I run a one-person consulting business out of Akron, Ohio. AI drafts my emails, chases my invoices, updates my customer records, preps my meetings, and writes the first draft of almost everything I publish. The company has real customers and no employees.
Over the past few months I have also built similar setups for 11 other small businesses. An insurance firm, a scrap metal recycler, a Medicare agency, a marketing shop, solo consultants. This piece is what I would tell you over coffee: what works, what breaks, and the one rule that holds it all together.
The timing matters. The U.S. Chamber of Commerce found that 40% of small businesses now use generative AI, nearly double the year before, and Stanford’s 2025 AI Index puts organizational AI use at 78%. Yet most owners I meet use AI like a smarter search box. The value sits in the gap between asking a chatbot questions and letting AI run your operations.
What running on AI looks like day to day
Yesterday I dragged a customer card into the “won” column on my pipeline board. About ten seconds later my phone buzzed with a confirmation that the follow-up work had already kicked off. I never opened another app.
Another moment from this week: from my phone, over text, I asked my system to check my contact list against the live database. It matched all 35 records, flagged nothing out of place, and I moved on with my day.
This is plumbing. Boring, reliable plumbing that happens to read and write English, wired into the tools the business already uses: the inbox, the calendar, the payment system, the customer list.
The mental shift that took me longest: treat AI like an operations employee you train, supervise, and build guardrails around, rather than a genius you interview.
The one rule that makes it safe
My business runs on a single rule. AI drafts, I approve.
My system can write any email it wants. It is physically blocked from sending one. I built a hard gate into the sending tool itself, so even when the AI decides an email is ready, the send fails until I have read it and said yes. The same gate covers invoices and anything else a customer would ever see.
Good research backs the rule up. A large field experiment with BCG consultants found AI made people meaningfully better at most tasks, and worse at tasks past the AI’s limits, because they trusted confident answers that were wrong. The researchers called it falling asleep at the wheel.
I aim for a business where the machine does the typing and I do the judging.
Real numbers, honestly framed
The most dramatic result I have seen came from an insurance firm. One employee built quote comparisons by hand, and by her own time breakdown, each one took about three and a half hours. The system I set up does the same comparison in about five minutes, and she reviews the output before it goes anywhere.
Those are her numbers, from her own accounting of her day. I am careful about that, because this industry has a fabrication problem. If an AI consultant quotes you a stat nobody measured, run.
The durable value story is quieter. The wins come from forty small automations, like the follow-up that goes out the same day instead of next week, or the invoice that gets chased without anyone dreading the task.
Where it broke
Most articles about running a business on AI skip this part.
In my first client setup session, a workflow I had planned died on the spot because a website threw up a login wall the automation could not get through. The client watched it happen. Live builds teach humility fast.
Worse: I asked AI to attack the billing code it had written for my own membership site. The audit found a real bug that could have double-processed a payment. AI wrote the bug and AI caught the bug, but only because I made adversarial review a standing step instead of trusting the first output.
My favorite failure is the embarrassing one. For a stretch, my own site had a payment path that looped customers in a circle, so a person who wanted to pay me could not find the way to do it. I had reviewed the site the way a builder does instead of the way a stranger does, and the machine had no way to know the difference.
The pattern across all three: AI concentrates judgment instead of removing it. I now spend the hours I once spent typing on checking outputs and asking better questions.
What I would tell another owner
Start with one workflow you hate. Pick a recurring task that makes you groan, automate it end to end, and keep yourself as the approval step.
Then let it compound. One workflow becomes five, five become a system, and six months later you look up and the business runs differently. Consistency over intensity, the same rule as everything else I have built.
Keep your hands on the wheel while it compounds. The owners I have watched get real value from AI stayed awake while automating and treated every confident answer as a draft.
I am one person in Ohio who decided the tools were finally good enough to run a real company on, then did the unglamorous work of wiring them in. The tools are the cheap part now. The judgment is the job.
