The first run

In September 2026, we built our first AI content worker and ran it for the first time. The idea was simple. Feed it a written brand profile, let it draft a week of content, and have a human review everything before it goes out.

The worker runs as a scheduled workflow inside an automation platform called n8n. It reads from a brand profile, writes blog posts and social posts in one pass, and hands them off for review. No part of it publishes anything on its own. That rule has not changed since day one.

The stray character

On one of the early runs, the output broke. Not because the writing was bad. The formatting cracked on a single stray character buried in the response. One character, and the whole structured output failed to parse.

We could have told the AI to be more careful. That fix would not have held. Instead we switched to structured output that the API validates before it ever reaches us. Now the system checks its own formatting against a strict shape before anything moves forward. If it does not match, it fails loudly instead of breaking quietly downstream.

That is the kind of fix you want from a vendor who builds systems for a living: not a patch that asks the AI to try harder, but a change that makes the failure impossible to miss.

What the first drafts actually sounded like

The bigger problem was not a technical bug. It was the writing itself.

The first drafts used phrases that sounded like every other AI-generated post you have scrolled past. They announced themselves instead of starting in a moment. They hedged. They wrote the same way for a law firm managing partner as they did for a property manager, as if both readers wanted the same thing from the same sentence.

We read those drafts and wrote down exactly what was wrong with them. A list of banned words and phrases. A rule that each week tells one story, not five. A rule that every channel gets written for its own reader, because the person scrolling Facebook on a Tuesday afternoon is not the same person reading LinkedIn on a Wednesday morning.

That list became the voice rules the worker runs on now. They did not come from a style guide someone wrote in a conference room. They came from reading bad output and naming the problem.

Why a human still reads everything

Every one of these posts, every blog, every social caption, goes through a human before it publishes. That is not a temporary step while the system is young. A worker that drafts content is not the same as a worker that should be trusted to publish content without someone checking it first.

If you run a law firm, you already live by a version of this rule. Texas lawyers have a professional duty to supervise the nonlawyers and systems that do work on their behalf. An AI drafting tool is no different. It can save you time on the writing. It cannot sign off on what goes out under your name, and it cannot replace your judgment about what a client needs to hear.

The same logic applies if you run a property management company or a local service business. The tool can draft the callback text, the review response, the weekly post. You still read it before it goes to a customer.

What this cost

Nothing about this was free. It cost the time to notice the broken output, trace it to one stray character, and rebuild the formatting step so it validates before anything moves forward. It cost a full read-through of bad drafts to build a list of what not to do. It cost the discipline to keep a human reviewing every single post instead of turning the system loose once it looked good enough.

That is the actual price of building a system like this. Not a subscription fee. The hours spent reading your own machine’s mistakes closely enough to fix the real problem instead of the symptom.

What is next

A chief of staff layer is planned next, one that reports back in a daily phone briefing instead of a dashboard you have to remember to check. We will write about that one once it is built and tested, the same way we wrote this one: after it broke, not before.

If you are weighing whether AI belongs in how your firm or your business runs, start with the part nobody wants to admit. Something will break first. The question is whether you build a system that catches it, or one that just hopes it does not happen again.