The CMS that followed me through a career
Content management systems have basically followed me through my whole career. Systems built from given modules: for a new page or a new piece of content you do not always need a developer. Someone learns the building blocks, which content fits which module, which elements make a landing page or a product page.
That was always a comfortable setup, because developing the pieces (new modules, maintenance, the design system) stays separate from everyday “website building.” Companies of every size use systems like this. And that is exactly why it felt so obvious that AI would kick the door in here first.
Two paths, and which one they are selling
I think there are two paths ahead of us.
Path one: the AI knows the company’s design system, voice, products, audiences. The input is a simple prompt (“we need a landing page for the new product”), and the output is a page assembled from the CMS modules, maybe in several versions, for several audiences, in several languages. The process stays in-house. The prompter might be a developer or a marketer, still signing off on the result (a person in the loop), responsible for the generated page that users actually see. Not HTML poetry on every session: structured building from blocks you already have.
Path two: skip the CMS and generate everything. Here the prompter is the user. Say they are shopping for a new laptop, and the output is a unique page generated for them, with the products that match what they want. Hyper-personalization. The experience is fully custom; everything is born in that session.
Long term, I think path two is the destination. We are not there yet, for a few reasons.
One is reliability. Personalization sounds great until the hallucination rate is above zero. What if the AI recommends a product that does not exist, or gets a little creative with the specs? Traditional pages can be wrong too, but finding and fixing the bug is much simpler than reaching into a model’s brain. Classical web is deterministic: you see what is in the database. An LLM is probabilistic: it writes what is most likely to come next. Marrying those two is not a marketing slide. It is risk.
The other reason a lot of pitches leave out: cost. If we generate everything in every session, the bill climbs fast. It is much cheaper to build a page once than to generate a unique one for every user (text and the full HTML skeleton every time).
And then there is speed. People expect a page to load in under a second. If the interface keeps changing, muscle memory (“the cart is top right”) melts too. Hyper-personalization is not only expensive. It has a cognitive tax.
The YOLO mentality
I see path one as the honest near term, and still, tech often thinks in extremes and leapfrogs.
People are surely working on structured generation too. That is not what looks like the trend. The other path is. I know why: it is less exciting on a slide to say “build on the architecture we have, fix the foundations, introduce AI gradually” than to say “from tomorrow you need neither developers nor designers, every user gets a unique experience, and ROI goes to the moon.”
It feels like the YOLO mentality is running the room. Reasonable arguments get swept aside, because one thing overrides everything: AI has to be introduced as fast and as hard as possible.
That is not a technology constraint. It is business and psychology. Venture capital is hunting unicorns, not a CMS workflow that is thirty percent faster. On an earnings call it is hard to say we are carefully making the existing system smarter: you need a flashy, sci-fi feature on the slide. Marketing’s narrative is currently louder than the senior engineer whose day ends with sub-second load times, scale, and error messages. And there is the Kodak fear: if we do not tear down our own working model, someone else will do it for us.
I understand the fear. YOLO is still not a strategy.
What the storefront hides
I have another data point for this, and this one is not theory.
For the past four months I have been analyzing Fortune 500 enterprise websites to understand how they are built, and how ready they are for this new AI era. I have been publishing what I found as I went (the latest one just last week). Without making you dig through all of them: what they share is that the pages are full of errors, and they are unoptimized. Structured markup, machine-readable hints like llms.txt, cleaner page structure, semantic HTML, load time: a thousand small things you could already fix. Meanwhile a lot of those same companies already ship well-sounding AI solutions.
Those foundations could be cleaned up much faster today with AI assistance, and the result would be tangible. Once they are stable, you can move toward the bolder, more complex stuff. Sometimes I feel like an old man mourning the good old days, but I think what I am saying is just common sense.
Showcase AI is fast. You can drop a chatbot on the site and leadership checks the “AI strategy” box. Page structure, semantics, machine-readability: that is invisible work. Hard to sell in a board presentation. And yet if people increasingly search through models, a website’s primary “readers” are often bots, and the bot does not read the hype. It reads the structure.
Long term is not tomorrow
So: are prompts replacing web UI?
I do not think so anytime soon. The webpage as a format stays, at least while reliability, cost, speed, and a learnable interface still matter. And there is a quieter reason we barely touch in these debates: bots and AI still need to get the information from somewhere. I said at the start that websites are the easiest digital source for training any model. Why would you unplug that?
How we treat the website is transforming too. It is not only our storefront for customers. It is becoming the company’s public knowledge graph: the structured, public record that models read, cite, and summarize when a person never opens a tab.
What changes is how the page gets made, and who reads it first. AI could already speed up page building from the blocks we have. And it can already make a company invisible if the machine layer (clean structure, markup, data) is a mess while the storefront is already flashing generative wonders.
Hyper-personalization can be the long game. The YOLO mentality only succeeds at mixing up the long game with tomorrow.
So here is the question I want to throw back: which of the foundations did you check off before the AI strategy made it into the quarterly report?