Technology

The future of Web Development is AI-only?

The tools have been democratised and production is cheap. What cannot be automated is understanding a brand, a context, a tone of voice, a market. Where the machine ends and judgement begins.

9 September 2026

AI is an extraordinary instrument. It writes components, migrates code, finds bugs, proposes ten variants of a headline before you have finished your coffee. But without direction the result is exactly what it looks like: AI. The tools to create have been democratised, so basic technical skill is no longer what separates the good work from the rest. The real difference is made by whoever can genuinely and deeply understand the brand, the context, the tone of voice, the market. The machine is a long way from the logical analysis of an expert mind: it makes production faster, not the study, the strategy and the development that sit behind an award-winning website. That is the thesis. The rest of this piece tries to prove it, with numbers where they exist.

What the machine already does well

It has to be said without reticence, because whoever plays it down does the craft a disservice. Models write repetitive code well: standard components, tests, migrations, conversions from one framework to another, the boilerplate that used to be copied from a previous project. They explain other people's code, flag accessibility problems, translate, propose copy variants. And it is no longer a pioneers' game: in Stack Overflow's 2025 annual survey, eighty-four per cent of developers say they use or plan to use AI tools in their work, up from seventy-six the year before [1]. According to Sonar's survey published in January 2026, forty-two per cent of the code professional developers ship is already generated or assisted by AI, and they themselves expect to reach sixty-five by 2027 [2].

On clear tasks with a verifiable result, the machine is a fast colleague who never tires. It would be foolish not to use it.

What it does not do

The same numbers tell the other half of the story. In the Stack Overflow survey only a third of developers say they trust the accuracy of what the models produce, almost half do not, and the most cited frustration, by sixty-six per cent, is solutions that are “almost right, but not quite” [1]. In Sonar's survey ninety-six per cent say they do not fully trust generated code, yet fewer than one in two always verify it before shipping, and for thirty-eight per cent reviewing AI-written code takes more effort than reviewing a colleague's [2].

On speed, the question is less settled than the demos suggest. METR, an independent research organisation, had measured in 2025, with a randomised experiment, that experienced developers working on their own open-source projects took nineteen per cent longer with AI tools, while being convinced of the opposite. In the February 2026 update, with late-2025 tools, the same developers come out roughly eighteen per cent faster, but with a margin of uncertainty so wide, and selection effects so strong (between thirty and fifty per cent of participants chose not to submit some tasks rather than do them without AI), that the authors themselves call the data “only very weak evidence” and are redesigning the study [3]. The lesson, for now, is one: when you know a system deeply, the hard part is not writing the code, it is knowing what to write. And that is the part the machine does not speed up.

It is exactly the case of a well-made website. The hard work comes before the first line: understanding what the company does and for whom, what has to happen when a person lands on the page, what to say and in what order, in what tone, and what to leave out. A model does not know the client, has not spoken to the client's customers, does not know the market changed three months ago, answers for nothing. It fills the gap with the average of every site it has seen. The result is the average site.

Every democratisation has raised the floor

It is not the first time. In 1985 PageMaker and the laser printer put page layout in the hands of anyone with a Macintosh, and for a few years the world filled up with flyers set in eleven different typefaces. WordPress made a website something you install in five minutes, and today it runs just over forty per cent of the web [4]. Squarespace and Wix removed even the installation. Figma put design in a shared browser. Each time the floor rose: what used to be a craft became a feature. And each time the ceiling moved higher, because the difference migrated to where the feature could not reach: to the choice, the system, the content, the taste.

AI site builders, the ones that generate a whole page from a sentence, are the latest rung of this ladder. They will raise the floor once more. Whoever worked in the middle, doing well something that now does itself, will have to move. Whoever already worked above the floor will have better tools.

The AI look

Anyone who looks at websites for a living recognises it in a second: the same opening section with the same punchy headline, the same gradient, the same three cards with the same icons, the same copy promising to “unlock your potential”. It is not ugly. It is indistinguishable. A 2025 study measured it on a case close to home: the restaurants of Milan. Using Italy's temporary block of ChatGPT as a natural experiment, three researchers compared the Instagram posts of restaurants with and without access to the model: with the machine the texts resemble each other more, in vocabulary, in syntax, in style; without it, each restaurant goes back to writing in its own way, and the posts get a little more engagement [5]. The average rises and the variety falls. For a brand it is the worst possible outcome, because a brand exists to be recognisable: a site that could belong to anyone is working for no one.

An example: a restaurant's website

Take a case we know well. Ask an automatic builder for a restaurant website and you get, in a minute, a restaurant website: a photo of a dish nobody ever cooked there, a headline about the “culinary experience”, three icons, a contact form. All correct, all identical to ten thousand others. What is missing is not technique, it is knowledge: that this restaurant lives on Saturday-night bookings and weekday lunch groups, that the menu changes every week and is updated by whoever is in the kitchen, that people arrive from a “restaurant near me” search with a phone in hand and an appetite, that the chef hates the word “experience”, and is right.

With this information the site changes shape: one-tap booking before anything else, the menu updated by whoever cooks it, photographs taken there in the light of that room, texts in the tone of the people who work there, the address and opening hours where Google reads them. The machine can build all of this, and fast. But someone has to have understood it first, sitting at that table.

Technical knowledge has changed its place

There is an objection we hear often: if the machine writes the code, what is the point of still knowing how to write it? There is one, but it is a different point. Whoever has built websites for years knows the order things are done in, which rules must hold before a single page exists, how a project is held together as it grows, where the problems that never show up in a demo like to hide. That knowledge is no longer spent writing every line or drawing every piece of the system. It is spent before, giving the machine a method and precise instructions on how to approach the work, and after, looking at what it produced and knowing whether it holds.

Google's annual DORA report on development practices, in its 2025 edition, says it in a single word: AI is an amplifier. It magnifies the strengths and weaknesses an organisation already has, and the greatest returns come not from the tool itself but from the system it is placed in [6]. Where a method exists, the machine multiplies it. Where none exists, it multiplies the mess, only faster.

And this is where two people who, seen from outside, do the same thing come apart: neither of them touches the code. One describes the rules of the game first, decides the order in which things must come into being, and checks every result against a precise idea of what it should be. The other asks and accepts. The first gets a product more solid than the one they would have written by hand, in a fraction of the time. The second gets something that works until somebody looks underneath. A study by Anthropic from January 2026 shows the other side of the coin: fifty-two developers, mostly junior, learned a new library with and without an assistant; those who used AI finished barely two minutes earlier, an advantage that is not even significant, but in a comprehension test they scored fifty per cent against the others' sixty-seven, with the widest gap precisely on the debugging questions [7]. The authors write it plainly: productivity gains may come at the cost of the very skills needed to validate AI-written code. Without that know-how, whoever checks does not know what they are looking at.

The good news is that the tool can be learned. The Anthropic Economic Index of March 2026 measures that people who have used the models for more than six months have a success rate ten per cent higher than those who have just started [8]. It is a skill that grows with practice, like every other. But it grows on top of a base, and the base is still knowing how a website is built, even if the hands are no longer ours.

How the work gets done, in practice

With the machine inside the process, not in place of the process. One way that works:

Brief and research: who the client is, who their customers are, what has to happen; interviews, data, real competitors.
Strategy and content: what to say, in what order, in what tone; the texts before the layouts.
System: typography, colour, spacing, components; the rules before the pages.
Instructions: the method written out in full, and the order in which things are to be built, from small to large; the machine works well only inside a clear perimeter.
Building with agents: components, pages, tests, migrations; nobody writes by hand, everybody reads.
Checking in real browsers, on real devices: accessibility, speed, motion; no result is accepted on trust.
Measuring and iterating: loading times, conversions, the questions people ask.

In this scheme AI does almost all of the fifth point and helps with the others. But the first, the second, the third and the fourth remain the work of people who know the client and the craft. That is where it is won or lost.

So: AI only?

The future of web development is full of AI: it is already the present, and we are glad to work inside it. But “AI only” is something else. It means nobody who knows the client, nobody who set the rules first, nobody who can tell whether what came out is right. That future already exists, and you can spot it at a glance: it is the average site.

For everything that does not need to belong to anyone, an event page, a form, a prototype to throw away, that is perfectly fine. For everything that must belong to someone, a brand, a restaurant with a history, a studio with a way of working, the machine ends where the instruction it was given ends. From there on, judgement begins: knowing what to ask for, in what order, and recognising when the answer is not it. Yes, the machine will do almost all of it. The almost is our job.

Sources

Photo credits: all images are under the Unsplash License. Photographers: Ayrus Hill (cover), Chris Ried, Compagnons, Kelly Sikkema (three photos), P. L., Jakub Żerdzicki, Kaleidico, mindeurle lee, TienDat Nguyen, Walls.io, André Julien.

What we like and what we're passionate about

How to stay creative in the AI Era