Google unveils the Gemini agent: a 'digital employee' with its own email that also uses Anthropic models

By Zendoric · October 9, 2026.

Google Cloud unveiled on Thursday, during the "Gemini at Work 2026" event, a single AI agent for the entire company. The announcement came from Thomas Kurian, CEO of Google Cloud, in a corporate blog post setting out his keynote. According to the company, it is "a single universal agent" that answers questions, drafts documents, analyzes data, generates images and multimedia content, and writes and runs code, all from the same text box.

The novelty is not that Gemini reasons or writes: it already did that. What is new is the architecture Google proposes for putting it into daily work. The agent lives inside Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar, with the same memory and the same permissions across every application. It can act as a personal assistant, as a team manager or —the most striking piece of the announcement— as an "agentic coworker": an identity of its own within the company, with an email address on the internal domain (of the @agents.company.com type), a calendar, space in Drive and a presence in the corporate directory. A human colleague can add it to a chat or mention it in a document, and the agent replies, edits and is logged in the version history under its own name, not that of whoever set it up.

To solve each task, Gemini can also create specialized temporary "subagents" on the fly and coordinate them in parallel or in sequence over hours or days, even after the user closes the laptop. And here is the second detail that interests us most: the model that carries out each job does not have to be a Gemini. Google explains that the agent "orchestrates between the Gemini family of models and Anthropic's Claude models today, and other leading models, private and open, in the future", choosing the most suitable one according to the cost and difficulty of each task.

Google cites aggregate adoption figures, with no external audit: close to 500 Google Cloud customers each processed more than a trillion tokens over the past year (the token is the smallest unit of text a language model processes); close to 80% of Google Cloud customers already use some AI product; and, according to the company, close to 90% of Fortune 100 companies use Gemini Enterprise, its corporate agent platform.

The blog also offers concrete cases, again with figures that come solely from Google or from the customers cited, without independent verification. Bradesco, one of Brazil's largest banks, says it has cut the review time for tax and contract documents from an hour to five minutes, and reduced risk inconsistencies by 60%. Japanese insurer SOMPO claims to have built more than 10,000 custom agents for its 34,000 employees —almost one for every three workers— and to have cut the development time for new internal models from a week to a day. Bunnings, Wesfarmers' Australian DIY chain, puts at half a million hours the administrative savings achieved with an internal agent. These are figures from one side, presented as part of a commercial launch: they are best read as the ceiling the vendor is showing off, not as an average any company adopting the tool can expect.

Google is also adding a data layer designed to stop the agent inventing business figures: a "Knowledge Catalog" that settles once and for all the definitions of terms such as "net margin" so that every agent uses the same one; a system for putting into context the 90% of enterprise data that, according to the company itself, remains unstructured (PDFs, scans, audio); and a "borderless lakehouse" that allows data stored in Amazon S3 or Azure to be queried without moving it or paying variable egress fees. Bloomberg Media, according to Google, raised the accuracy of its queries by 63% in the development phase by anchoring the agent to that catalog. Sector specialization —already in testing for banking and legal services, and planned for public administration, healthcare and retail— confirms that Google is targeting regulated verticals, the ones that pay most for reliability and compliance.

The most significant part of the announcement is not technical, it is strategic. That Google should voluntarily integrate Anthropic's models —its direct competitor in the race for the frontier model— into its own agent confirms something we already noted when analyzing other distribution deals between rivals in the sector: the battle for enterprise AI has shifted from the model to the platform that wraps it. If a company entrusts its corporate memory, its permissions and its workflows to Gemini as the orchestration layer, it matters little which engine solves each task underneath: the value —and the customer— stays with whoever controls the text box, the agent's identity and the data catalog. The plumbing wins, not necessarily the smartest model.

The customer figures, self-reported though they are, point in the same direction as our series on AI and employment: the first thing to fall is repetitive administrative work. Going from an hour to five minutes in document review at Bradesco, or saving 500,000 hours of administration at Bunnings, is not a marginal optimization: it is the compression of the reviewer or case-processor role we have been documenting sector by sector for months. In the short term that means less demand for that profile and more pressure to retrain toward tasks of judgment, oversight and client relationships, precisely the ones Google promises to free up. The transition is neither automatic nor free, and the companies that manage it best will be those that train those profiles before dispensing with them, not after.

The figure that should most worry security officers is a different one: an agent with its own corporate email, persistent memory and the ability to create subagents is, to all intents and purposes, a new identity inside the company network. Google knows it, which is why it devotes much of the announcement to identity controls, sandboxing (isolated execution to limit the damage if something goes wrong) and real-time spending caps. But when an insurer deploys 10,000 agents for 34,000 employees, the attack and error surface is no longer measured in user accounts: it is measured in different orders of magnitude. Governing that fleet —who can create an agent, what it sees, whom it represents— will, in the coming years, be a job as important as training the model that drives it.

None of this yet answers the underlying question: whether these agents free up human time for higher-value tasks, as Google promises, or simply compress headcounts without replacing that time with anything. That answer will not come from a corporate blog on the day of a product launch: it will come from what each company does with the hours it says it has saved. That is where to look a year from now, when the figures are no longer supplied by the vendor alone.