I coined the term “autoresponder” in 1994. I built the first web-based domain registration service in 1995. I led the worlds’ largest ecommerce webhosting service in 1997. Since then I have pioneered technology solutions and business models across a number of industries, including real estate and title. Today I run a title company and a company staffed largely by AI agents. Here are the 7 rules I follow before I let AI touch anything that matters in a business. 🧵

If your AI setup cannot be explained on one page — who does what, who is allowed to do it, who stops it — you do not have a setup. You have software sprawl with a new vocabulary.

Founders measure AI ROI in hours saved. Wrong scoreboard. Measure decisions improved, exceptions caught, and revenue per employee. Hours saved is how you justify a tool. The other three are how you redesign a company.

I do not want more agents. I want fewer, named, owned, and watched. Headcount vanity is how companies get sloppy. Digital headcount vanity is worse — it is cheap to add and expensive to unwind.

The first serious AI conversation in any company should not be about models. It should be about money: what the agent can spend, what it can promise, what it can break.

Everyone asks what I am automating. More useful: what I refuse to automate. A decision I cannot explain. Anything that spends money without a ceiling. A customer promise the company cannot keep manually. Anything built on a messy source of truth. Automation scales a good process. It also scales a lie.

I run a title company. It is one of the most regulated, process-heavy, zero-tolerance businesses you can pick. That is exactly why I test AI there first. If a workflow survives compliance and customers in title, it will survive in your business.

A founder asked me this week: “Should I hire an ops person or build an agent?” Wrong question. First: write down what the ops person would decide versus what they would do. The doing is agent work. The deciding is still a person — maybe you. Most founders discover the job was 80% doing and 20% deciding, and they only ever needed to hire the 20%.

The most expensive AI implementations I see are the ones that skipped the boring layer. Data cleanup. Permissions. Exception handling. Human review. The model was never the hard part.

Character.AI had roughly 12 million active users. It was also losing about $8–9 million every month. Now it has filed Chapter 11. That is an important AI business lesson. I call it The Growth-Math Gap. More users do not automatically create a stronger business. If the cost of serving each customer rises faster than the value that customer creates, growth can actually make the problem worse. AI makes this especially important because usage has real marginal costs: Compute. Infrastructure. Storage. Engineering. Safety. Compliance. Then add another variable: Legal risk. Character.AI is also facing lawsuits and investigations involving alleged harms to young users. Those claims remain subject to litigation, but the cost and operational burden are real. The lesson for founders isn't “don't grow.” It's: Know what the growth is worth. What does a customer cost? What does that customer produce? Does margin improve as you scale? What new liabilities does the product create? Growth is validation. It is not a business model. Don't let growth hide the math.

Every AI rollout I have watched fail had the same autopsy: No owner. No permission model. No definition of done. And a model told to “just handle it.” That is not an AI problem. It is a management problem with new vocabulary.

Free money (API credits) surprise... @AnthropicAI just started adding a $200 monthly API credit to my account as a benefit on my $200 Max membership. Very cool since I am using Claude models via API on @OpenClaw.

Digital Worker Rule 2: If two runs of the same question produce two different answers, you do not have a system. You have weather. The fix is never a better model. It is a cleaner source of truth and a log of who changed what.

“Autonomous” is a marketing word. What I actually want from an agent is bounded, observable, and interruptible. If it is not all three, it is not working for you. You are working for it.