Here is a number worth sitting with. Of all the houses that will be built in the world this decade, only a small fraction will ever pass through an architect's hands. The rest get drawn by a contractor on site, copied from the neighbour's, or not drawn at all. That is not a failure of taste. It is a failure of access. Professional design costs more than most households can pay, and takes longer than most builders can wait.
We started as architects, and this bothered us more than any building we could have designed ourselves. Two people in a studio, working carefully, might see forty houses through in a career. The city needs millions. However good the forty are, the arithmetic says the profession is not where the built environment actually gets decided.
So when people ask what AI means for architecture, our answer is not renderings. It is reach.
We have already run one version of this. A design-service network in South Asia wanted to sell professional house plans in small cities at a price a family could pay — a few thousand rupees for a plan, sections, elevations and a 3D model, rather than a fee that only a developer could justify. No studio could do that work by hand at that price. What we built instead was an engine inside the CAD tools the architects already used: requirements came in through a structured form, the system generated layout options, and the architect selected and refined rather than starting from a blank sheet.
The part that mattered was not the generation. It was the checking. Every plan the engine produced was scored against rules the reviewers had been holding in their heads — minimum room sizes, natural light and ventilation in every habitable room, wet areas stacked over one another so the plumbing runs straight, a staircase wide enough to carry furniture up. Those checks used to depend on which reviewer opened the file. Now they ran before anyone opened it. The architect's judgement moved to where it earns its keep: choosing between good options, not catching bad ones.
That is the shape of what we think 2030 looks like. Not an AI that designs the city, but a layer of automated competence underneath every design decision, so that the baseline of what gets built is professional even where the professional never visits.
Take the word accessible literally for a moment, because it has two meanings and both apply. The first is the one above: design that people can afford and actually reach. The second is the older one — buildings and streets that a wheelchair, a pram, a bad hip or a pair of tired eyes can use. In most of the world the second is regulated, weakly enforced, and checked at the end, when changing anything is expensive. It is exactly the kind of rule that belongs in the engine rather than in the review. A door width, a turning circle, a step-free route from the street to the bed: these are geometry, and geometry is what the tools are good at. Put them in at generation time and they cost nothing. Find them at inspection and they cost a rebuild.
Cities work the same way, one scale up. The questions a planning department asks about a block — daylight to the neighbours, distance to a bus stop, how many homes a piece of land can carry without overwhelming the drains — are answerable from data that already exists. Today they get answered slowly, by a few people, for the projects big enough to pay for the study. The tools now make it possible to answer them for every plot, before the drawing starts, and to show the trade-offs to whoever is going to live with them.
We should be honest about what is not solved. A model that has seen ten million floor plans will produce a plausible eleventh, and plausible is not the same as good. It will reproduce the mistakes of the ten million along with their virtues. Which is why the rules matter more than the generator: someone has to decide, explicitly, what a good home is in this climate, on this budget, for this family — and write it down where the machine can read it. That is design work. It is arguably the most important design work there is, and almost nobody is paid to do it yet.
There is also the question of who owns the layer. If the checking engine belongs to a vendor, the standard of what gets built belongs to the vendor too. The version we built sat inside the client's own tools, on rules they could read and change. We would argue for that every time. The rules are the public interest; the software is just where they live.
So, 2030. We do not expect the architect to disappear. We expect the job to split. A smaller part of the profession will keep doing what it always did — the bespoke, the civic, the one-off — and that work will be better for the tools. A larger part, we hope, will turn to writing the rules and building the engines that carry professional judgement into the millions of homes that never had it. That second job is less glamorous and matters more.
We would rather be in the second group. Not because the buildings will be beautiful — some will, most will be ordinary, and ordinary is fine — but because a city where the baseline is competent, and where the door is wide enough, is a more accessible city than the one we have. That seems worth the next five years.