I sell AI systems to other businesses. It would be a bit rich if I did not run them on my own. So: an AI agent writes and publishes most of what goes out under my name, including a good deal of this journal. People find that either obvious or unsettling, depending on how they feel about the technology. Here is the honest version of how it actually works.
What the agent actually does
Every morning it reads the content calendar I approved at the start of the month, writes the two posts scheduled for that day, builds the share cards, runs a set of checks, publishes, and writes a log of what it did and what it skipped. I did not sit down at a keyboard this morning. It is not a tool I prompt. It is a process I set up once and it runs whether I am in a client meeting or not.
The important word there is "approved". Once a month I look at the plan: what we are writing about, which questions we are trying to answer, what is off limits. That is the decision. Everything after it is execution, and execution is exactly the part a machine should be doing.
The rules matter more than the writing
Here is what took me longest to understand. The valuable part of this setup is not that the agent can write. Lots of things can write. The valuable part is the list of rules it cannot break.
Some of mine: never publish a number without a source. Never quote our own prices in a public post. Never touch certain projects. Never invent a client result. Those are not suggestions in a prompt, they are scripted checks that run before anything goes out, and they fail loudly. The agent does not get to decide whether today is an exception, because it is not asked.
That is the same thing I tell every business owner who wants to point AI at their work: the model is not the hard part any more. The limits are. Write down what must never happen, then make it impossible rather than unlikely.
Two things it caught this week that I would have shipped
This is the part I did not expect, and it is the honest reason I keep going with it.
We published a piece about which documents South African law says you cannot sign electronically. It is a legal topic, so getting it right actually matters. The agent went and pulled the text of the Act itself rather than trusting the write-ups. Good thing. One of the sources it read stated a threshold of ten years for long leases. The Act says twenty. That number is repeated on a lot of South African pages, and I would have quoted it without blinking, because it is everywhere and it looks right. Reading the primary source is not clever. It is just discipline, applied every time instead of when you happen to feel careful.
The second one is more embarrassing. Our site is a single-page app, which means asking a web address whether it exists gets you a cheerful "yes" for absolutely any address you invent. Earlier posts had linked to product pages that do not exist under those names. The links looked fine, returned a healthy status, and quietly went nowhere. Nobody noticed for a week. The agent noticed because it compared a real page against a made-up one and saw they were identical, then went and fixed the posts that were already live.
Neither of those is genius. Both are just checks that got run because a process ran them, not because someone remembered to. That is the whole pitch for this stuff, and it is much less exciting than the demos.
The part I do not hand over
Judgement, and anything with my name on the emotional end of it.
The agent does not decide what we stand for, what we will not say, or which promise we make to a client. It does not pick the month's themes. It does not talk to anyone who is upset. When a post touches something that actually happened to me, I write it, because there is no source for that and no model that can fake it honestly.
And I read what goes out. Not every draft before it publishes, or I would just be doing the job slowly with extra steps. But I read the log, and I read the posts. If something is wrong, that is mine. "The agent wrote it" is not a defence I get to use, and I would not respect anyone who did.
Would I recommend it
To a business owner, yes, with one condition: not yet. Not on day one. Do the boring work first. Find the thing you do the same way every single time, write down the rules for it, and only then hand it over. If your process is a mess in a person's hands, automating it just gets you a faster mess with better grammar.
I did not start with an agent that runs my marketing. I started with a morning where I sat down and worked out what was actually eating my week, which is where this whole thing started. The agent came a long way after that, and it only works because the thinking happened first.
Frequently asked questions
Does an AI write your blog posts? Yes, most of them, including a fair portion of this journal, from a calendar I approve monthly. I set the themes, the rules and the limits. It does the writing, the checks and the publishing. I read what goes out and I own it.
How do you stop it publishing something wrong? Scripted checks that run before anything goes live and fail loudly: no unsourced numbers, no invented results, no off-limits topics. Rules the agent cannot decide to skip, because it is never asked whether today is an exception.
Doesn't that make the writing generic? It would if I handed over judgement too. The agent executes a plan I set and cites real sources. Anything that needs a point of view, or that actually happened to me, I write myself. The rules are what keep it specific rather than the tone.
Should I automate my marketing this way? Eventually, not immediately. Automate a process you already do the same way every time and can write the rules for. If the process is a mess done by a person, it will be a faster mess done by a machine. Do the thinking first.