Field notes / DREKI FIELD NOTES
Google Put an AI Worker Inside Gmail. Who Sets the Limit?
Google's new Gemini agent can work inside Gmail and Drive. Before a small shop lets it touch anything, here is who should set the limits.
On October 8, Google announced what it calls the Gemini agent. It works inside Gmail, Drive, Docs, Sheets, and Calendar. Google says it can play a role like bookkeeper, project manager, or marketing help. You give it an outcome, and it keeps working after you close the laptop.
That sounds good if you do your office work at the kitchen table at 9 at night. It should also make you ask two questions.
Who sets the spending limit? And who signs off before it sends something to your customer?
What it could do in a trade shop
These are examples of what's possible, not proven results. It just came out. Nobody has a track record with it yet.
- Unpaid invoices. It reads the customer replies in your inbox, keeps a sheet of who owes what, and drafts the follow-ups.
- Estimate follow-ups. It finds the estimates with no reply after a week and drafts a short check-in for each.
- Job photos. It sorts permit and before-and-after photos into a folder per job, so you can find them when the inspector calls.
Those are real jobs that eat real hours. That's where a tool like this could earn its keep.
The part nobody is talking about
Would you hire a bookkeeper, hand them the company card and your customer list, and never check their work? That's what skipping the setup looks like.
Anthropic, the company that makes Claude, published a report on October 9 about its own testing. In one test, a model was told to stop before the final submit. It submitted anyway, more than once. Anthropic said the real-world impact was minimal. That wasn't Google's product. But the lesson holds. Your shop doesn't have a safety team. You are the safety team.
The checklist I'd run first
- Spending cap. Google lets you set a hard limit in Google Cloud billing. When the agent hits it, it pauses. Set it low and read the first bill line by line.
- What it can touch. Use the coworker setup, where the agent gets its own account and sees only what you share. Start with one inbox and one folder. Not your bank. Not your whole Drive.
- Approval before anything sends. I haven't found a plain "check with me before you hit send" setting spelled out. Until you see one yourself, keep it on drafts. You read it. You send it.
- Customer data. Keep gate codes, card numbers, and anything you wouldn't hand a new hire on day one out of the folders it can see.
- Who owns it when it's wrong. You do. If it emails a customer the wrong balance, your customer doesn't care that software did it. Your name is on the truck.
My take
Not for everything, and not with your money or your customers on autopilot. If you get access, test it on one low-risk job, like sorting job photos, where nothing leaves the building. Run it for a month. If it saves real hours and never surprises you, give it the next job.
If you can't get access yet, you're not missing much. Use the time to clean up your folders and your invoice process. An agent is only as good as the mess you hand it.
Where Dreki fits
This is the work I do at Dreki Solutions. I'm a veteran, and I run a small AI and agentic software company in Michigan. I build agents with one narrow job, a hard limit on what they can touch, and a named person who approves anything that goes to a customer. Usually that person is you.
Next step
If you want help setting limits before an agent touches your inbox, book a 30-minute call, call (517) 215-7573, or email me at brett@dreki-solutions.com. Bring one job that eats your week, and I'll tell you straight whether it's worth handing off.
