An internal AI agent at Bunnings has saved staff half a million hours of administrative work, according to Google, as parent company Wesfarmers deploys agents across its retail portfolio. The figure was cited at Google Cloud's Gemini at Work 2026 event, where the company announced the Gemini agent, a single agent for knowledge work, content creation and coding.
The announcements are detailed in Gemini at Work 2026: Introducing Gemini agent, adapted from Google Cloud CEO Thomas Kurian's keynote.
Google said shopping agents that help customers discover products and build their carts have lifted conversion rates up to three times at Kmart and Officeworks. It did not say what tasks the Bunnings agent performs or how the saving was measured.
Zip, another Australian company, is using Gemini Enterprise to build what Google calls an AI-Native Product Factory for its US business. Cross-functional teams will use it to connect customer research, design, engineering and compliance.
Kurian said nearly 500 Google Cloud customers each processed more than one trillion tokens in the last year. Nearly 80% of Google Cloud customers now use its AI products.
"Work now starts in the prompt window," Kurian said.
The Gemini agent takes objectives rather than instructions. Users delegate an outcome and come back to finished work. Tasks can be assigned directly, scheduled, or triggered by events.
The agent runs in the cloud and keeps one set of memory and context across web, mobile, desktop, command line, Google Workspace, Microsoft 365 and Slack. Work that runs for hours or days continues after a user closes their laptop.
Gemini chooses the model for each job, orchestrating across Google's Gemini models and Anthropic's Claude models. Google said other private and open models will be added.
Coworkers with their own inbox
The agent can create temporary sub-agents, each with its own identity, to run parallel or sequential steps over hours or days.
It can also act as a "coworker agent" with a persistent role. Coworker agents get their own @agents.company.com email address, persistent storage, and a Workspace account with calendar, Drive and directory presence.
Colleagues can add a coworker agent to a Chat space or @mention it in a document comment. Its suggested edits appear under its own name in version history. Google said it sees only what is shared with it, and access follows existing sharing and membership.
Gemini connects to Confluence, Microsoft Office, Teams, Slack, Jira, Salesforce, ServiceNow, BigQuery, Databricks, Postgres and Snowflake, and to files on a user's desktop. It can also work with any Model Context Protocol (MCP) server inside or outside the company network.
Teams can publish reusable skills and tools to shared company registries. Google said Gemini keeps four kinds of memory: session, semantic, procedural and episodic.
For unstructured content, a new Smart Storage capability enriches PDFs, images, scans and audio recordings in place, writing context back onto the object itself. Google said the intelligence stays where the data resides and inherits existing security settings.
A Knowledge Catalog maps business definitions once so all agents use them. Google said Bloomberg Media lifted SQL query accuracy by 63% during initial development by grounding its data agents in the catalog.
Industry versions for financial services and legal are in preview, with government, healthcare and retail to follow. The legal version inherits matter-level permissions and ethical walls from document management platforms NetDocuments and iManage.
Governing agents like employees
Google framed agent governance around four questions: who the agent is, what it may do, what it did, and what it should never touch.
Each agent receives a cryptographically attested identity with least-privilege permissions, governed like an employee. Every action is written to an audit trail attributed to the agent rather than a person.
Agents run inside a sandbox with its own network boundary. All traffic in, out and between agents passes through Agent Gateway, which Google describes as an AI network firewall enforcing organisational policy in real time.
Policies are written once and applied to every agent in the company. Google's example is "agents may not open documents classified Need to Know."
Administrators can also set hard spending caps per project in the Cloud Billing Console. A project's agent pauses when its cap is reached, and per-project tracking lets organisations charge AI costs back to departments.
The approach to agent identity follows OpenAI's move last week to give always-on agents their own identity.