On 8 October 2026, Google Cloud unveiled Gemini Agent, a single agent for work in companies. You no longer give it instructions one by one: you hand it a goal, and it organises itself to reach it in Gmail, Docs, Sheets or Salesforce, sometimes over several days.

‍

The ambition is real. But as of 12 October, the agent is open to only a handful of customers, with no price and no date, and nothing yet says when it will arrive in France. 😅

‍

At AI Sisters, we train teams to work with AI, tool by tool and role by role. Here is what Gemini Agent really does, what remains unclear, and how to prepare your teams now. Let’s go 👇

‍

“You give the agent goals, not just instructions,” summed up Thomas Kurian, CEO of Google Cloud, at the launch.

‍

In brief

‍

  • Announced on 8 October 2026 by Google Cloud, Gemini Agent brings conversation, autonomous tasks and coding together in a single window.
  • It works alone and for a long time: the task carries on in the cloud with your computer closed, with sub-agents running in parallel.
  • It acts in your tools: Gmail, Docs, Sheets, Calendar, but also Salesforce, SAP, Slack or Microsoft 365.
  • “Coworker agents” make their debut, with their own email address, calendar and name in the company directory.
  • It is not available yet: private preview, no published price, no firm date, nothing on France.
  • Preparation, however, starts now: choosing the tasks to delegate, framing the data, training the teams.

‍

‍

What exactly is Gemini Agent?

‍

Gemini Agent works on a different logic: it receives a goal and chains the steps together itself to reach it.

‍

Google describes it as “a universal agent for work”. It is available on the web, on mobile, on Windows and Mac, and directly in Workspace with the @Gemini mention. It also works in Slack and Microsoft 365.

‍

What Gemini Agent can do, according to Google
CapabilityIn practiceWhat you need to know
Receive a goalYou describe the expected result, the agent plans the stepsNo systematic human approval described by Google
Work for a long timeThe task carries on in the cloud, computer closed, for hours or daysA spending cap can pause the task
Multiply itselfIt creates temporary sub-agents that work in parallelTheir cost is not published
Act in WorkspaceGmail, Drive, Docs, Sheets, Slides, Chat, Calendar, called with @GeminiIt sees whatever your access rights open up to it
Connect to your toolsSalesforce, ServiceNow, Workday, SAP, Jira, Slack, Microsoft 365, BigQuery…Each connection is approved by your administrators
RememberMemory of the session, your preferences and the procedures it learnsThe retention period is not specified

‍

Google’s example: you ask it to organise a meeting with “the usual team”. The agent identifies the participants from your past conversations and meetings, checks calendars, suggests a slot and writes to the external guests.

‍

How Gemini Agent works, according to Google’s description. One: you set a goal, rather than a series of instructions. Two: it plans the steps and chooses the right tools and AI model. Three: it acts in Gmail, Calendar, Drive or Salesforce, and launches sub-agents if needed. Four: it reports back in a tracking inbox, even with your computer closed. Google does not describe systematic human approval before each action: it is up to administrators to set the rights.
From goal to report: the four stages of a task handed to the agent.

‍

💡 In plain terms: assistant or agent. An assistant is a satnav that tells you the way: you drive. An agent is a driver you give the address to: it chooses the route, stops for fuel and lets you know when you arrive.

‍

On models, Google says the agent chooses “the most suitable model” for each task, from the Gemini family and Anthropic’s Claude models. The default model is not specified. Gemini 4 Argon, covered in our article, remains in restricted access.

‍

“Coworker agents”, the real novelty

‍

The newest part of the announcement is the permanent agents, which Google calls “coworker agents”. Each one has a role, an email address, a Workspace account, a calendar, a Drive space and an entry in the company directory.

‍

In practice, you can add a “supplier follow-up” agent to a Chat space, mention it in a Google Docs comment, and find its edits under its name in the version history.

‍

The profile of a coworker agent. An email address such as @agents.your-company.com. A calendar and a Drive, with its own Workspace account. A directory entry: you add it to a Chat space or mention it in a Docs comment. Signed actions, visible under its name in the version history. It only sees what is shared with it.
What a permanent agent receives, just like a new starter.

‍

💡 In plain terms: an agent with an email address. It is like welcoming a very fast intern: they have their badge, mailbox and access, they work on the files you hand them, and everything they do is signed with their name.

‍

⚠️ What Google is not saying yet. VentureBeat notes that Google has not confirmed whether all agents get a mailbox or only permanent agents. The cost of running dozens of agents continuously remains unknown.

‍

Availability, pricing, France: what we know and what we don’t

‍

This is where the announcement parts ways with reality. Gemini Agent has been presented, but is not on sale yet.

‍

Availability, pricing, France: where things stand on 12 October 2026
QuestionWhat we knowSource
Can it be used today?No, except by a selection of customers in private previewGoogle, CIO Dive
When will it be available?“Soon”, according to Google. Late October or early November, according to executives quoted by Constellation ResearchA single source, to be confirmed
How much does it cost?No published pricing. “At no extra cost” for Gemini Enterprise customers, according to a spokesperson quoted by VentureBeatUsage-based billing mentioned by Constellation
With which subscription?Some Business and Enterprise plans, with no precise list9to5Google
Available in France?No source says so, either wayNot confirmed
With which AI model?Gemini models and Anthropic’s (Claude), chosen according to the taskGoogle

‍

To keep control of the bill, Google provides a spending cap per project, set by the administrator. When it is reached, the agent pauses until someone restarts it with a click.

‍

Be careful not to confuse it with Google’s consumer agent, Gemini Spark.

‍

Gemini Agent and Gemini Spark: two products not to confuse
Gemini AgentGemini Spark
For whomCompanies, business accountsIndividuals, AI Pro or Ultra subscribers
StatusPrivate preview, announced on 8 October 2026Experimental, launched in May 2026
In FranceNot specifiedExcluded from the European Economic Area
Confirmation before actingNot described in general termsYes: sending messages, purchases, forms, editing shared documents

‍

Discover our AI training & book a call

Prepare your teams to work with agents, on their real tasks.

📚 See our AI training📞 Book a discovery call

‍

What it could change for three roles

‍

Google cites many customers, including BNP Paribas. But Futurum and VentureBeat note that most of these results come from earlier Gemini Enterprise deployments, not from the new agent. The figures are Google’s, with no independent measurement.

‍

Results cited by Google, to be read with caution
CompanyUseAnnounced result
CommerzbankDocument reviewFrom 20 hours to 1 hour
BradescoDocument reviewFrom one hour to 5 minutes
NokiaIncident resolutionUp to 80% less time
BNP ParibasInternal assistant for more than 65,000 employees (credit memos)No productivity figures
Orange SpainBuilding custom agentsMore than 1,000 agents

‍

Three roles, three first tasks. Finance: summarising an annual report, with figures checked one by one and rights limited to the strict minimum. Legal: gathering the clauses of one contract type on a closed matter, read-only at first. Operations and customer service: sorting incoming requests, with a person approving each reply. Outside the private preview, none of these uses can be tested today.
The first task to delegate, role by role.

‍

For a finance team: preparing the files, not making the decision.
Google is launching in preview a specialised version with more than 50 business skills and data from FactSet, LSEG and S&P Global. First task to delegate: summarising an annual report, with figures checked one by one. Watch out: an agent that reads your financial data must have rights limited to the strict minimum.

‍

For a legal team: research and review, not signing.
The legal version, also in preview, mirrors the access rights and ethical walls of law firms’ document tools, such as NetDocuments and iManage. First task to delegate: gathering the clauses of one contract type on a closed matter. Watch out: an agent’s mistake on a shared document commits the team, hence the value of starting with read-only tasks.

‍

For operations and customer service: tasks that span several tools.
A request arrives by email, the agent checks the record in the CRM, prepares a reply and schedules a callback. First task to delegate: sorting incoming requests, with a person approving the replies. Watch out: long, multi-tool tasks are precisely the ones whose reliability has not yet been independently measured.

‍

Security and data: the questions to ask before you start

‍

Google highlights several safeguards, but some answers are still missing.

‍

✅ What Google provides. Each agent has a verified identity and access kept to a minimum. Connections to external tools are approved by your administrators, and every action is logged under the agent’s name. According to Google, Workspace does not use your data to train its models without your permission.

‍

⚠️ What still needs clarifying. Google has not said what happens to your data when a task is handed to a Claude model. The Gemini Enterprise documentation provides for hosting in the European Union, but not for every feature or every model. And earlier versions of Gemini have already been tricked by instructions hidden in emails or calendar invitations.

‍

In July 2026, the CNIL (the French data protection authority) published a note on agentic AI with the CIANum. It recommends access limited to the strict minimum, real-time monitoring and human approval for important decisions.

‍

How to prepare your teams now

‍

Gemini Agent is not available yet, but the companies that benefit from it fastest will be those that have prepared the ground. Here are five steps, useful whichever agent you choose.

‍

Five steps to prepare, whichever agent you choose. One: list five repetitive tasks that span several tools. Two: rank them by risk, from read-only to irreversible action. Three: map your data, because an agent inherits the rights it is given. Four: write your rules. Five: train your teams to delegate.
Five useful steps, whichever agent you choose.

‍

  1. List five repetitive tasks that span several tools: follow-ups, meeting notes, sorting requests, preparing meetings.
  2. Rank them by level of risk: read-only, reversible action, irreversible action. Always start with the first category.
  3. Map your data: who has access to what in Drive and in your business tools. An agent inherits the rights it is given.
  4. Write your rules: which data can be entrusted to an agent, which actions require human approval, who monitors. Our guide to integrating AI into a company sets out this approach in detail.
  5. Train your teams to delegate: setting a clear goal, breaking down a task, checking a result. It is at the heart of our role-based approach to training.

‍

Already on Workspace? Ask your Google contact whether your plan will be included, and where your data will be processed.

‍

Book your 30-minute call

A no-obligation conversation to build your AI roadmap.

👉 Tell us about your AI project 👈

‍

Key takeaways

‍

  • Gemini Agent moves from assistant to agent: you hand it a goal, and it chains the steps together in your tools.
  • Coworker agents are the real novelty, with their own account, calendar and signed actions.
  • It is not available yet: private preview, no price, no firm date, no commitment on France.
  • The customer results cited are Google’s, and most concern earlier tools.
  • Prepare now: tasks to delegate, written rules, trained teams.

‍

Sources