Can AI replace a typical CMS?
For a small site, an AI assistant can do the work a content management system usually does.
Instead of a content management system, I am letting an AI assistant create, edit, and publish the pages directly — the content lives in plain files, and I ask for changes in plain words. The question is whether that holds up against the everyday CMS, or only looks tidy until the first awkward edit.
Right now
Under inquiry. Currently running — the full report follows as it concludes.
The findings
The whole report, open: what it cost, what broke, step by step — and what I’d do differently.
Overview
This experiment asks a plain question: for a small site, can an AI assistant do the job a content management system usually does? Instead of a CMS, the content lives in plain files and I ask the assistant to create, edit, and publish the pages directly, in plain words.
This very page is the proof. It was written, styled, and published by asking — no CMS admin, no dashboard, no clicking through menus. So the report below is not a forecast; it is a description of how the site you are reading is actually run.
The setup
The whole stack lives on Google Cloud
Every technical part of the site — hosting, the domain, the database, the deploy — runs on Google Cloud. I taught the assistant to operate the platform itself: a written skill that lets it use the Cloud modules, tools and features it needs to keep the site actually running, from the server it is hosted on to the integrations around it.
Design comes first, in Figma
The thing that changed everything: I do not say “build me a website.” I design the UI in Figma first, then hand that over as the reference. With a picture to match, the colours, the type and the layout stay aligned with the brand instead of drifting. Design first, build from it — in that order, every time.
Existing code is welcome
It also works well on a site that already exists. For one project I exported the code from Webflow and told the assistant to take that codebase and manage it from there. It does not need a blank page to start; it can adopt what is already built.
What worked
Structure is the lever
The better the site is structured up front — clear sections, named, with a layout that holds together — the better the assistant understands what I want. Once that scaffolding exists, a request as small as “add a blog article with this content” is enough: it knows where it goes, builds it, and deploys it.
A staging environment, not the live site
Changes land on a staging copy first, get looked at, and only then get promoted to the real site. The same discipline a careful team would keep — just driven by plain requests instead of a release manager.
Where it breaks
Undefined = drift
The honest downside: when something is not clearly pre-defined, the assistant will invent it — and an invented section tends to land just outside the brand. It fills the gap, but not always in a way that fits. The failure is real and it is mine to prevent, not the tool’s to guess.
The rule that fixes it
So the process is not optional: design the section in Figma first, make it look exactly as it should, and only then build from it. Skip that step and you get something plausible but off-brand. Follow it and the drift disappears. The lesson of the whole experiment is that one sentence.
Why it matters
If you have basic technical skills, you no longer need the middle layer
You do not need a website builder or a hosting-control panel to run your own site. With a little technical footing you can manage it in natural language — in German, too. “Add this, change that, publish” is the whole interface.
If you delegate, it changes what you are paying for
And if you would rather not do it yourself and hand it to an agency: an agency still working by hand will cost you far more, because they are doing manually what a request can do in minutes. The better deal is one where you send what you want — even a WhatsApp voice message — and the agency’s only real job is to make sure the AI is used the right way. That is where this is going.