Case study · Our own systems · Kenya
The AI system that runs our own publishing.
Before we advise anyone on AI, we run it in our own business. The blog, the news desk and the social media for this site run on a pipeline we designed and built. This page documents how it works and what it has and has not achieved.
The numbers
What the system has published.
Counted from this site's content on 29 September 2026. The social figure covers our own pages and one client's Facebook page.
What this system does not do: it does not replace judgement, and it does not guarantee an audience. Publishing every day builds the library, and readers follow more slowly. We will add traffic results to this page once they are large enough to mean something.
How it works
Four steps, with a check before every publish.
Choose the topic
Topics come from a researched keyword list mapped to what Kenyan organisations actually search for. A rule in the pipeline keeps at least two of every three articles AI-led.
Draft with a language model
A language model drafts the article against a fixed brief: Kenyan context, worked examples with the arithmetic shown, and no named clients.
Check before anything is saved
Automated checks reject a draft that claims a client for us, names a regulator that does not exist, or comes back too short. Sentences that attribute a statistic to an outside body are stripped.
Publish and share
The site rebuilds on its own. A scheduler then turns published guides into branded posts for Facebook and Instagram at 07:00 and 17:00 East Africa Time.
People still set the rules. We choose the topic list, write the checks and spot-check what is published. The system does the repetitive work.
What we learned
Three lessons we now bring to client work.
A language model will invent a case study if you ask for one
So the pipeline never asks. Examples are written as unnamed worked examples, and any draft that names a client of ours is rejected before it reaches the site.
Schedulers are less reliable than they look
The built-in scheduler ran hours late, which is useless for a 07:00 post. Each slot now fires from a separate clock and is recorded, so nothing posts late or twice.
Keep the model away from the last step
Social captions and images are assembled from the published article itself. No language model runs at posting time, so nothing new can be invented on the way out.
The same thinking applies to any repetitive work in your organisation. Start with a free AI Business Assessment, or read how our AI consulting and AI automation engagements run.
Talk to us
Want a system like this in your business?
Tell us the repetitive work that eats your week. We will tell you honestly whether AI is the right tool for it.