The Self-Driving Small Business

Last week I wrote about Replit’s Self-Driving Company post. Michael, a colleague from a group of founders and builders I’m part of, left a comment. He believes self-driving businesses are the future. His question: do smaller organizations wire this into the processes they already have, or rewrite everything so it’s self-driving from the ground up?

Quick recap. Replit’s CEO published data showing 2.9x more code per engineer after they connected an internal agent to every system the company runs on. They also canceled a seven-figure software contract because an internal replacement, wired into all those systems, beat the commercial product.

Here’s the thing though. Replit’s six months sit on years of infrastructure. An agent harness they built themselves. Engineers who think in orchestration. A data team. If you run a fifteen-person business, none of that is available to you. So no, you can’t run their playbook natively. You also don’t need to. The small business version looks different, and in a few ways it’s easier.

No Process Debt

In The Invisible Interface I tell the story of Maria, a bakery owner I use to explain what happens when this technology doesn’t reach everyone. A single wholesale quote costs her ninety minutes: researching the buyer, checking a spreadsheet she updates by hand, drafting a price, chasing delivery details. Meanwhile the croissants come out of the oven in twelve minutes and she’s the only one who can do that too.

Rereading that chapter after the Replit post, something jumped out at me. Maria has no process debt. Nothing is written down. Her processes are habits in her head. Every consultant would call that a weakness. I think it’s her biggest advantage.

A 5,000-person company has to fight its own documented processes, and the people whose jobs are those processes. Maria has nothing to unwind. No SOP to defend, no committee to convince. The rewrite Michael is asking about is cheap for her because nothing was ever written.

The data agrees. The AIDB benchmarking study found companies of 1 to 50 employees report the best ROI from AI: 3.49 out of 5, against 2.94 for enterprises over 5,000. Simple reason. In a thirty-person company, the person deciding to use AI is the person using it. The feedback loop is seconds, not meetings.

The Math Is Interruptions, Not Hours

The enterprise pitch is a cost story. Hours saved. That math misleads a small owner.

Maria’s quote isn’t expensive because it’s ninety minutes. It’s expensive because it interrupts the only person who can run the ovens. For an owner-operator, the scarce resource is unbroken attention. So don’t hand an agent your most frequent task. Hand it your most interruptive one.

And for a small business this is a revenue story first. The customer who asks at 9 pm whether you deliver to Cambridge and gets silence doesn’t show up on any expense line. She buys from someone else. Agents at the edges capture sales that were quietly dying.

Where I’d Start

Replit went where work was verifiable. Code runs or it doesn’t. A small business needs a second filter: blast radius. Replit can absorb a bad pull request. Maria is one wrong answer about the almond torte away from an allergy incident.

So rank tasks by how checkable the output is and how much damage a mistake does. Internal and checkable goes first. Customer-facing stays in draft mode: the agent prepares, you approve. Self-drafting before self-driving. For most small businesses that’s the right design, maybe permanently.

Three tasks you can hand off end to end today: answering inbound questions from one document you control, drafting quotes for your approval, and chasing follow-ups that die in your inbox.

The One Decision You Can’t Undo

Realistically, most small businesses won’t build any of this. It’s coming inside the platforms you already pay: your storefront, your books, your scheduling tool. Replit says a product version is coming.

That makes vendor choice the one decision that’s hard to reverse. The agent running your business accumulates the memory of how your business works. Leave after two years and the memory stays behind.

In the book I lay out a three-part test I call the principle of dignity: your data, your defaults, your off switch. Can you export what it learned about your business? Do you decide what happens automatically, or does the vendor? Can you turn it off and still operate tomorrow? You can’t negotiate an enterprise contract. You can walk from a vendor that fails those three. Ask before you connect anything.

Meanwhile, do the one thing no vendor can lock in. Write your business down. The price list. The answers to the twenty questions customers actually ask. What a good quote looks like. That’s the substrate every agent will need, it lives in your head, and it costs a few evenings. Whatever platform you pick later, the memory you wrote down yourself is the one piece you’ll always own.

If you’re 50 to 500 people, your path bends back toward Replit’s. You have process debt and people whose jobs are the process. Start in your most verifiable function, run the agent in a public channel where everyone can watch it work, and redesign the workflow around it. Bolt it on and you’ll get a pilot that impresses everyone and changes nothing.

So: neither retrofit nor rewrite. One loop at a time. Most interruptive task first. Drafts before autonomy. The dignity test before you connect. And get the business out of your head while the market builds the rest.

For the owners reading this: what’s the task that interrupts you most, the one that pulls you off the work only you can do?

Harry Glorikian is the author of The Invisible Interface: How AI Turns Intentions Into Actions—And Who Wins ( Ideapress Publishing / Simon & Schuster , June 2026).