Essay

An AI coworker takes one sentence to hire. Who fires it?

Google is giving AI agents their own email address and a place in the company directory. Creating a digital coworker will soon take a single sentence. The hard part comes after: who keeps track of them, and who switches them off?

October 12, 2026 · Levy van Bremen · Lees in het Nederlands

On October 8, Google showed what an AI coworker looks like. At its Gemini at Work event, the company introduced the Gemini agent: an AI assistant that does not just answer questions but carries out work on its own, from drafting a document to an assignment that runs for days. Software that finishes tasks on its own like this is called an agent.

Google calls the most striking part a 'coworker agent'. You describe the role you need, and Gemini creates that coworker. It gets its own account with an email address, a calendar, its own storage and a place in your company directory. Colleagues hand it work in a group chat, the way they would with a person. According to Google, it acts under its own name rather than yours, and it sees only what you share with it.

The feature is still in a private preview with a select group of customers. According to the tech publication VentureBeat, Google says the agent will be included at no extra charge for Gemini Enterprise customers, including through existing Google Workspace business subscriptions. Google promises more pricing details, including for long-running work, closer to the wider release.

My argument: once an AI coworker can be created in a single sentence, the scarce part is no longer creating it but keeping track of it. If you plan to use this, you need for agents what you already have for staff: an owner, a purpose and a moment when they leave again.

Hiring gets easy, keeping track does not

Hiring an employee is an event in any business. There is a vacancy, an interview, a contract, a laptop and an account. Precisely because it takes effort, everyone knows who works there and who they report to. And when someone leaves, their account is closed and the key comes back.

An AI coworker skips all of those steps. It asks for no salary and no desk, and if your subscription includes it, not even a separate purchase. Then the usual thing happens when something becomes easy and cheap: you get a lot of it. In the same announcement, Google mentions the Japanese insurer SOMPO, which built more than 10,000 custom AI agents across its 34,000 employees, for tasks such as searching and summarizing documents. That is roughly one agent for every three or four people.

Security specialists already see what that means. The Cloud Security Alliance, an organization in the field of digital security, summarized its research in June. In 68 percent of organizations, it is not clear what an agent did and what a person did. Only 28 percent can trace an agent's action back to a responsible person across all their systems. And among organizations with more than a hundred approved agents, more than half reported at least as many agents running that had never been approved.

Those figures come from large organizations with their own security department. A company of twenty people has no such department, but it will soon get the same button.

What its own account does and does not solve

Google has thought much of this through. Giving an agent its own identity is an improvement on what often happens today: an AI tool that logs in with an employee's account, so that afterwards nobody can tell who did what. According to Google, every action goes into a log, under the agent's name. You can set a hard spending limit per project; once it is reached, the agent pauses. And access follows the same sharing settings your team already uses.

But read carefully what Google promises: every action is attributed to the agent rather than to a person. Technically, that is clean. For your organization, it is a gap. You cannot hold an agent to account. It does not feel responsible when a quote with the wrong price goes out, and it does not notice that the project it was created for finished three months ago. A log tells you what happened, not who should have been paying attention.

Security vendors see that gap too. RSA, a company that secures accounts and access, recently presented a product for registering and managing agents. RSA summed up the problem briefly: agents skipped every process built for people, with no registration, no owner and no accountability. In that product, every agent gets a designated owner and a lifecycle status, and when it is retired, its access is revoked.

HR policy for software

The solution is less new than it sounds. Every business already knows how to bring people in, guide them and let them go. Much of that can be reused.

That last point is the one most often forgotten. The Cloud Security Alliance warns that agents are taken out of use without being formally signed off. When an employee leaves, you close their account, revoke their access to the books and ask for the key back. When nobody uses an agent anymore, usually nothing happens.

How I do it myself

Under the name VB Tech, I build software and AI agents on my own. For my own work I run a system of thirteen services and a number of AI agents. They collect the numbers for my apps, keep an eye on the costs of my database and check whether my apps are still in the App Store. Together they carry out a task more than six hundred times a day.

For each service I see a color: green, yellow or red. Problems reach me bundled together via Telegram. And in early October I deliberately switched one off, rather than leaving it on the list.

With thirteen services and one owner, that overview fits in a single glance. It is a different story when twenty colleagues each create a few, because it costs nothing and takes one sentence.

The strongest counterargument

You could say this is overblown. Agents are software, and you manage software as software: the IT provider keeps a list and that is that. Treating them as staff makes them bigger than they are.

There is something to that. An agent has no rights, no feelings and no notice period. But Google itself chooses to present it as staff: an account, an email address, a place in the directory and, according to the announcement, governed like an employee. That puts it in the same systems and the same habits as your people. Colleagues send it work, share documents with it and start relying on it. A list kept by your IT provider does not capture that behavior. Besides, most businesses have no IT department of their own. What they do have is experience with people who come and go.

What you need to decide now

The Gemini agent is not yet available to everyone, and Microsoft and OpenAI are working on similar AI coworkers. So you have some time. Use it for one decision, before the button appears in your subscription: who in your business may create an AI coworker, and what they must record when they do.

Keep it small. One list, with an owner for each agent, a purpose, what it may see, a budget and a date on which someone decides whether it stays. If nobody can fill in a name under owner, the agent does not get created. That sounds strict for something that costs nothing. But precisely because it costs nothing, the rule is needed.

Sources

Is there work in your business that comes back every week? That is what I build software for.

Book a call →