Give It Its Own Login

Pull up the activity log on whatever AI tool you trust most and look at who did what last week.

I'll bet it says you did. You sent that follow up. You moved that file. You updated that deal stage at 2:14 in the morning while you were asleep.

You didn't, of course. Your agent did, wearing your login like a borrowed jacket. And as long as everything goes fine, nobody cares whose name is on the log. The day something goes sideways, that's the first question anyone asks, and "well, it was sort of me, but also the robot" is a terrible answer to give a client.

Last week the two biggest office software companies on the planet said, more or less out loud, that this has to stop.

On Thursday at its Gemini at Work event, Google announced what it calls coworker agents. Each one gets a real Workspace account. Its own email address on a separate agent subdomain, its own calendar, its own Drive, and a listing in the company directory. Colleagues can add it to a Chat space, @mention it, or tag it in a comment on a Doc. Everything it does gets logged to an audit trail under the agent's name, not the name of whoever switched it on. Google also added spend caps per project, so when an agent hits its budget, it pauses until somebody clicks resume.

One day earlier, Microsoft made its Execution Containers generally available. That's a way to run an agent inside a sealed box on a Windows machine, with a written policy that spells out which files, which network addresses and which programs it can touch. The policy sits outside the agent, so the agent can't talk itself into more access halfway through a job. And over on the identity side, Microsoft already requires every agent account to have at least one named human sponsor.

Put those side by side and they're saying the same thing. The big players have stopped treating agents as a feature you click. They're treating them as staff, with a badge, a desk, a manager, a spending limit and an exit interview.

You don't need Google's enterprise pack to work this way. You need about two hours and the willingness to treat your agent like the new hire it actually is.

Three problems show up when an agent runs under your account, and they get worse the more useful the agent becomes.

You can't tell your work from its work. When a client asks why they got two invoices, you're digging through a log where every line has your name on it. Was it you, the agent, or the agent running a step you approved three weeks ago and forgot about? Nobody can say, including you.

It inherits everything you can reach. Your account can see payroll, your bank, every client folder and every private message you've ever sent. Your agent only needs the support inbox and the shared calendar. But when it borrows your login, it gets the whole keyring. Every permission you have becomes a permission it has.

You can't fire it cleanly. Shutting it off means changing your own password and breaking three other tools. So people don't. The old agent keeps running long after anyone's paying attention.

Give it its own login and most of that goes away. You can see what it did, it starts with access to nothing, and you can shut it off without breaking your own stuff.

Before you create a single account, write one paragraph that answers four questions. What is this agent responsible for? What is it never allowed to do? Who does it report to? How will you know it's doing a good job?

Here's a real one from a home services company I helped last month:

That paragraph does more work than any setting you'll configure later. It tells you what accounts to create, which permissions to grant and what to measure. And when the agent does something odd, either the description allowed it and needs tightening, or it didn't and you've got a bug to chase.

If you can't write the paragraph, you're not ready to build the agent.

Give the agent its own account in every system it needs. Not yours, and not one the whole office shares.

Email. Create a dedicated address like [email protected]. If you're on Google Workspace or Microsoft 365, that's a regular user account. Yes, it might cost a seat. Pay for it. It's cheaper than one bad week of cleanup. Put "AI" in the display name so nobody on the receiving end mistakes it for a person.

Calendar. The agent gets its own calendar, and you share the specific calendars it needs to see or book on. It doesn't get yours.

Files. Create one folder the agent owns, plus shared access to exactly the folders in the job description. If the job description says "approved FAQ," share that document and nothing else from your drive.

CRM. Add the agent as its own user. In Go High Level that means a separate user with a custom role that only sees the pipelines and contact fields it actually works with. Every note and stage change it makes will now show its name.

Automation platform. In Make.com, create the connections the agent uses with the agent's credentials, not yours. Name them clearly, something like "Gmail, Inbox Agent," so six months from now you know exactly which scenario is acting as whom.

One naming rule makes all of this easier: the agent's name goes in front of everything it owns. Inbox Agent, Inbox Agent Calendar, Inbox Agent Drive, Inbox Agent Make connection. When you're auditing later, you search one word and see the whole footprint.

Microsoft got this exactly right by requiring a named human sponsor for every agent. Copy that rule.

Every agent has one person whose name sits next to it on a list. That person reviews its work, answers for its mistakes and decides when it gets more access. Pick a backup too, because the sponsor will eventually take a vacation, and agents don't pause just because their manager went fishing.

The sponsor's job is light but real. A fifteen minute review once a week. Read ten random actions from the log. Check the success metric from the job description. Decide whether anything needs tightening.

The thing you're avoiding is the orphan agent. Someone builds it, the builder leaves, and now there's a bot emailing customers that nobody is watching. Every business I've audited with more than three agents running has at least one orphan. Usually they find out when a customer forwards them something strange.

A login tells you who did it. A box limits what it can do in the first place. Get both.

Here's the minimum box for a small business agent:

Read and write are separate decisions. List everything the agent touches and mark each one read only or read and write. Most agents need to read far more than they write. The Inbox Agent reads the FAQ but should never edit it.

External actions get a gate. Anything that leaves the building, an email to a client, a text, a post, a payment, either goes through an approval step or is limited to a template you've already approved. Internal actions like tagging, filing and drafting can run free.

Spend has a ceiling. Google's new project spend caps pause an agent when it hits its limit. You can get the same effect on any stack. Put the agent on its own API key with a monthly limit set at the provider, and add a counter in your automation that stops the run once a daily number of actions is reached. If your agent normally sends forty replies a day and suddenly tries to send four hundred, you want it stopped at fifty, not discovered at four hundred.

Learning mode first. Microsoft's containers have a learning mode that blocks anything outside the policy and writes a report of what the agent tried to do. Steal the idea. For the first week, run the agent with every external action routed to a draft folder instead of sending. Read what it would have done. Then open the gates one at a time.

With its own identity, the logs mostly write themselves. You still want one place where a human sees the whole picture.

Set up an activity sheet, one row per action: timestamp, agent name, action, the record it touched, and whether a human approved it. In Make.com that's one extra module at the end of each scenario. Ten minutes to add, hours saved the first time someone asks what happened.

If an agent drafts follow ups from calls, keep the transcript next to the draft. Fathom records who said what, so every promise in that email traces back to the moment someone actually made it.

Here's the review prompt I give sponsors for their weekly check. Paste in the last week of rows from the activity sheet along with the job description:

That last line matters. You don't want a cheerful recap of four hundred normal actions. You want the six weird ones.

FROM THE AI NEWSROOM

The exact systems behind everything in this issue. The audit sheets, the routing logic, the templates and the review cadences, built out step by step so you can copy them straight into your own stack. One time, forty seven dollars.

Get the Blueprint for $47

.....

Everyone skips this part. Even Google's announcement doesn't yet say what happens to an agent when its creator leaves, or where a paused agent's email goes. Answer both for your own business before you need to.

Write a three line offboarding plan into the job description:

Do a fire drill once. Turn the agent off on a quiet afternoon using only the plan. If it takes longer than ten minutes or you discover a scenario you forgot about, fix the plan. Then turn the agent back on.

Here's how this played out for the home services company with the Inbox Agent.

Tuesday afternoon we wrote the job description. Took twenty minutes, mostly because the owner and the office manager disagreed about whether the agent should quote price ranges. They settled it: ranges yes, numbers no.

Wednesday we created the identity. A Workspace account named "Inbox Agent (AI)," a dedicated calendar, a shared view of the estimates calendar, read only access to one FAQ doc, and a Go High Level user that could see the leads pipeline and nothing else. Its own Make.com connections. Its own API key with a monthly limit that's about three times what we expected it to use.

Thursday through the following Wednesday it ran in learning mode. Every reply went to a drafts label. Maria, the office manager, read them each morning with her coffee. She corrected eleven of about one hundred eighty, and nine of the eleven were the same problem: the agent was too eager to promise "first thing tomorrow" when the calendar said otherwise. We added one line to the job description. Fixed.

The week after, we opened the gate on replies but kept bookings in draft. A week after that, bookings went live too.

Total cost, a Workspace seat, a few dollars of model usage and maybe four hours of human time spread across three weeks. What they got: nights and weekends answered in under fifteen minutes, a log where every line says either "Maria" or "Inbox Agent," and an off switch everybody knows how to use.

The owner's favorite part surprised me. When a customer complained about a reply, he pulled up the log, saw it was the agent, saw the job description didn't cover that request, and fixed it in five minutes without having to figure out who to blame.

For two years, most of us have treated AI like software. You install it, you configure it, it runs as you.

The coworker idea is a better mental model for anything that acts on its own. An agent that sends email and updates records is doing a job. Jobs need descriptions, managers, limits, reviews and a way out. You wouldn't hand a new hire your personal login on day one. Don't do it to the agent either.

Once you've done this for one agent, the second takes half the time. The third is almost boring, and boring is exactly what you want from the thing running your inbox at 2am.

FROM THE AI NEWSROOM

For operators who would rather build it with someone than build it alone. The full stack, installed alongside you, with the decisions made in the room instead of guessed at. Ninety seven dollars.

Join the Accelerator for $97

.....

Pick the agent or automation that does the most on your behalf without you watching. The one that sends things, moves things or books things.

Write its job description. Four questions, one paragraph.

Create its own email account and move its connections off your login, one system at a time, starting with whichever one sends things to customers.

Name a sponsor and a backup. Put both names in the job description.

Add the activity sheet row to the end of each scenario.

Then run the fire drill. If you can turn it off cleanly in under ten minutes, you've got an employee. If you can't, you've got a liability with good manners.

Jordan

The AI Newsroom | Practical AI for people with a business to run.