Google Cloud CEO Thomas Kurian introduced the Gemini agent at Gemini at Work 2026 and named the sports brand On among its early testers. Google calls the Gemini agent a single, universal agent for work.
On, the sportswear brand headquartered in Zurich, used AI agents to migrate 24 core services to Google Cloud. Migration time fell from 3 months to 2 weeks per service. An external vendor had initially quoted $500,000 to migrate a fraction of the microservices. Google Cloud and On announced these figures on 8 October 2026.
Kurian backed the announcement with scale figures. Nearly 500 Google Cloud customers each processed more than 1 trillion tokens in the past year. Nearly 90% of the Fortune 100 use Gemini Enterprise.
In brief: The Gemini agent combines chat, autonomous task execution, media generation and coding in 1 agent. Press reports describe it as in private preview, and Google has not published prices. Constellation Research reports that general availability is expected around the end of October or early November 2026. The DACH companies named are On, Commerzbank, Deutsche Telekom, Deutsche Bank, Berenberg and Siemens. Open points for regulated buyers are portability of agent memory and skills, data residency in Switzerland, retention controls and cost per task.
What Google Announced at Gemini at Work 2026
According to Google, the Gemini agent runs in Google’s cloud, and tasks that take hours or days continue after the user logs off. It keeps 4 kinds of memory: session, semantic, procedural and episodic. Semantic memory is a knowledge base built from documents and conversations. Episodic memory records the agent’s past actions.
For multi-step jobs, the Gemini agent creates temporary sub-agents. Coworker agents hold a lasting role and get their own Workspace account with email address, calendar and Drive. A router selects the model for each task. The pool covers Google’s Gemini models and Claude models from Anthropic, with further models announced.
The Gemini agent is available on web, mobile, desktop and the command line, and inside Google Workspace, Microsoft 365 and Slack. Connectors link it to enterprise systems and to any Model Context Protocol (MCP) server. Each agent gets its own attested identity with least-privilege permissions. Actions are logged under the agent’s name, and all traffic passes through a policy gateway.
Cost controls for the Gemini agent consist of smart routing and hard spend caps per project. When a cap is reached, the agent pauses until someone resumes it. Industry editions are in preview for Financial Services and Legal. Government, Healthcare and Retail are announced as coming soon.
The DACH Companies Named in the Announcement
As was mentioned, On is rolling out Gemini Enterprise to every employee and has tested the agent’s dynamic model selection.
Commerzbank uses Gemini Enterprise for document quality assurance. Google reports that manual review work at the bank fell from 20 hours to 1.
Deutsche Telekom built a unified data lakehouse across more than 40 legacy on-premises systems. Google reports that its teams now work 10 times faster while addressing European data sovereignty requirements.
Deutsche Bank is named as a user of the Financial Services edition. Google’s release of 25 August 2026 describes the bank as a key design partner for that edition’s Financial Research agent. Berenberg and Siemens are listed as customers without further detail.
All figures in this section are reported by Google or by the customers themselves, without published methodology or independent verification.
What the Gemini Agent Means for DACH Entrepreneurs
The DACH companies named at Gemini at Work 2026 send a clear signal. Commerzbank, Deutsche Telekom, Deutsche Bank, Berenberg, Siemens and On do not run pilots for publicity. These are regulated, risk-aware organizations, and they have moved Google’s agents into document review, data infrastructure and company-wide rollouts.
For an entrepreneur in Switzerland, Germany or Austria, the question has changed. It is no longer whether agentic AI is ready for serious business. It is which processes to hand over first, and on whose infrastructure.
The barrier to entry is falling at the same time. Google has said it does not intend to charge a per-seat license for coworker agents, and pricing is expected to be consumption-based. A 10-person company can now give an agent a defined role, such as preparing offers or answering supplier requests, without adding a hire. On’s case shows the economics: work an outside vendor priced at $500,000 for only part of a migration was done by a small internal team with agents.
The same announcement also shows where the risk sits for smaller companies. Google’s data residency documentation for Gemini Enterprise lists no location in Switzerland or Germany. The closest option is the EU multi-region. The announcement does not say where the agent’s memory is stored or where requests routed to Claude models are processed. The agent builds a lasting memory from documents and conversations, so client and employee data becomes part of that memory. That brings revDSG and GDPR duties with it: access requests, impact assessments, and deletion. Google has not described controls for any of these. A large bank negotiates these terms with its own legal team. A 20-person company usually accepts the default settings. Lock-in deserves the same attention. Model choice stays open, but skills, memory, and policies accumulate inside Google’s platform, with no stated export path. For a small company, the most valuable thing an agent builds is its knowledge of how the business works. Decide early where that knowledge lives.
Where Lab51 Closes the Gaps
Lab51 builds custom AI agents for the processes where these gaps matter most: client-facing work, sensitive data, and the knowledge that defines your product. Where Google’s documentation lists no Swiss location, Lab51 offers Swiss-hosted models, with self-hosted options for regulated industries such as pharma, finance and government. Where the announcement gives no export path for agent knowledge, the Lab51 client owns the model and the knowledge base. Google’s customer figures are self-reported. Each Lab51 agent is instead tested against a benchmark dataset of must-get-right questions, approved by your team before launch. FADP and GDPR compliance is built into every layer of the system. Many companies will run both: a Workspace agent for scheduling, drafting, and internal research, and a custom agent for regulated and client-facing processes. The same benchmark dataset tells you which route fits each process.
Planning your first AI agent, or unsure which processes belong on which platform? Contact Lab51 for a consultation. We’ll map your use case, your data, and your compliance requirements, and show you where a custom build pays off.
