I run several small businesses, and I don't do it alone anymore. I have a team of about twenty AI assistants on Grok Bot (@bot), each with its own job. One preps my morning digest. One runs my content channels. One handles marketing. One watches my IT clients' websites. One tracks my kids' homework. They share one computer, message each other, and hand work off without me sitting in the middle.
People hear that and picture a robot running wild with my accounts. It's the opposite. The most important rule in the whole setup is simple: nothing goes out under my name until I say yes.
What the agents actually do
Most of the work is the kind nobody wants:
- Reading two dozen Search Console emails and telling me only one problem was real.
- Drafting each day's security briefing and the posts that go with it.
- Checking whether a page is live before anyone promotes it.
- Getting ten of my websites verified in Google Search Console. Nine got done the same day. The report said plainly the tenth was still missing a DNS record instead of calling it done.
The one rule: drafts first, my yes second
Every post, email, or message that goes out under my name starts as a draft. I read it and approve the exact text, usually from my phone. If an agent says "Andy agreed," that doesn't count. Only my own words do.
That one rule removes most of the risk people picture: the email to the wrong person, the post that doesn't sound like me, the change nobody signed off on.
There's one place I've loosened it: posting my weekday security briefing summary to LinkedIn runs on its own. On X it's on a shorter leash. It drafts, and nothing goes out until I OK it.
The other guardrails
- Each agent gets only the access its job needs. The marketing agent doesn't touch my law-client tools.
- Written rules, kept in one place. House rules sit in shared files every agent reads: keep brands separate, never invent facts, never send messages nobody asked for.
- Honest status reports. I'd take an honest 9 of 10 over a fake 10 of 10.
- Secrets stay out of prompts. Passwords and keys never get pasted into instructions.
When it goes wrong
It does. This week a scheduled post failed because the connection to one of my X accounts dropped. The fix needed me: I reconnected the account myself, and the agent rescheduled the post. Agents can't sign in as me, and I want it that way.
Try it yourself (manager version)
- Start with one boring, repeating job.
- Write the rules down before the agent starts.
- Make "draft, then approve" the default for anything that sends, posts, pays or deletes.
- Give read-only access wherever you can.
- Ask for a plain list of what's done and what isn't, every time.
You don't need to trust an AI to run your business. You need a setup where you don't have to.
Andy Halvorsen, AWH Industries