There is a question people in this line of work hear more and more often, in a tone that sits somewhere between curiosity and pity: if artificial intelligence can generate interfaces, copy and code in a matter of seconds, what exactly are designers for? It is a fair question, and it deserves a serious answer — because the stock reply, "AI is just a tool", stopped being good enough quite some time ago.
Years ago I wrote that function is a commodity: once everything works, value shifts to the experience. AI is now repeating that shift, one rung up the ladder. This time it is the solution itself that is becoming a commodity. Generating a prototype, rewriting a text, producing a component, assembling an interface: work that used to take days of specialist effort now takes minutes. And whenever making things becomes easy, value migrates towards whatever remains scarce.
So what remains scarce? Two things, one upstream of production and one downstream: understanding which problems are worth solving, and proving that the solution worked. The middle stretch of the process — generation — is being automated. The two ends are rising in value. And it is the profession's good fortune that these are precisely the two stretches design has always looked after.
The proof: an "AI project" read as a design project
At INPS we ran what everyone calls an artificial intelligence project: simplifying administrative texts with generative AI. I regard it as a design project that happens to use AI, and the difference between those two readings is what this essay is about.
Consider what actually happened, move by move.
We identified a need. Public texts were written at a level of complexity that shut millions of people out of understanding their own rights. A measurable problem — the readability index put a figure on it — and one worth solving, because a text nobody can understand is a barrier to accessing a benefit.
We designed a solution. Manual rewriting takes time and does not scale: across hundreds of service pages, the artisanal route was a dead end. AI removed the constraint of scale, but on its own it was not enough: the process needed a human in the loop checking every text, a framework of metrics to validate readability and fidelity to the original content, and a guarantee that the simplified text retained full legal validity. All of that architecture — what to automate, where to keep the human, what "good" actually means — has a name, and the name is design.
We measured the impact. The average Gulpease readability index rose from 57 to 64, bringing the texts within reach of roughly ten million more people. And in a blind validation with 1,620 users, 71.4% preferred the machine-rewritten version, which beat the originals on every perceived dimension: clarity, fluency, even trust. The full story is in the dedicated essay.
We made the result communicable to leadership. Simple figures, direct comparisons, an experiment anyone can grasp. That is the step that turns a good project into a precedent.
By the end of the journey, the question "what has design got to do with AI?" turns itself on its head: identify a need, design a solution, measure the impact, communicate it. This is design. Even when it sounds as though we are talking about something else.
What AI did, and what design did
It is worth being precise about the division of labour, because that is where the answer to the opening question lives.
AI did the part that previously refused to scale: generating the rewrites. Valuable work, and no human editorial team could have sustained it at that volume. Everything else — choosing that problem out of a thousand candidates, setting the quality criteria, deciding the human stayed in the loop, designing the experiment that makes the result credible to any audience — came from the discipline, not from the model. AI made the solution scalable; design made it reliable and provable. Neither, on its own, would have reached production.
And the direction of travel is already set: the next step we are working on goes beyond simplifying what exists, towards generating texts from scratch out of the legal sources alone, with every passage anchored to the provision that justifies it. The more generative capacity grows, the more weight the design questions carry: what should it produce, within which constraints, and how will we know whether it works.
The designer in the age of AI
All of this leads to two consequences, one uncomfortable and one liberating, and both need saying.
The uncomfortable one: anyone who has built their professional identity on a single step of the process — producing the artefacts — is right to worry, because that is exactly the step being automated, genuinely and fast. Defending the territory of artefacts is a losing battle.
The liberating one: design is a method. Identify a need, give shape to a solution, measure its effect, make the result communicable: artefacts occupy one segment of that journey, and the method runs through all of it. The designer is the interpreter of the whole process — the person able to conduct it from one end to the other, whatever tools happen to fill the steps in between. For those who define themselves that way, this is the best era the profession has ever seen: AI has supercharged precisely the segment that used to cost the most, and running the full cycle — from identified need to proof of impact — has never been faster or cheaper. Every hour freed from production goes back to where the interpreter's judgement is irreplaceable: spending time with users, understanding problems, defining what quality means, building the proof. A process with AI inside is still a design process. And a process always needs someone who knows how to conduct it.
Design has been the custodian of that method since long before the tools that now appear to threaten it existed. AI has made production abundant. Conducting the process — choosing the right problems, and proving we have solved them — remains the job. And it has always been the best part.