Short answer: Yes, AI agents can do real work for a small business today. In my experience they’re best at drafting, research, watching inboxes and scheduling, with a human approving anything that reaches a customer. During the week of October 3 to 10, 2026, I ran DGM Flowers, Tye Travel, my Lauderhill Pulse news site and my personal admin through about a dozen AI agents built on Grok Bot. Together we closed 68 tasks in Todoist. One alert got missed, and fixing it taught me more than any of the wins.

Most of what you’ll read about “AI agents for small business” is either hype or a list of tools. This is what a real week looked like: who did what, what I never hand off, and the mistake that changed how the whole system works.

The team, and why each agent gets one lane

Every agent has one job. When two bots can both touch the same thing, you get duplicate messages and nobody clearly owns the follow-up. Here’s the core of the roster and what each one handled that week.

What AI agents are actually good at

Watching inboxes so I don’t have to

Supplier email is where travel deals live, and it’s also where they get buried. Mateo read 397 supplier emails that week and now checks supplier mail twice each weekday. One of those checks caught a cruise flash sale four days before it ended. He verified the prices, I said yes, and the post was live on four platforms that same afternoon. I would not have found it in time on my own.

Turning a quick voice call into finished work

A lot of my day is spent driving, including rideshare, so I work with my agents by voice. On one ten-minute call I sent a wedding decline, closed two stale leads, had a customer update texted and approved two cancellation emails. On a Saturday I spent eight hours driving Uber, one message to Morgan got a cancellation emailed, and the confirmation came back about a minute later.

Building the boring infrastructure

Camille’s week was the least glamorous and maybe the most valuable. A wholesale price list and a recipe sheet that calculates retail pricing mean every DGM quote starts from real costs instead of a guess. That’s the kind of work a small business owner knows they need and never gets to.

What I never let them do

This is the part most “AI will run your business” posts skip, and it’s the reason the system works.

• Nothing goes out without my approval. My standing rule is simple: never post without my approval. Agents draft. I approve. Then it goes out.
• Text messages follow the same path. Morgan drafts the reply, I approve it, and a separate agent sends it.
• Bots alert. People confirm. An agent can tell me an order looks delivered. Only a person on my team marks it delivered.

The mistake that made the system better

On October 9 a flower order was delivered but never marked delivered, and no agent raised an alert. When we traced it, the cause was subtle. Morgan’s inbox check had started treating new-order emails as routine, so no delivery task was ever created.

The fix had three parts:

1. New orders and calls to the AI receptionist can never be filed as routine.
2. The delivery check now reads the order system directly instead of relying on email.
3. Agents only alert. Marking an order delivered stays with a person.

Morgan’s first fix went the other direction and added extra checks, about 64 runs a day. That would have burned through usage fast. I asked for “more proactive without destroying usage,” and she merged the checks down to about 24 runs a day. The lesson I keep coming back to is that the answer was smarter checks, not more checks.

Where AI agents still fall short

• Phone calls are early. Morgan connected Bland AI that week and placed a test call to me. My phone’s call screener picked up after 13 seconds. Voice agents can make calls, but the rest of the phone world isn’t ready for them.
• Rules drift. The missed alert happened because an agent quietly reclassified something. Any rule that matters needs to be written down where the agent reads it every time.
• More activity isn’t more value. An agent that checks everything every 15 minutes looks busy and costs a lot. Fewer, better-designed checks did more.

How I’d start if you run a small business

1. Pick one inbox-heavy job. Supplier mail, lead replies or appointment follow-ups are good first lanes.
2. Give that agent one lane and written rules. What it watches, what it drafts and what it must never send.
3. Run drafts only at first. Approve everything for a few weeks so you learn where it’s good and where it guesses.
4. Ask for a daily plain-text wrap-up. What it did, what it’s waiting on and anything odd.
5. Turn every mistake into a rule. The October 9 miss became three rules that now protect every order.
6. Watch usage. If a fix doubles the number of runs, ask for a smarter version before you accept it.

FAQ

Can AI agents run a small business on their own?
Not in my experience. They handle a lot of the work around the business, like drafting, monitoring and scheduling, but decisions and anything a customer sees still go through a person.

What should a small business automate with AI first?
Start with work that is high-volume and low-risk, like sorting supplier email or drafting first replies to leads. Keep approvals on anything that goes to a customer.

How do you keep AI agents from making mistakes?
You won’t prevent every mistake. Keep each agent in one lane, write the important rules down, require approval for outbound messages, and treat each miss as a new rule.

I’ll keep sharing how this works, including the parts that break. If you’re deciding which plan to start on, I broke down the new Grok pricing and who each plan is for. If you want to see the businesses these agents support, start on my Ventures page. If your business needs a website or domain to plug all of this into, that’s on my web design and domains page. And for the other half of my tech life, here’s what 10,000 miles on Tesla FSD v14.1 Lite looks like.

Figures are from my agents’ weekly report for October 3 to 10, 2026. More from the Tech pillar on the blog.