My bots built two websites. The code wasn't the hard part.

On the morning of October 19, my web development bot had an update ready for isabelladuarte.com. Before uploading it, it compared its copy of the site with what was actually live on the internet. They didn't match.

The hero photo, the salon section, the globe button, and several photos were already different on the live site. Someone had made tweaks directly on the published site, and the "official source" had fallen behind. If the bot had uploaded its version, it would have silently erased changes no one remembered making.

This is what working with an AI team actually looks like. It’s not an explosion of code. It’s a boring question someone has to ask before every single change: Is what I have the same as what’s published?

What they built

fabiandelafuente.com is my personal site. What the team put together on top of it in just a few days:

• A gallery of my patents, inspired by Google Arts & Culture exhibits: 61 cards, of which 22 are granted patents, 27 are pending, and 12 have been published. Below that, an inventions section featuring the products that came out of them, each with its company logo.
• University: lessons on how I run my AI team, featuring my free book «Build Your AI Team» in PDF in both Spanish and English, plus a 23-minute podcast version.
• The Spanish version of University so Google can show it to anyone searching in Spanish, and short links for each social network that lead straight to the book.

isabelladuarte.com is Isa’s site. She chose the design concept on October 6 from several proposals the team prepared. It features her bio, a map with 47 countries and her travel photos, and a blog. On October 21, her post on the Spritz went live, and that very same day, the entire blog was available in Spanish, English, and French.

The cast: my AI chief of staff set priorities. A web development bot designed, built, and published everything. My communications bot wrote and illustrated. And an assistant bot emailed Isa for the one thing a machine can't invent: her bio, her photos, and her exact titles.

The hard decision

On Isa's site, there were nine pieces of text marked as "provisional": her bio, her contact email, photos of her dog, her exact title at her company. A bot can write a convincing bio in ten seconds. That is precisely the problem.

The rule was that the site would not be presented as finished with made-up data. The gaps were left marked until Isa replied, even if it delayed everything. When her answers came in, the bot changed "CEO and founder" to "CEO and co-founder." It’s one word, and it’s the difference between a site that sounds good and one that is true.

The same went for my patent gallery. Seventeen 2026 applications aren't public yet. I decided they should appear on the list, but without drawings: just a black-and-white card reading "patent pending." What isn't public doesn't get shown, no matter how pretty it looks.

What broke

Quite a bit, as always.

The machine didn't have the key to the door. My bots' main server couldn't connect to upload files to my hosting provider; that route was blocked. For days, every publication depended on someone uploading it by hand. The fix was to move publishing to another of the team's computers that did have a way out, with its own key, and a rule: before every upload, a backup of what was there is saved.

The source wasn't the source. What I mentioned at the beginning. The bot ended up downloading the live site and applying changes to that copy. Ever since, on Isa's site, only the files that changed are uploaded, never a fully regenerated page, because the tool that regenerates it would wipe out hand-made fixes.

The checker makes mistakes too. When the team checked whether everything loaded, a test tool hung forever waiting for a connection that never finished. Another bug: the Spanish folder on my site returned "access forbidden". The day's changes hadn't caused it; it had been sitting there for heaven knows how long. No one had noticed because no one had looked.

The principle

Back in the 1950s, newspaper print shops had a rule: no one touches the printing plate without a galley proof in hand. It didn't matter how much of a rush they were in. You print a sheet first, you read it, and then you run the press.

Bots are frighteningly fast. But speed isn't the risk. The risk is speed applied to the wrong copy. That's why every one of my bots' publications now carries the same routine: a backup beforehand, a check that every page loads afterward, and a timestamped note in the log.

Building a site with AI is already the easy part. The hard part is knowing what is actually published and not letting anyone—human or bot—invent what's missing.

---

The sites: fabiandelafuente.com · isabelladuarte.com

If you want to learn how I run my AI team, download my free book: fabiandelafuente.com/university/