Security controlsProfessional Cloud Security EngineerProfessional
Delivery pipelinesProfessional Cloud DevOps EngineerProfessional
Agent systemsProfessional Agentic Architect (Google says the general availability exam launches in mid-November)Professional
Business leadersGenerative AI LeaderFoundational

The names come from Google’s certification page. Google also says its exams are being updated to reflect product changes announced at Google Cloud Next ’26.

Checks that matter more than a badge

  • Status and expiry. Google says Professional certifications are valid for two years. Associate and Foundational ones are valid for three. See Google’s renewal FAQ.
  • Renewal route. Since July 2026, some certifications can be extended by one year through courses and skill badges on Google Skills. Ask how the candidate renewed. See Google’s renewal post.
  • Dates and names. A candidate profile that lists Gemini Enterprise Agent Platform work dated before April 2026 deserves a follow-up. The work may have been done on Vertex AI under its earlier name.
  • Hands-on questions. Ask how they would enable Data Access logs, choose a region for Canadian data, or move off Deployment Manager. Real experience shows in the details.
  • Scope and dates. Compare the claimed role, project size and employment dates across the resume and the interview.
  • Work eligibility. Confirm the person can work in the country where the project sits.

Where Augmnt fits

Augmnt is a marketplace that connects businesses with verified AI, data and tech consultants. Every consultant is checked for resume fraud, skills, work eligibility and background.

Businesses can use the software platform, which has AI matching, ATS integrations and a dedicated Customer Success Manager. They can also use the staff augmentation service. There, the Augmnt team sources, verifies and delivers onshore consultants in the US and Canada. A Partner Program covers referral, integration and services partners.

Next steps

Frequently asked questions

Is Vertex AI still available on Google Cloud?

Google introduced Gemini Enterprise Agent Platform on April 22, 2026 as the evolution of Vertex AI. Google says all Vertex AI services and roadmap items are now delivered through Agent Platform, not as a standalone service. Google’s documentation still uses Vertex AI names in places, such as its deprecation list. Check the current docs for the SDK and API names your code uses.

Which Google Cloud region should a Canadian talent platform use?

Google Cloud has regions in Montréal and Toronto. Pick the region before you load candidate data, because it controls where your resources run and where your data is stored. Check that each service you plan to use can be set to a specific data location. Then confirm your needs with your privacy counsel.

Does using Google Cloud make a platform PIPEDA compliant?

No. Google says its customers are responsible for their own due diligence under PIPEDA. Google also says cloud providers cannot formally certify that their customers comply with privacy laws. Google publishes a PIPEDA whitepaper and assessments that can help your review. This is general information, not legal advice.

What does Google Cloud encrypt by default?

Google says it encrypts customer content at rest by default, with AES-256 used for storage-level data keys. It lists a small number of Persistent Disks created before 2015 that use AES-128. Google also encrypts requests in transit with TLS by default. You can add customer-managed encryption keys through Cloud KMS if you need control over keys.

Are Cloud Audit Logs turned on by default?

Some are. Admin Activity and System Event audit logs are always written and cannot be disabled. Data Access audit logs are off by default for every service except BigQuery. To record who read candidate data in other services, enable Data Access logs and restrict who can view them.

Which Google Cloud certifications should I look for in consultant candidate profiles?

Google lists Professional certifications for Cloud Architect, Data Engineer, Machine Learning Engineer, Cloud Security Engineer and Cloud DevOps Engineer, among others. Professional certifications are valid for two years. A certification is one signal. Ask for the active status and expiry date, and test hands-on skills with real scenarios.

Do I have to tell candidates when I use AI to screen them?

It depends on where you hire. In Ontario, employers with 25 or more employees that use AI to screen, assess or select applicants for a publicly advertised job must say so in the posting. In New York City, employers and employment agencies that use an automated employment decision tool need a bias audit within the prior year, public audit information and candidate notices. Other places may have rules too. This is general information, not legal advice.

Sources

These are the primary sources used for the facts in this guide.

Reviewed on November 10, 2026. Google Cloud changes often. Check each linked page before you act.