Some problems hide in plain sight because everyone assumes someone else owns them. Our process for getting a new product live on the website was one of those. It worked, mostly, but it had grown to the point where a handful of people touched every listing, and none of them fully owned it. So when I decided to build an AI automation agent to take it over, the first surprise was not the technology. It was realizing how tangled the process had quietly become.
I want to walk through why we built it, why I chose to build it myself instead of handing it off, and what it actually changed.
What was broken with our old product page process?
The core problem was accountability. Too many people touched a single product listing, and when everyone is accountable, nobody is.
A product would come in, and then it moved through specs, images, inventory, pricing, and sales channels before it ever went live. Each step had a person, sometimes two, and the handoffs were fuzzy. If something shipped wrong, or worse, never shipped at all, there was no single owner to point to. It was not any one person's fault, which is exactly the problem.
When we later pointed the agent's audits at our own catalog, it surfaced roughly $50,000 of product that was sitting unlisted and hidden, never actually for sale. That is the real cost of a process with no clear owner. Real product, already paid for, sitting in the dark.
How did I figure out the old process before automating it?
I started by asking a lot of questions and asking people to overshare.
Before I wrote a single instruction, I chatted with everyone who touched the process and had them explain their piece in more detail than felt necessary. I wanted the boring stuff, the exceptions, the little "oh, and sometimes we do this" steps that never make it into a written SOP. Because so many people had a hand in it, that oversharing was the only way to see the whole picture.
Then I handed all of it to the agent and had it analyze and assess what I had gathered. It gave me a clean breakdown of the process we actually had, broken parts and all, and then proposed a better one. Seeing our own mess laid out plainly was worth the exercise on its own.
What does our AI automation agent actually do?
Once a product is ready to go on the site, the agent builds the listing, runs its checks, and puts it in draft. It never flips anything live on its own.
We named it Leo, after Leonardo DiCaprio, because every agent we build gets a name and a personality. It makes them easier to talk to and, honestly, more fun to work with. Leo takes a product that is ready for the website, builds the page, checks the balances and details that used to get missed, and leaves it in draft. From there, it hands off to Kent, who reviews the draft and flips the switch to make it live.
That last part matters. The human still owns the go-live decision. The agent does the heavy, repetitive, error-prone work, and a person signs off. There are a lot of intricacies under the hood, but that is the shape of it: build, check, draft, hand off, human approves. Clean lines and one clear owner at each step.
The result is that a process that used to eat hours and hours a week, spread across several people, now runs on a defined path with real accountability.
Why did I build it myself instead of delegating?
I built it myself on purpose, even though I probably could have handed it off, because I wanted to learn how to build with AI, and I did not want to lead from a distance.
We are a small company, around 50 people. My day-to-day is not on the product side of the business. I have not personally built a product page in six or seven years. By any normal logic, this was a delegation candidate, and you could fairly argue it should have been.
I saw it differently. I saw a chance to actually understand how to build these tools instead of just approving them. I did not know all the little intricacies of the process I was replacing, so it took me a while, and I had to ask a lot of questions. But by the end, I had everything I needed, and I understood the build in a way I never would have if I had watched from the sidelines.
There is a version of leadership where you stay in the ivory tower and let everything happen below you. I think that is a cardinal sin. Sometimes you have to jump in, get to the front of the line, and lead by example. Owning this build kept me sharp and kept me fluent in how we actually build with AI.
What did building it actually give us?
Beyond the hours saved and the cleaner accountability, the real payoff was the learning and being able to spread it.
Since I built Leo, I have had multiple calls with team members, walking them through how to build things like this themselves. That only works because I did the reps.
When Leo was finished and the dust settled, I had it do one more thing. Its understanding of how we operate, how the business works, and what we do had gotten strong enough that I asked it to build a copy of that as a shell I could export. The idea was simple: hand that shell to a teammate, let them import it, and give them a running start on building their own agents instead of starting from a blank page.
It worked. We have already built two more agents I was not directly involved with, using the exact learnings I picked up and passed along. I also built Kate, an inventory-audit agent, which watches for listings with no stock, no reorder point, and no open purchase order. The pattern is spreading, which was the whole point.
What is the lesson for other founders and operators?
If you want your team to build with AI, build something yourself first. Not everything, but something real.
You do not have to become the person who builds every agent. You do have to understand the work well enough to teach it, unblock it, and know what good looks like. The fastest way to get there is to pick one genuinely messy process, own the build end to end, and feel the intricacies yourself. This is also where the boring governance advice actually pays off, because building the agent forces you to assign clear ownership for every decision the system makes. Leading by example is not a slogan here. It is the difference between a team that waits for permission and a team that builds.
Frequently Asked Questions
What is an AI automation agent?
It is software, built on a large language model, that carries out a real business process on its own with defined rules and checkpoints. Ours builds product pages, runs checks, and leaves the final go-live decision to a person.
Why did you name it Leo?
We name every agent we build and give it a personality. Leo is short for Leonardo DiCaprio. Names make the agents easier to talk about and more pleasant to work with day to day.
Should a CEO build AI tools themselves or delegate it?
Both can be right. I delegate plenty. For our first serious agent I chose to build it myself so I could learn the craft and teach it. If nobody on your leadership team understands how these are built, that is a good reason to own one yourself.
How does the AI agent avoid publishing mistakes?
It never goes live on its own. It builds the listing, runs its checks, and leaves it in draft for a human to review and approve. The agent does the repetitive work, and the person owns the final call.
What do you mean by "lines of accountability"?
It means every step of a process has one clear owner. Our old process had many hands and no single owner, so mistakes had no home. Building the agent forced us to define exactly who owns what.
Final Thoughts
The technology was the easy part. The hard part, and the valuable part, was being honest about how messy our own process had become, then owning the fix instead of assigning it. We got back hours every week and a catalog with clear accountability. I also got something harder to measure: I actually know how to build these things now, and so do more of my people.
When everyone is accountable, nobody is. The way out is to pick up the work yourself, define who owns what, and lead from the front.
