Google Cloud AI for Enterprise: Platform Overview

If your company already runs on Google Cloud, BigQuery, or Workspace, you do not need a separate AI vendor. You need to know which Google Cloud AI service maps to which problem, and how the 2026 rebrand to the Gemini Enterprise Agent Platform changes how you buy and deploy.

How Google Cloud structures its AI services in 2026

Google's AI portfolio breaks into four layers, and picking the right layer matters more than picking the right model.

The bottom layer is infrastructure: TPU v5p and v6e pods, plus NVIDIA GPU instances on Compute Engine. The next layer is the Gemini Enterprise Agent Platform, which absorbs everything that used to be called Vertex AI: Model Garden, Agent Builder, Agent Engine, Vertex AI Search, training, tuning, and the model gateway. The third layer is the application AI APIs: Document AI, Speech-to-Text, Text-to-Speech, Translation, Vision AI, and the Customer Engagement Suite. The top layer is end-user products: Gemini for Workspace inside Gmail, Docs, Sheets, Meet, and the Gemini app, plus Gemini Code Assist for developers.

Most enterprise teams should treat the platform layer and the API layer as the default. You only drop to raw TPUs or GPUs when training a custom foundation model or running an unusual inference workload at extreme scale.

The Gemini Enterprise Agent Platform replaces Vertex AI

Vertex AI is gone as a brand, but the services are not. At Google Cloud Next 2026 in April, Google announced that all Vertex AI services and roadmap items now ship under the Gemini Enterprise Agent Platform, and that the older Agentspace product has been folded in. Existing customers do not have to migrate. The same APIs, the same SDKs, the same IAM bindings.

What did change is positioning. Google now sells one platform for the agent lifecycle: build agents in Agent Builder, run them on Agent Engine, store memory and sessions in managed services, govern tool access through the new Tool Governance controls, and observe everything through Cloud Logging and the Agent Platform console. If you were already using Vertex AI Search, Vertex AI Pipelines, or Vertex AI Studio, you are already on the new platform.

Model Garden: the default catalog for foundation models

Model Garden is the Google Cloud equivalent of a model marketplace, and it is where almost every enterprise generative AI project should start.

In May 2026 Model Garden hosts more than 200 models. Google's first-party lineup includes Gemini 3.1 Pro for the heavy reasoning work, Gemini 3.1 Flash for cheap high-throughput requests, Gemini 3.1 Flash Image for visual generation, Lyria 3 for audio, and the open Gemma 4 family for self-hosted use cases. Third-party models include Anthropic's Claude Opus, Sonnet, and Haiku, plus Meta Llama, Mistral, and select specialty models from partners. Everything goes through a single Vertex AI endpoint with one IAM permission model and one billing line.

The win for enterprises is not the breadth of the catalog. It is that you can swap models behind the same API call without redoing security review every time a new frontier model ships.

Agent Builder is where you build agents

Agent Builder, which still appears in the docs under the Vertex AI namespace, is the no-code and low-code surface for building agents. You define the agent's instructions, attach tools (functions, APIs, BigQuery queries, search indexes), connect a data store, and pick a foundation model. The agent runs on Agent Engine, Google's managed runtime that handles concurrency, sessions, memory, and tracing.

Two things make Agent Builder stand out from generic agent frameworks. First, it ships with first-class connectors to Google Workspace, BigQuery, Cloud Storage, and a long list of SaaS systems through Application Integration. Second, the new Tool Governance layer, announced in 2026, lets a central platform team approve which tools an agent is allowed to call before that agent reaches production.

For multi-agent systems, Google released the Agent Development Kit (ADK) and supports the Agent2Agent (A2A) protocol so agents from different vendors and frameworks can talk to each other.

Document AI handles the boring intake work

Document AI is the OCR, parsing, and document understanding service, and it earns its keep on intake-heavy workloads.

You point Document AI at a PDF, image, or scanned form, and it returns structured fields. Out of the box it has processors for invoices, receipts, W-2s, 1099s, contracts, drivers licenses, passports, and a long tail of industry-specific forms. You can also train a custom processor on your own document type with as few as a hundred labeled examples. Pricing is per page, with separate rates for the general processors versus the specialized ones.

The realistic enterprise pattern: an inbound document hits Cloud Storage, a Cloud Function triggers a Document AI call, the structured output drops into BigQuery or your ERP, and a Gemini-backed agent flags any low-confidence extractions for human review.

Speech, Translation, and the Customer Engagement Suite

For contact centers and any customer-facing voice or text workload, three APIs do most of the work.

Speech-to-Text supports more than 125 languages and dialects, with both batch transcription and streaming for live calls. Text-to-Speech goes the other direction with hundreds of voices powered by WaveNet and Studio voices. The Translation API translates text in 100-plus languages and can also translate documents directly in Docx, PPTx, XLSx, and PDF while preserving formatting.

The Customer Engagement Suite stitches these together. A real-world deployment looks like this: a customer calls in, Speech-to-Text transcribes their words, the Translation API converts the text into the agent's language in real time, the agent replies in their own language, the response gets translated back, and Text-to-Speech delivers it as synthesized speech. Welocalize, one of the largest enterprise localization providers, runs hundreds of millions of words per year through the Translation API.

BigQuery is the data layer that makes the rest worth it

The reason large enterprises pick Google Cloud over equivalent AI features on AWS or Azure is usually BigQuery, and the integration with the Agent Platform is what closes the loop.

Gemini and other Vertex AI models are exposed directly inside BigQuery as SQL functions. You can call ML.GENERATE_TEXT, ML.GENERATE_EMBEDDING, AI.GENERATE_TABLE, and a growing list of generative functions on top of any BigQuery table without moving data to a separate inference service. Agent Platform notebooks, including Colab Enterprise and Workbench, are natively integrated with BigQuery so a data scientist can move from SQL to Python to model training in one surface.

For agents, this matters because the data agents need to do useful work usually already lives in BigQuery. With the Vertex AI BigQuery connector, an agent can query the warehouse with natural language, get back grounded results, and act on them, with row-level security and IAM permissions still enforced at the warehouse layer.

Gemini for Workspace covers the productivity tier

Gemini for Workspace is the part most of your employees will actually touch every day, and the pricing changed materially in 2026.

Google retired the standalone Gemini Business and Gemini Enterprise add-ons (which used to cost 20 to 30 USD per user per month) and folded Gemini directly into Workspace plans. As of May 2026, Business Starter is 7 USD per user per month on annual billing, Business Standard is 14 USD, Business Plus is 22 USD, and Enterprise is custom. Google raised list prices by roughly 17 to 22 percent to absorb the bundled AI.

What you get: Help me write in Gmail and Docs, Help me organize in Sheets, Take notes for me in Meet, the Gemini app for chat and research, and admin controls in the Workspace admin console. Business Starter is intentionally limited (mostly the Gmail side panel). Business Standard is where the meaningful Docs, Sheets, and Meet features unlock. Enterprise plans add Vault, S/MIME, and the strongest data residency controls.

Pricing model: consumption everywhere except Workspace

Pricing on the Agent Platform side is almost entirely consumption-based, which is good for pilots and dangerous for unbounded production usage.

Foundation model calls are priced per million input and output tokens, with rates that vary by model tier. Agent Engine runtime is roughly 0.0864 USD per vCPU-hour and 0.0090 USD per GB-hour of memory, billed per second. Session and memory storage runs about 0.25 USD per 1,000 events. Vertex AI Search ranges from 1.50 USD per 1,000 queries on the basic tier to 6.00 USD per 1,000 on the enterprise tier with advanced features. Document AI is per page. Speech-to-Text and Translation are per character or per minute of audio. New Google Cloud customers get 300 USD in free credits valid for 90 days, and Express Mode lets you try the Agent Platform without enabling billing.

The expensive surprises usually come from three places: high-volume token usage on a Pro-tier model when Flash would have worked, idle Agent Engine deployments left running between pilots, and Vertex AI Search queries that scale with traffic faster than anyone modeled.

Compare the main Google Cloud AI services

Use this matrix to pick the right service for the workload, then layer them together.

Step-by-step: how to get started on Google Cloud AI as an enterprise

Here is the rollout sequence I recommend for an organization that has never run a serious Google Cloud AI project before.

Step 1: Stand up a clean GCP organization and billing structure

Before any AI project, get the org hierarchy right. Create a dedicated AI folder under your Google Cloud organization, with separate projects for each environment (sandbox, dev, staging, prod) and each business unit. Wire billing to a single billing account and turn on detailed usage export to BigQuery on day one. This is what makes chargebacks and per-team cost reports possible later.

Step 2: Enable the Agent Platform and pin down access

Enable the Vertex AI API (it still surfaces under that name in the API library) and the Agent Builder API in the relevant projects. Set up an IAM group for AI platform admins, a separate group for AI developers, and a third group for AI consumers. Pin model access at the org level using Org Policy: most enterprises should explicitly allow only the model families they have already cleared through legal and security review.

Step 3: Pick one workload and ship a pilot

Resist the urge to roll out everything. Pick one painful workload. Good candidates: an internal Q and A agent over the policy library, an invoice intake pipeline using Document AI, a multilingual support agent on top of the Customer Engagement Suite, or a BigQuery analytics agent for a single department. Build, deploy on Agent Engine, get real users on it, and measure.

Step 4: Wire it into Workspace and BigQuery

Once the pilot works, integrate it where employees and data already live. Surface the agent in Gemini for Workspace via custom Gems or Workspace Add-ons. Connect it to BigQuery so it can query live data with row-level security intact. Push activity logs to Cloud Logging and structured events to BigQuery for monitoring.

Step 5: Add governance and scale horizontally

Once one workload is in production, turn on Tool Governance to gate which tools new agents can call. Define a model approval list. Stand up a small platform team that owns the Agent Platform, and let business units self-serve agents on top of that platform. This is the same playbook the strongest enterprise AI teams use on AWS Bedrock and Azure AI Foundry — the platform team owns the rails, the business units own the workloads.

For more on the org-level side of this playbook, the enterprise AI adoption roadmap and the guide on how to build an AI center of excellence cover the team and process patterns that make the platform actually pay off. If you are evaluating Google Cloud against alternatives, the AWS AI services overview is the closest direct comparison.