Why AI reached law before architecture
Architects took to software early and to AI late. The order had a reason, and this fall the reason went away.
Investors ask us some version of the same question. If AI is this capable, why hasn't it changed architecture the way it changed law?
Part of the answer is diffusion. A wave like this needs founders who focus hard on one industry, and some industry has to go first. Which one went first wasn't random, though. Law was the first profession whose core work matched what language models could do.
Law was text in, text out
In 2023, Harvey and others saw that language models could already digest large volumes of text, which is most of legal review, and would soon draft structured, clear documents. Law was a huge industry, a natural fit and a clear path to revenue. As agents matured, the same companies started calling their products agentic operating systems and moved into inbox management, expenses and the rest of a lawyer's admin.
Two layers of work
An AI system for any profession has two layers.
The first takes administrative load off the professional: a proactive assistant that handles routine tasks and the mental overhead around them. That only became buildable around the start of this year, which is why assistants like Instinct and Dots are getting attention now. It is also what we explored with Collie, our first product. On its own it is generic, and a weak way into any one profession.
The second does meaningful parts of the actual work, and this is where architecture differs. Before 2023, architects were arguably more software-savvy than lawyers. Today it's the reverse. Lawyers have excellent AI-native tools, while architects still rely on CAD, 3D and rendering software built on conventional machine learning. That isn't because nobody tried. Until very recently, AI couldn't do anything useful inside that stack.
What we saw in the studios we visited
This month we've been sitting with small architecture and landscape practices. Their staff already use ChatGPT and Claude, then carry the results into AutoCAD, InDesign and Illustrator by hand. One landscape practice spends about 40 hours on each proposal and writes one most weeks. A single image of a potential project takes them around six hours, partly because the site information is missing and steps get done twice. What they asked for was tools built for architecture software, so they wouldn't have to go through a chatbot at all.
What changed this fall
The unlock is an AI that can operate architecture software directly, reliably enough not to break things, with everything about a project in one place. Two capabilities had to arrive for that. Models had to become useful with 3D: reading a model, reasoning about geometry, setting up a render. And they needed general computer use.
My co-founder Liam Shepard made the second point. The architecture stack is so entrenched that nobody would adopt an AI-native product that asked them to replace it up front. You have to work across the software studios already own. Connectors help where they exist, and computer use is what makes an agent general enough to cover the rest.
Fable 5 made this possible, at a cost no studio would pay. Opus 5.5 made it practical.
2023
Models read and write
AI goes to law first, because review and drafting are text in and text out.
Early 2026
Assistants take on admin
Inbox, expenses and scheduling can be handed off. Useful in any office and specific to none.
Fall 2026
Models operate design software
Computer use and 3D reasoning get reliable and cheap enough to work inside CAD, BIM and rendering every day.
Next
One assistant across the whole project
The admin and the design work of a project run from the same memory, inside the software a firm already uses.
Where it goes
The next step is breadth: one assistant that handles both layers of a project from the same memory, inside the tools a firm already runs, so nobody has to carry files between programs by hand.
In 2023 it was clear that AI would transform legal drafting and review. Today it is as clear that AI will transform architecture, and nobody could have built a meaningful AI-native stack for architects before this fall. That's the window we're building Alvar in.