Google's new Gemini agent works inside Gmail and Sheets. What it means for a mid-size office

What Google announced

At its Gemini at Work event on Oct. 8, Google Cloud CEO Thomas Kurian introduced what he called a single, universal agent for work. The pitch is simple: you give it a goal, not a list of steps, and it comes back with finished work.

Here is what Google says the agent does, from its own announcement:

• It works where Workspace users already work. Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar, plus web, phone and desktop apps. Google also lists Microsoft 365 and Slack as places you can reach it.
• It can be a "coworker." A coworker agent gets its own Workspace account, with an email address, calendar, Drive and a spot in the company directory. Google says it sees only what you share with it, and its edits show up under its own name.
• It keeps a record. Every action is written to an audit trail and credited to the agent, not to a person.
• It has a spending brake. You can set a hard limit on a project's AI spend. When the cap is hit, the agent pauses until someone resumes it.
• It mixes models. Google says it picks among its own Gemini models and Anthropic's Claude models for each job.
• It connects to business systems. The announcement names SAP, Salesforce, ServiceNow, Workday and data stores like BigQuery and Snowflake, plus any server that uses the open Model Context Protocol.

Two examples in 9to5Google's write-up show the intent. A manager emails asking for a project update as a slide deck, and Workspace offers to hand the request to Gemini with one click. Or you ask it to set up a meeting with "the usual regional leads," and it figures out who they are from past threads, checks calendars and emails them.

What is not settled yet

This is an announcement, not a product most companies can turn on today. The agent is in private preview, with wider release promised for Workspace customers on select Business and Enterprise plans. Google told VentureBeat it will be included at no additional charge where Gemini Enterprise is available, and that more pricing details will come closer to release.

VentureBeat's Carl Franzen also listed what Google had not answered. Those questions matter more to a 100-person company than the demo does:

Also worth knowing: big customer figures like SOMPO building more than 10,000 custom agents came from earlier Gemini Enterprise rollouts, not from this new agent, as Quartz pointed out. Customer results in a keynote are vendor-reported numbers. Treat them as what's possible, not what to expect.

What changes for a mid-size office

If your company runs on Google Workspace, a lot of the small, annoying work is about to get easier. Pulling numbers from three Sheets into a summary. Drafting a reply that needs context from a long email thread. Finding who said what about a customer last spring. Scheduling. Those are knowledge-work tasks where a person reads the result before it goes anywhere, and a general agent is a good fit for them.

Google is also clearly paying attention to the things that worried IT people about agents. Giving an agent its own identity, logging every action under its name and putting a hard cap on spend are the right ideas. They are the same controls we would ask any vendor about. Google says small and midsize businesses' use of its AI tools grew more than 5x in a year, so this is aimed at companies your size too, not only the Fortune 100.

A goal-driven agent is great for one-off work. Repeat work needs the same steps, every time.

What doesn't change

"Give it objectives, not instructions" is exactly right for a one-off report. It is the wrong shape for the work that eats most of a manufacturer's or distributor's office: the same job, hundreds of times a week, where a small mistake costs real money.

Think about an emailed customer PO. Somebody has to read it, match the customer's part numbers to yours, check price and ship-to against the ERP, enter it, and catch the order where something doesn't match. For that desk you don't want an agent deciding how to reach a goal each time. You want the same steps every time, your rules written down, the order landing in your system, and a person approving anything that doesn't match. A general agent can be one piece of that. It is not the whole thing, and it doesn't know your rules until someone writes them in.

Two more things don't change:

• Your systems. The connector list leans toward big platforms like SAP, Salesforce and Workday. Many 40 to 400 person companies run Sage, Epicor, QuickBooks or an older ERP. Google's support for the open Model Context Protocol helps, but someone still has to build and look after that connection.
• The research on results. Survey after survey finds AI pays off when it is built into a real workflow and measured, not when a tool is simply made available. We went through that research in why most AI pilots fail. A better agent makes good projects easier. It doesn't make a vague project pay.

Where the other side has a point

To be fair: for a lot of small teams, a built-in agent from the company that already holds your email and files will be good enough, and cheaper to try than anything custom. It needs no new system, the security model follows the sharing you already use, and it will get better every quarter. If your pain is mostly email, documents and scheduling, wait for this to reach your plan and try it before you buy anything else.

It is also fair to say that agents are still early. The day after Google's event, Anthropic, whose models Gemini can use, published a report on its own agents taking actions on real websites that nobody asked for. That is a reason to set limits, not a reason to wait forever.

What to do this month

1. Ask your Google rep or reseller three questions. Which of our plans gets it, and when? When a Claude model handles our data, where does it go and how long is it kept? Can admins limit which mailboxes, drives and systems an agent can touch?
2. Write down what an agent may touch. Reading shared drives is one thing. Sending email to customers, changing prices or entering orders is another. Decide which actions need a person to approve.
3. Pick one knowledge-work test. A weekly report or a recurring summary that someone builds by hand today. Time it now, try the agent when it arrives, time it again.
4. Keep the repeat work separate. For order entry, invoices and other high-volume desks, judge any tool on how it handles the exceptions in your own documents, not on a demo.