Google’s latest enterprise AI product has an unusual feature for something called Gemini.

It can use Claude.

Google Cloud introduced a new Gemini agent on 8 October, describing it as a single work agent that can answer questions, create content, write code, use company tools and carry out longer jobs.

The model underneath it does not always have to be a Google model.

Google says the agent can choose between models in the Gemini family and Claude models from Anthropic, depending on the job. Other private and open models are expected later.

For European companies already testing several AI vendors, that is more interesting than another model launch.

The question is no longer only which model employees should use.

It is whether the company wants one agent sitting above several models, business systems and data sources.

Before rolling that out widely, there are six things worth checking.

1. Decide whether model choice should be automatic

Google separates the agent from the model doing the underlying work.

The company says Gemini can route jobs across its own models and Anthropic Claude models. It can also combine different models inside larger projects.

The pitch is straightforward. A large model is not necessary for every task, so routing simpler work to a cheaper model can reduce cost while reserving more capable models for difficult jobs.

That is useful, but it creates a governance question.

Who decides which model is acceptable for which type of work?

A marketing team might be comfortable using several models for:

  • first drafts;
  • campaign summaries;
  • spreadsheet analysis;
  • market research;
  • meeting preparation.

The same company may want stricter rules for:

  • customer data;
  • legal documents;
  • pricing decisions;
  • unreleased products;
  • regulated information.

Do not make “best model for the task” the only rule.

Add your own rule for which models are allowed to see which data.

2. Map the systems Gemini can reach

Google is positioning Gemini as more than a chat window.

It can connect to company tools and work with information across them.

Google lists integrations and connectors for systems including:

  • Google Workspace;
  • Microsoft Office;
  • Microsoft Teams;
  • Slack;
  • Confluence;
  • Git;
  • Jira;
  • Salesforce;
  • ServiceNow;
  • BigQuery;
  • Databricks;
  • Postgres;
  • Snowflake.

It can also connect to MCP servers.

For marketing and ecommerce teams, that could put campaign documents, sales records, analytics data, customer-service information and internal planning material within reach of the same agent.

That makes connector access one of the first things to audit.

Create a simple inventory:

System Data available Read access Write access Owner
Drive Campaign files Yes Limited Marketing ops
Salesforce Lead records Yes No Sales ops
BigQuery Performance data Yes No Data
Slack Internal discussions Selected No IT
Jira Project tickets Yes Limited Operations

The important part is not how many connectors can be switched on.

It is how few the agent actually needs.

3. Check what Gemini remembers

Google says the agent can maintain context across devices, sessions and longer-running jobs.

It describes several kinds of memory, including information about the current task, accumulated knowledge, remembered procedures and a history of previous work.

That can make the system more useful because employees do not have to explain the same project every time.

It also gives administrators another set of questions to answer.

What should persist?

For how long?

Who can see it?

Can an employee inspect or correct it?

What happens when someone changes teams or leaves the company?

What happens when a project includes confidential client material that should not become useful context for unrelated work?

A marketing team may like the idea of an agent remembering brand guidelines and campaign history.

A client services team may be less comfortable if information from one account appears in another context.

Memory should therefore be reviewed as a data-control feature, not just a convenience feature.

4. Review permissions before giving the agent jobs

Google says Gemini can work as a personal assistant, a team member or a more persistent coworker-style agent.

It can also create or coordinate temporary sub-agents for larger tasks.

That increases the importance of permissions.

A person opening a document is easy to understand.

An agent opening several systems, creating sub-tasks and continuing to work after the employee closes the laptop is a different operating pattern.

Before handing over real work, test:

  • which files the agent can open;
  • which systems it can query;
  • whether it can change records;
  • whether it can send messages;
  • whether it can create files;
  • whether it can trigger another tool;
  • what happens when an action is denied;
  • what gets logged.

Start with read-heavy work.

A research task that reads approved sources carries less risk than an agent that can edit a live CRM record or send a customer message.

5. Put a hard limit on spending

One of the more practical parts of Google’s announcement is cost control.

Google says Gemini can route work to different models based partly on cost and performance.

It also supports project-level spend caps.

When a project reaches its configured limit, Google says the agent can be paused until someone decides to resume it.

This is important because agentic work can be harder to price than a single prompt.

One employee request may trigger:

  • several model calls;
  • searches;
  • tool use;
  • sub-agents;
  • code execution;
  • longer-running background work.

A monthly licence price does not necessarily tell the whole cost story.

For an initial deployment, track:

Cost per completed task

Not just tokens.

If a competitor-analysis job costs €4 and saves an hour, that may be reasonable.

If a simple weekly report triggers €40 of agent activity, routing or workflow design may need work.

Also split spend by project or department where possible.

That makes experimentation easier to stop before it quietly becomes permanent infrastructure.

6. Check where data is processed before calling it a European setup

European companies should be careful with one assumption.

Using Gemini Enterprise does not automatically answer every question about where every underlying model processes data.

Google Cloud supports regional and multi-regional services, and its documentation shows that model availability and processing arrangements can differ by model and location.

That becomes especially relevant when the agent can choose models from more than one provider.

Before approving a workload, record:

  • the model that may be used;
  • the Google Cloud region or endpoint;
  • where model processing occurs;
  • retention settings;
  • logging;
  • connected systems;
  • subprocessors where relevant;
  • whether the workload contains personal or confidential data.

Do this for the actual configuration being deployed.

Do not rely on the words “Google Cloud”, “Gemini” or “European region” as a substitute for checking the service settings.

The Workspace connection will make adoption easier

Google is also bringing the same agent directly into Workspace.

Gemini can work inside Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar while keeping the same context, skills and controls.

That lowers the barrier for employees because they do not have to move into a separate AI application for every task.

A marketer could research a market, analyse data in Sheets and prepare a presentation without repeatedly rebuilding the context.

That convenience is also why rollout controls matter.

When an AI tool sits inside software people already use every day, experimentation can spread faster than a formal pilot.

Administrators should know which teams have access, which connectors are active and which tasks are moving from assistance to autonomous action.

European companies are already part of Google’s rollout story

Google’s own launch material names several European organisations using Gemini Enterprise.

It says BNP Paribas is deploying Gemini Enterprise within an internal assistant used by more than 65,000 employees.

Commerzbank is using it for document quality assurance.

Lloyds Banking Group is using Gemini-based infrastructure for internal AI development.

Google also cites Nokia, UK-based Arden University and Swiss sportswear company On.

These are useful examples of the kinds of workloads Google is targeting.

They should not be treated as independent proof of the results Google reports.

Customer case studies tell us what companies are trying. They do not remove the need to test the same workload inside your own systems.

Run a small rollout before a company-wide one

A useful pilot does not need hundreds of employees.

Choose one team and three repetitive jobs.

For example:

  1. weekly campaign reporting;
  2. market research;
  3. internal briefing preparation.

For each job, record:

  • time taken before Gemini;
  • time taken with Gemini;
  • models used;
  • connected systems;
  • human corrections;
  • failed actions;
  • cost;
  • sensitive data touched;
  • final output quality.

Run the same jobs for several weeks.

You will learn much more from that than from counting how many employees opened the agent.

The useful adoption metric is not logins.

It is whether a specific job became faster, cheaper or more reliable without creating a new data or approval problem.

This is not a Google Search change

The name Gemini can create confusion for search teams.

This announcement concerns Google’s enterprise work agent.

It is not an announced change to:

  • Google Search rankings;
  • AI Overviews;
  • Google AI Mode;
  • Google Ads auctions;
  • Merchant Center visibility.

Do not turn enterprise-agent adoption into an SEO claim.

For Search developments, NEMO tracks those separately in Search & AI.

For businesses evaluating AI tools and platform changes, see Platforms & Regulation.

The interesting part is not Gemini versus Claude

Google putting Anthropic models inside the Gemini agent may look like a model rivalry story.

For businesses, the more important change is that model choice can move behind the interface.

Employees may deal with one agent while the system decides which model, tool and company data should handle the job.

That can make AI easier to use.

It also means companies need better control over permissions, data, memory and cost.

Before rolling Gemini out across a European organisation, answer six questions:

  1. Which models can handle which data?
  2. Which company systems can the agent reach?
  3. What does it remember?
  4. What can it change?
  5. What can it spend?
  6. Where does the work actually run?

If those answers are clear, the model picker becomes useful.

If they are not, adding more models only adds more uncertainty.

Primary sources