Architecture may be one of AI’s hardest pursuits. Unlike coding or writing, the data that could be used to train an architecture AI model is not widely available (or stealable) online. Rather, it’s stored away in the servers and physical filing cabinets of individual architecture firms. Digital drawings, 3D models, and even hand sketches are the blood and guts of an architecture project, and they’re all vastly more complicated than the résumé writing or HTML coding that AI tools have quickly mastered.
None of the big AI labs are currently attempting to tackle this challenge. That’s leaving the job up to the companies that actually hold all the data: the architecture firms themselves.
Architecture firms, both small and large, are actively building out their AI capabilities. They’re hiring data scientists and machine learning specialists. They’re running Shark Tank-style AI ideas competitions, vibe-coding bespoke plugins and apps to automate highly specific tasks, and even developing their own hyper-niche large language models that can help them create building forms and floor plans that reflect their signature style.
All this effort amounts to a broad recognition from across the industry that it’s a sink-or-swim moment for architecture. As the practitioners change the way they work—and as client expectations shift—some architecture firms are starting to realize that their own portfolios contain exactly the kind of information that can help them stand out in a business that’s only getting more competitive.
“Everyone has their little gold mine they’re sitting on, with all this data of past projects they’ve done,” says Faizan Zaidi, director of design technology at the architecture firm Spectorgroup. “But the question comes down to which firms are willing to build the tools on top of it.”
“Four years ago I got introduced to Dall-E, which I think was, for a lot of us, an awakening,” says Matthias Hollwich, cofounder of the 15-person New York-based architecture firm HWKN. “I decided to throw the whole office into an AI exploration for three months.”
Since that watershed moment in 2022, the number and variety of AI tools has exploded, and Hollwich has built on his early adoption. His firm is now regularly using many AI tools for its office, residential, and hospitality projects. Here are a few:
· NYC Zoning AI for zoning analysis and building massing that gives projects an overall size and shape
· Midjourney, MNML, XFigura, and Nano Banana for conceptual designs and renderings
· [TestFit](https://www.testfit.io/) and [Forma](https://www.autodesk.com/products/forma-site-design/overview) for site plans and analysis
· [Maket](https://www.maket.ai/) for floor plans
· [PermitProof](https://www.permitproof.com/) and [UpCodes](https://up.codes/) for building code compliance
· [Swapp](https://swapp.ai/) for construction documents
· Architectures for building information modeling (BIM) to digitally link 3D models and construction documents
Hollwich contends that the tools have fundamentally changed his firm’s approach to design, as well as what his firm actually produces. “Currently, most architects are trying to use AI to optimize the process, or maybe intensify some of their designs,” he says. “What we have done after these experiments is almost take a step back and say, no, this is actually bigger than just using these tools to come up with a better way to design the buildings. . . . There’s a new ideology that is emerging.”
The tools speed up the process of developing design concepts and ease the complexities of translating initial designs into wood and steel on a permitted construction site. But Hollwich says they also make room for deeper analysis that can make those projects better serve their intended users. AI tools have allowed his firm to do more deep research during the proposal and conceptual design phase, tuning designs to be more specific and contextual.
For one recent proposal, an office building in Amsterdam, HWKN used AI to analyze who the future tenants of the building would be, based on other companies and industries in the region. The tools helped the team quickly understand where those potential tenant companies are located and what types of buildings they occupy. AI also provided detailed information about their spaces, including floor plans, ceiling heights, and building amenities. “We created this whole matrix, and we basically said, ‘If you design this building according to this information, you’re going to be ahead of the game because every one of these companies you would like to attract will see themselves in it,’” Hollwich says. HWKN won the project, and now uses AI for other project lead searches and analyses. It’s also building a custom RFP (request for proposal) response tool using Claude.
“For us, there’s nothing the same from how we worked four years ago,” Hollwich says.
Spectorgroup, the New York firm that Zaidi is part of, has been dabbling with its own AI plugins and tools to work alongside existing design programs. Zaidi says the firm has been using AI-assisted coding tools to build custom software to tackle small but time-intensive parts of the design process. One example is a tallying tool used during the creation of test fits, the example floor plan designs architects mock up to give clients an idea of how, say, an office may be laid out.
“It looks at your plan, does the math, which it’s really good at, and tells you you have five conference rooms, you have two pantries, you have 300 desks for people, and it spits out this legend,” Zaidi says. “That was a manual, tedious task that’s now being replaced by AI.”
Zaidi says the firm has explored whether AI might be able to do the actual test fit itself, deciding where to place those conference rooms, pantries, and desks. But so far the technology hasn’t been able to match the expertise of human designers, who know not to place a noisy pantry right next to a conference room, for example.
“When I’m evaluating AI tools in their current state, what I’m seeing is, yes, they are good, they’re on the right trajectory, but they’re not there yet,” he says.
He and many other experts in the field believe, however, that it’s only a matter of time before these tools move beyond the tedious tasks and early-stage concepts. And this evolution is likely to start happening inside some of the biggest, oldest, and best-known firms in the world.
Gensler, the world’s largest architecture and design firm, has built what co-CEO Jordan Goldstein calls an AI sandbox. Less a space than a way of working, the sandbox launched in the early days of generative AI going mainstream, and has functioned as a kind of mirror design realm to experiment with AI tools.
Designers across the company’s 58 offices and 33 practice areas have been encouraged to take real client projects and firewall them into the sandbox, where they can try out AI tools to see where, if at all, they can be useful to the actual project at hand. At the same time, Gensler has been expanding its AI talent pool, hiring specialists in machine learning and data science.
“We sensed that there was a wave coming, and we started to really invest in experimentation,” Goldstein says.
In the early days, this was about learning which off-the-shelf tools could augment the work of Gensler’s designers, shaving time off conceptual renderings or accelerating site analyses. But as the technology evolved, so did the sandbox, eventually turning into a kind of AI breeding ground.
Designers began turning their experiments into bespoke tools, vibe-coding systems and interfaces that solve specific problems or speed up portions of the design process. That prompted the firm to launch its Innovation Challenge, an annual juried pitch showcase, now in its fourth year, that takes the best of those ideas and turns them into tools for the firm to use.
This focus on experimentation and creation led to a suite of internal tools Gensler has rolled out across the firm. The first, named gBlox, is a design and digital massing AI plug-in that uses text and image prompting to develop strategies for project sites. Another, gFloorz, applies predictive analytics to early-stage floor plans to give clients a window into how spaces will be used and what that means in terms of their sustainability efforts, operations, maintenance, and even return on investment.
Additionally, the firm is using AI-driven filmmaking tools to add custom short films to project proposals, offering clients a predictive vision of a day in the life of someone using a building that, at the proposal stage, is still just an idea. “That’s an incredibly engaging way to connect with a client in a marketing pitch,” Goldstein says.
For Gensler, which made $1.87 billion in revenue in the 2025 fiscal year and regularly ranks as one of the top-earning architecture firms, deep pockets make these kinds of efforts possible. But Goldstein argues that spending alone won’t be what gives architecture firms the edge in putting AI tools to use. “We believe that the organizations that will have the lasting advantage, certainly over the next decade, will likely not be those with access to the best models, because those capabilities are becoming increasingly available to everybody,” he says. “The differentiation will come from how effectively a firm like ours continues to use a deep domain of data with the additional creativity that we’re bringing to the table to create lasting value for clients.”
KPF, a 50-year-old firm of 650 people, has been using AI to operationalize its vast portfolio, turning old projects into sources of data that can inform or even shape the projects of today.
Luc Wilson, director of global design technology at KPF, says AI tools have been used to analyze and describe every image in the firm’s database, an estimated 200,000 in all. The past four years’ worth of project proposals can be queried or prompted to create new proposals. Complex design specialties, like the core designs used in tall buildings, are cordoned into their own specific databases to provide more detailed responses to queries.
These vast pools of data are now accessible through an AI system KPF built, called KAI. “It’s very much in the vein of Claude or ChatGPT,” Wilson says.
But unlike those tools, which run on large and complex models, KAI runs on several focused databases, as well as KPF’s entire project portfolio and intranet. A designer can use KAI to find images of every office lobby with hardwood furnishings that KPF has designed, or precedents for the placement of elevator shafts in 60-story buildings, or details on which people in the firm have worked with a specific client before. “It’s saving a lot of digging,” Wilson says.
KPF also added a vibe-coding subagent to the system, allowing people to use simple text prompts to build interfaces and web apps, all of which get deployed to KPF’s controlled servers, not the broader internet. Within the first month of launching that subagent, more than 400 apps were created. And it wasn’t just the youngest, most tech savvy employees coming up with new ideas.
“A principal vibe-coded her own interactive web map to help track leads in a new region, which was awesome,” Wilson says. “One of the things that’s exciting for us is now the person closest to the problem can solve it.”
The firm is also using these tools to power image creation, tapping what KAI digs up to then image- or text-prompt various image diffusion models, all while aiming toward outcomes that look like the kind of work KPF has built its business on.
“We’ve spent a lot of time on the system prompts . . . the thing that the model reads before it executes any task,” Wilson says. “That’s where you can really embed organizational standards, best practices, even aspirational goals in terms of how it directs the model to respond. It has made it much more KPF.”
Foster+Partners, the U.K.’s largest architecture firm, has been exploring AI and machine learning since 2018 and currently has more than 120 AI models at its disposal.
“We are using large language models, diffusion models, statistical methods, and optimization techniques across the entire design process, from early design explorations and performance feedback to applications that boost knowledge dissemination, productivity, and operational insights,” explains Martha Tsigkari, the firm’s head of applied R&D.
Through the firm’s AI Portal, Foster+Partners staff can query 60 years of design guides and projects in natural language. They can use various AI-augmented configurators for urban plans, stadiums, and adaptive reuse designs. AI-assisted regulation comparison helps with document ingestion and report generation for the firm’s work around the globe. AI agents can generate parametric models from simple images.
“The most visible shift is AI moving from tool to actor, with AI agents ultimately orchestrating entire design pipelines,” Tsigkari says. “We are excited about what is coming: an API-first approach combined with Model Context Protocols will enable agentic workflows, with AI moving fluidly between tools and data sources without human intervention at every step.”
But with the entire toolbox the agent will struggle, Tsigkari says. “Too many capabilities, too little clarity.” Instead of building a monolithic, top-down system, Foster+Partners has developed a modular set of components that designers can plug into when needed, automating small pieces of the design process. Tsigkari says this is all about optimizing the process without losing control of the design.
“We have made a conscious choice to not just automate but completely reinvent the production aspect of our design process, rather than the creative one,” she says. “We have been process-mapping our workflows to identify which parts of the process can potentially be automated.”
So much of this is built on the firm’s long history, but Tsigkari says the unending flow of new project data is essential to keeping the tools trained and productive. Foster+Partners has been developing a strategy for automating data collection, organization, and processing across disciplines, helping its AI tools’ ability to use new sources of data while cutting down repetitive work for staff and designers.
“The same source feeds websites, reports, AI models, and design workflows. It is cross-referenced, agent-traversable, and live,” Tsigkari says. “We want to author knowledge once and use it across workflows while we continuously train the system through everyday work.”
For the 35,000-person design, engineering, and consulting firm Arcadis, tapping institutional knowledge can be daunting. It’s taking on thousands of projects a year, ranging from designing ports and building semiconductor factories to planning water treatment facilities. Name any type of building project, and the company has almost certainly done at least a few. But the knowledge behind those projects is often buried somewhere inside a database or archive, out of sight and out of mind. “When we hire someone new, they don’t really have any direct way to access or learn from what happened five years before they joined,” says Jason King, a principal of architecture and urbanism at Arcadis. “So many problems that they’re going to face have been solved over and over and over and over and over.”
As a specialist in computational design, King has been using AI to try to make the firm’s deep institutional knowledge more readily accessible. Focusing on specific practice areas, like residential architecture, he’s helped build AI systems that can take lessons from the projects of the past and apply them to new ones.
For example, King led the creation of a database of about 7,000 of the firm’s residential projects from the past 20 years—which itself is only a portion of the total—populated with key details, thumbnail images, and CAD drawings. That data became the training material for machine learning models that can now be used to accelerate initial designs. “When we have a new project, we can upload the specs of that project, and we can have a generative tool make a floor plate for that project site, down to unit layouts,” King says.
Designers need only upload a sketch or basic CAD drawing of a new project, and the system will pull up work from the portfolio that most closely matches it. “It’s a way for people who don’t have any programming skills to interface with thousands and thousands of floor plates that have been done over the past 20 years,” King says.
Arcadis is deploying these kinds of tools across its practice areas, as well as using more standard AI tools for optimizing business tasks like creating expense reports and setting up construction drawing sheets. King says the utility of these tools is obvious, but the firm is proceeding cautiously with design-focused tools like the floor plate generator, rolling them out only to small teams for evaluation. “It’s still something I would say is an R&D tool,” he says.
It likely won’t be long before these generative design tools become essential to the way the firm works, tapping into a gigantic portfolio of experience that few other firms can match. The result, King says, will be projects that build on the strengths of past projects. “I’m interested solely in how do we design better buildings and better cities, not how do we print faster,” King says.
For all the power firms are finding in their often massive project databases, the biggest moves in AI for architecture may be coming from the company behind the industry’s design infrastructure: Autodesk, the $53 billion software company that builds industry-standard digital design tools like AutoCAD and Revit. The company has built AI models from scratch, using its own CAD files as well as many offered up by its client base, which is basically every architecture firm in the world. (In exchange for accessing Autodesk’s advanced AI tools, clients agree to allow some or all of their project data to be used to train the models.) Unlike the LLMs behind ChatGPT and Claude that rely primarily on text, Autodesk has developed AI models trained solely on the geometrical data of the physical world.
These models, dubbed neural CAD, are designed to understand the three-dimensional nature of buildings and the interconnected systems within them. Mike Haley, SVP of research at Autodesk, has been leading the development of neural CAD for nearly a decade, mostly in stealth. It’s work that started years before generative AI went mainstream and captured the imagination of architects.
Haley says Autodesk opted to take a more deliberate approach by building an AI model that could actually understand space rather than just describe it or try to translate a description into an image. “It’s not a language model. It doesn’t look like a language model. It solves a different problem,” he says.
By training on actual CAD models showing millions of buildings and rooms, neural CAD was developed to understand where architects place things like walls, columns, stairwells, windows, and doors, and how they relate to each other, as well as how they relate to the structure of the building, from slabs to trusses to beams.
“It’s not trying to reproduce a specific person’s building, but it’s looked at all of those repeated fundamentals that are inside every single building, and it’s learned those patterns,” Haley says.
Autodesk has created specific neural CAD models for the three industries it builds most of its software for: architecture, structural engineering, and manufacturing. Its architecture-focused model can essentially be prompted to design a building at a high degree of fidelity, which could represent a dramatic evolution for the architecture industry and the people who work in it.
AI is not replacing architects, and neural CAD is hardly the new Claude or ChatGPT, but it does have a very focused depth of knowledge that is unlikely to be matched by the more well-known AI models.
“If you look at the size of some of the frontier models, our models are several hundred times smaller,” Haley says. “But they’re actually better at solving these very specific problems. And not only are they better, but the actual cost of them is far, far less.”
In the architecture industry, Autodesk is filling a hole most firms didn’t even realize existed until generative AI broke out in 2022. “I remember a whole bunch of companies coming to me [in 2022] going like, ‘What do we do? This is a new thing. Should we use ChatGPT? We don’t know,’” Haley says. “And then the next year, everybody came along and said, we’re all going to build our own models.”
Haley was quick to burst that bubble. “Really none of them, even if you’re an architecture firm that’s been around for 100 years, have enough data. I’m sorry,” he says. “It requires not just a volume of data, but a variance. You want that breadth of different designs and different approaches. So you’re just not going to have that, even in a very large company.”
Autodesk, as the main digital medium for how the architecture industry operates, had that power. And while big firms like Gensler or KPF can fine-tune their tools to produce concepts and designs that match their own styles and project types, neural CAD seems to be able to answer more general questions about buildings while also being more specific about how they actually get designed and built.
Haley says Autodesk’s AI models are trying to do something more transformative for the industry. The AI tools that run on neural CAD have only just started rolling out to AutoCAD users over the past few months. Haley says new developments are imminent. But at the same time, he doesn’t see AI taking out large swaths of the architecture industry. Nor does he see AI eliminating the need for the design software Autodesk sells.
“The ability of these AI models to solve every problem is not evenly distributed today,” Haley says. “There are some of these things we can totally nail and some of them we’re not going to have solutions for for years to come.”