Gemini agents get company accounts, and staff-built agents need named owners

A manager describes the job, and Gemini creates a coworker agent with its own email address, calendar and Drive. For now only selected customers can use it. When it opens up, it will come at no extra charge with Google Workspace business and enterprise plans. Microsoft already gives agents their own mailbox through Agent 365, so Google is bringing a known idea to Workspace.

The Gemini assistant inside Gmail works as you and can open anything you can open. A coworker agent works under its own account and sees only what is shared with it. So giving it a job means sharing a folder, the same step a team takes for a new colleague.

For example, a procurement team could share one folder of supplier forms with an onboarding agent. New suppliers send their documents to the agent's address. It checks that each tax form is there, chases what is missing and books the setup call. A buyer still approves the bank details, because a false bank account is where supplier fraud usually starts. At 30 new suppliers a month and two hours of chasing for each, the team gets back 60 hours, a week and a half of one person's month.

Before the first agent arrives, ask IT whether its account will count as part of the company when a file is shared with everyone in it. If it does, every file that someone once opened to the whole company is open to the agent too.

About one in six knowledge workers who use AI have begun building their own agents. Spreadsheets spread through companies the same way: someone builds a tool for their own job, colleagues come to rely on it, and nobody else checks it. In 2012, JPMorgan lost $6.2 billion on trades whose risk was measured by a model built in spreadsheets, with figures copied in by hand.

Banks answered with a list of their important spreadsheets, each with a named owner and a regular test. Agents need that rule even more, because an agent acts instead of handing a number to a person. An agent that drafts replies for staff to send needs an occasional look from its owner. One that changes prices or pays invoices needs its owner to read a sample of its work every week.

In 41% of contact centres, a manager reviews fewer than four calls per agent each month, so one bad call can decide a review. A model that scores every call shows the pattern instead, such as an agent who solves problems quickly but explains the refund rules badly every time.

The same scores also rank everyone on the team. Scoring feels like coaching when a low score starts a conversation, and like surveillance when it goes into the file for pay decisions. Staff will learn which one it is from the first time a manager acts on a score.

The experimental app records both the microphone and the sound that the Mac plays, so it works with any video call. During the meeting you type bullet points, and a Google model running on the laptop turns them into full notes from the transcript. The model that searches the transcript is free for others to use, and any app maker may build it in.

So a monthly fee per person for turning recordings into notes will be hard to defend. Notetaker companies will have to charge for what one laptop cannot do, such as searching a whole team's meetings.

Anthropic's new small model costs $0.10 per million input tokens and $0.50 per million output tokens, a tenth of Haiku 4.5's prices. Comparing one invoice with its purchase order and delivery note means reading about 5,000 tokens. For 5,000 invoices, that is $2.50 of reading plus about 50 cents for the model's answers.

Checking a sample by hand then gets a different job. The model reads every invoice and sends the mismatches to a clerk. The clerk also checks a few dozen invoices that the model passed, and that sample now measures how far the model can be trusted.