The small studio advantage

Every few years somebody announces that the small practice is about to win. It keeps not happening, and the reason is dull rather than dramatic. Big practices can afford to keep people on the payroll that small ones cannot.

RIBA's benchmarking makes the point better than an argument would. Seventy-eight per cent of UK chartered practices have between one and nine staff, and between them they take 13 per cent of the revenue. Practices of fifty and up are 4 per cent of the profession and take 61 per cent. Same job title, two different industries.

What the money buys is permanence. Get big enough and there is reliably enough work coming through the door to keep somebody in computational design, somebody on bids, an environmental analyst, a BIM lead, a communications person, possibly a software developer. Five people cannot carry ten specialisms against the possibility that each one gets needed in March. So breadth of capability has always meant breadth of payroll, and that has held for as long as anyone reading this has been practising.

That equation has started to wobble, and faster than I expected. RIBA puts AI use in UK practice at 41 per cent in 2024, around 59 the year after, 74 this year. Adoption curves are boring on their own. What caught my attention is where it has turned up: not only early design images, but technical resolution, compliance checking, project and information management, and the administration of running a business. That last one is quietly the important one, because the overhead of being a practice is a decent part of why practices need to reach a certain size.

Then GPT-6 Astra landed on 3 September, still on a staged rollout, and that is the thing that made me want to write any of this down.

What matters is not that it is cleverer. It uses a computer. OpenAI's own examples are mundane in a way I find far more convincing than a demo reel would be: filling in forms and updating records, researching something and writing the result into a document, analysing data and plotting it, building a website and running the frontend QA on it, installing and testing software, working out what has gone wrong on screen, producing a spreadsheet or a deck that follows a template you already use.

For thirty years the chain has run person, software, output, and the software has been the bottleneck. Not because the tools are bad. Because somebody has to know them, and that person's time is the ceiling on what anyone else can attempt. Put an agent in the middle of that chain and the ceiling moves.

One number stopped me properly. Astra is reported at 95.9 per cent on BenchCAD, which shows a model a part and asks it to write the parametric CAD code that rebuilds it. Before anyone gets carried away, BenchCAD is bevel gears, compression springs, twist drills and threaded adapters, drawn from engineering standards. It is not architecture and it is not close. But look at the loop it describes. See the geometry, infer the parameters, write the code, run it, look at what came out, go again. That is a fair description of what a computational designer does all day, and it has always required somebody who spent five years inside one application.

The same announcement carries its own corrective, which I appreciated. Astra manages 72.6 per cent on OSWorld 2.0, a computer-use benchmark, and 41.4 per cent on AutomationBench, which sits nearer to sustained professional work. So it can drive the machine. I would not leave it alone with a deadline.

There is a BCG study I keep coming back to here. Consultants with no coding experience, handed generative AI, improved by up to 49 percentage points on work well outside their own discipline, and reached 86 per cent of the benchmark set by BCG's own data scientists. Eighty-six per cent is a lot. It is also not a hundred, and the missing fourteen is roughly where the expertise lives. So an architect who has never opened Houdini does not become a Houdini person. An architect who understands geometry and fabrication and can think computationally might get an agent to do part of a Houdini job without losing a decade to learning the node graph. Two very different sentences, and only the second is true.

The word has always described a payroll. You were multidisciplinary if you employed the disciplines, and the organisation chart was the evidence.

What changes is the tissue between them. A small team can move from architecture to visualisation to film to computation to interactive work to research to software without setting up a department at each stop. Worse than a specialist would do it, usually. But well enough to carry an idea across a line that used to cost a handover, a fee and three weeks.

I should be upfront that this is not a neutral observation from us. Afterform already works across architecture, visualisation, animation, real-time environments and AI with a very small number of people, and the honest account of how that happens is not that we have departments. The gaps just got cheap to cross. Some of the machinery is in the piece on running our models locally, and what comes out of it is in selected work.

Here is the finding that keeps the whole thing honest, and the one I would put in front of anybody trying to sell you an AI strategy.

In RIBA's 2026 report, three quarters of AI users say their productivity improved. Fifty-seven per cent report a positive return on the money. Seventeen per cent think their designs are better.

Seventeen. After all that.

You could read it cynically, as a profession defending its own mystique. I don't think that is what it is. It reads to me like people accurately reporting that they are producing more and not producing better, which is exactly what you would expect from tools that are extremely good at execution and have nothing to say about whether the thing was worth executing. So the claim I would defend is a narrow one. AI does not make five people as good as fifty. It gives five people some of the capacity that used to need fifty, and none of the judgement.

Nothing in the right-hand column is new. Architects have been claiming judgement and taste as the value for a century, usually while sending the invoice. What is new is that it has stopped sitting downstream of headcount.

Two things stop me getting carried away. The first is that none of this is democratising anything yet. In RIBA's 2025 data, 83 per cent of practices with fifty or more staff were using AI against 48 per cent of those under ten. Big firms have more money, more data, research time and people whose entire job is working these systems out. Whatever advantage is going here has to be taken. Nobody hands it over. Although I would argue the edge is not capital but how fast you can turn: five people can throw out a workflow over a weekend, five thousand have procurement, security review, licensing, training, governance and fifteen years of legacy decisions in the way. That is an argument rather than a statistic and I am presenting it as one.

The second is uglier. In the same survey, 59 per cent expect AI to reduce headcount across the profession, and 61 per cent think it will make it harder for early-career architects to pick up the skills. The work agents absorb first is the drawing sets, the schedules, the coordination, and that is the work through which people learn to be architects. If the smallest workable studio becomes three seniors and a pile of agents rather than three seniors and seven juniors, nobody has explained where the next seniors are supposed to come from. I have not got an answer either. It is the part of this I find uncomfortable, and I notice how rarely it comes up in the enthusiastic version of this conversation.

Everybody reaches for the generalist at this point and I want to be careful about it, because the generalist who knows a bit about everything is as useless as they have always been. More dangerous now, if anything, since the output turns up looking finished.

The useful version knows enough across several fields to ask the right question, tell whether the answer is any good, work out who or what should take it on, name the constraints that bind, spot the part that is wrong, recognise when they are out of their depth and need a real expert, and hold the whole lot together as one piece of work.

Which is architecture, described slightly differently. It has always been an integrating profession, the discipline that holds everyone else's expertise in an arrangement and carries the responsibility for how it turns out. What is changing is how many things one person can hold at once.

And none of this touches what large practices have. Liability. Accreditation. Institutional memory. Quality assurance that works. Procurement credibility, client relationships built over decades, and the ability to put two hundred people on something enormous. RIBA is blunt that professional judgement and competence remain necessary wherever design decisions affect safety, sustainability and wellbeing, and it is right, and the distinction matters most exactly where buildings can hurt people. Competing with some of the range of a multidisciplinary firm is a different thing from replacing what it is licensed to do, and I would rather say that plainly than let the argument run further than it deserves.

The direction still seems clear enough. The odd consequence of better AI might not be bigger companies at all. It might be much smaller ones having a go at considerably larger things.