Google Cloud introduced a Gemini agent for business work on October 8 at Gemini at Work 2026, designed to combine research, writing, data analysis and coding. Several of the capabilities shown remain in early access, according to the official product page.

Delegating multi-step tasks that run in the background, using Gemini on mobile and desktop, and choosing third-party models are not yet available to every customer. Google says access will expand, but the product page does not provide a date for these capabilities to become generally available.

Businesses can contact Google Cloud sales or try the Gemini Enterprise Plus edition free for 30 days. This trial of the existing edition does not automatically provide access to every newly announced capability. Google says business customers' data, including prompts and outputs, is not used to train its models.

Three people working together with laptops and documents at an office desk, viewed from above
Mikhail Nilov

Connections to workplace applications and spending limits

The official connector catalogue separates integrations by Gemini Enterprise edition. It lists Gmail, Drive and Calendar alongside tools such as Outlook, Teams, Slack, Jira and GitHub. Supported operations vary: some connectors search and summarise information, while others also allow records to be changed or tasks to be managed.

For Standard and Plus editions, the catalogue includes BigQuery, Cloud SQL and Cloud Storage, as well as Salesforce and ServiceNow. Custom connectors and connections to Model Context Protocol servers are also provided for. This list describes the editions' existing integrations; it does not announce general availability for every capability of the new agent.

In its presentation, Google describes an agent that can continue working in the cloud after a laptop is closed and coordinate specialist agents. The announced architecture includes agents' own identities, records of their actions, isolated execution in Agent Sandbox and enforcement of organisational policies through Agent Gateway.

Spending controls build on a set of tools announced separately on August 26. Google allows monthly project limits to be set in Cloud Billing Console. When a limit is reached, the agent's API calls are temporarily paused. Administrators can resume work or allow additional consumption, billed according to usage.

The financial documentation describes alerts at 50%, 80% and 100% of the limit and centralised usage reporting. It distinguishes per-user subscriptions from the consumption-based edition, initially available to selected customers. Deferred execution for some workloads is described as a forthcoming capability, rather than a generally available service.

Infrastructure migration and agents used in telecommunications

Google accompanied the presentation with announcements about Gemini Enterprise customers. Swiss sportswear brand On says it migrated 24 critical services to Google Cloud using a multi-agent architecture. According to the joint announcement, migration time fell from three months to two weeks per service.

The internal team completed 15 migrations entirely in-house, with less than five minutes of planned downtime for each service. Engineers reviewed generated code and authorised production cutovers. On is also a testing partner for the new Gemini agent; the migration figures describe its own project, rather than performance guaranteed to every user.

Orange Spain and Google Cloud also announced an expanded partnership on October 8. The operator says it has deployed more than 1,000 custom agents with Gemini Enterprise. They are used in human resources, IT, sales and customer service, for work carried out across the company's operations in Spain.

The listed tasks include classifying and routing incidents, analysing technical logs, retrieving internal information and managing service requests. The announcement also attributes expansion to tools that let employees without programming expertise build agents. Data is organised in BigQuery and Cloud Storage, while responses draw on semantic search and the company's documents.

Lawyers retain responsibility for reviewing legal documents

Law firm Cooley announced a collaboration with Google to develop an agent that identifies confidential information in documents intended for courts. The tool proposes passages for redaction before filing, including personally identifiable data, technical information and other potentially confidential material.

The announcement states that the agent remains in development. Lawyers make every decision about what information should be redacted before documents are submitted to court. Cooley is participating as a launch partner, helping shape the product and providing feedback from litigation practice. The announcement does not describe a system that autonomously files court documents.

In banking, BNP Paribas entered a five-year partnership with Google Cloud on September 24. The agreement provides for integrating Gemini into LLM@CIB, the Corporate & Institutional Banking division's internal AI assistant, available to more than 65,000 employees. Preparing corporate credit analysis documents is one of the workflows targeted.

The bank says integration follows its own security and data governance rules, and certain categories of information will not be stored in a public cloud. September's announcement refers to planned deployments within CIB. Separately, at Nickel, Gemini already supports an assistant that helps 200 advisers consult internal procedures and operating guidelines.

Access for university research and smaller businesses

Another October 8 announcement concerns the State University of New York. Google Public Sector and SUNY presented an artificial intelligence and high-performance computing platform built on Google Cloud with Gemini Enterprise. Initial access covers nine of the university system's most research-intensive campuses.

The platform brings together storage, computing resources and more than 200 AI models. Research involving weather forecasts, water quality and biomedical data processing is mentioned. SUNY separately offers students and faculty free Google AI courses; these do not automatically extend access to every agent capability.

Google also describes the new Gemini as being in early access for small and midsize business customers. In material addressing those customers, the company presents selection between its own models and Claude according to the task, while retaining working context. The general product page nevertheless lists third-party model choice among capabilities with limited access.

Examples attributed to Gemini Enterprise include Finland's Eficode, which indexes internal documentation for technical support, and Singapore's Koufu, which automates sales reporting. Google also mentions Sunhouse, where teams retrieve archived design projects from Drive. These uses are described separately from capabilities of the new agent that remain in testing.