Giving an AI a chat box is easy. Getting it to do useful work across a company’s apps is the harder proposition Google is now testing with a new Gemini Enterprise agent. The tool is in private preview for business customers, and Google plans to bring it to consumers later.
From prompts to objectives
More than answers: Users can give the agent a goal, and it can plan the work, draw on custom skills and tools, and connect to company systems. Requests can include files, folders or projects assembled for a particular workstream.
The agent can work inside Google Workspace apps such as Gmail, Drive, Docs, Sheets and Calendar. It can also connect to Microsoft 365, Slack, Jira, Confluence, Git, BigQuery, Databricks, Postgres, Snowflake and other business systems. Google says it can work with Model Context Protocol servers inside or outside a company’s network.
One point of entry: Gemini Enterprise is designed to give users a single place to assign and follow work, while making the agent available through mobile devices, desktop computers, the web, a command-line interface and services including Workspace, Microsoft 365, ServiceNow and Slack. Because it runs in the cloud, Google says the agent can retain context across connected devices.
A task inbox shows what the agent is doing, including its progress, use of skills and code, and delegation to subagents. That is more visibility than a mysterious loading spinner, though it still leaves businesses to decide what level of access an AI worker should have.
A coworker with an account
Its own identity: The agent can have a dedicated Workspace account and email address, including an address in the format @agents.company.com. Users can contact it by tagging or emailing it, sharing material with it, or adding it to a group conversation.
Google says the agent can use company context such as team membership, time zones, approval requirements and calendars. Actions are recorded in an audit trail under the agent’s identity, rather than appearing as if a human employee performed them.
That gives the agent a recognizable place in workplace systems, along with a record of its actions. It does not remove the need for companies to set permissions and decide which tasks should require human approval.
Model choice: Gemini will select a model for a task by default. Users can also choose one themselves, including Anthropic’s Claude models. Google says it plans to add open-source and other private models to the picker.
The company is also offering spending controls, including multi-model orchestration, smart routing and real-time spending caps. Those controls matter because an agent that can take on more work can also generate more usage, a connection enterprise buyers tend to notice before the invoice arrives.
Business first, consumers later
Google is starting with businesses, saying that enterprise deployment gives it a chance to address security, scale and performance challenges before a wider release. Early testers include On, Shopify and PayPal.
The new agent is separate from Gemini Spark, Google’s consumer-facing agent launched in May. Spark is available to subscribers through a different interface in the Gemini app. The planned consumer rollout for the new agent should not be mistaken for the arrival of Spark itself.
Google CEO Sundar Pichai said Gemini has more than one billion monthly active users, and that nearly 90 percent of Fortune 100 businesses use Gemini Enterprise. Those figures give Google a large existing audience to build on, though using an enterprise AI product is not the same as handing an agent the keys to company systems.
For now, Google is testing whether one agent can move between apps, models and work contexts without making the user coordinate every step. The promise is a coworker that can act. The practical test is whether companies can give it enough access to be useful, without treating a task inbox as a substitute for judgment.