After years of impressive but ultimately rudimentary 3D demos, AI is moving from generating static assets to generating worlds. Image and video models and inference providers have started touting architects as end users, because the whole industry can see that prompting realistic 2D and 3D worlds is close. We may have finally arrived at the moment where software can understand space.

Running alongside these spatial advances is an equally consequential shift. LLMs and agents have improved to the point that they can orchestrate entire workflows while respecting the codified rules of structured domains. That pushes the prize beyond software seats and into the labour budget within those domains: work that once took months collapses into days, and an operating-scale layout that took an engineer the best part of an hour takes under a minute. Two frontiers are converging: software that can generate and manipulate space, and software that can perform the expert work of turning space into something buildable, compliant and real.

That matters because one of the largest software categories in the world is still built around specialist tools, heavy files, local workflows, and decades-old interaction patterns. Autodesk's own annual report says the company is rolling out AI capabilities built on frontier vision-language models and proprietary foundation models that understand “3D design and make.” While it’s unclear whether Autodesk will successfully be able to pivot to AI driven design; the question is no longer whether AI can generate something that looks like a room, a building, or a city. It can. The real question is whether AI can produce something that architects, engineers, designers, and builders can use.

Today, the answer is mostly no. Almost every architect and spatial design founder we speak to tells us the same thing: AI is magical, but it is not yet professional. It creates images, not workflows. It produces inspiration, not drawings. It generates worlds, but not yet buildings. Moreover, Architecture, Engineering and Construction (often abbreviated as AEC), is not just about images. It also involves multiple text files, as well as spatial constructs that demand 3D editable geometry.

At Dawn, we’ve been spending time with founders and practitioners across architecture, CAD, BIM (building information model), spatial design, construction workflows, and world models. In this piece, we share our thesis on why incumbent software may finally be rebuilt, and what needs to be fixed before AI-native spatial design becomes enterprise-grade. Prompt to image was the first act. Prompt to world is the second. Prompt to buildable system is the real opportunity.

CAD stands for computer-aided design. At its simplest, it is software that helps people create, modify, analyse, and document designs. In AEC, CAD is the system through which physical intent becomes technical representation: a plan, a section, a component, a tolerance, a material, a machineable part, or a construction drawing. CAD can be represented in 2D or 3D formats.

BIM, or building information modelling, is related but different. If CAD is often about drawing or modelling geometry, BIM is about creating a digital representation of a built asset that includes both physical and functional information. A BIM model contains walls, doors, windows, materials, quantities, systems, schedules, and relationships between components. In theory, when one part of the model changes, the documentation and downstream views update with it. BIM is more than just a model: it is a coordination system for a building’s lifecycle, from design and construction through operation.

MEP (mechanical, electrical, and plumbing) is the engineering layer that makes a building function: HVAC, power, water, drainage, fire alarm, life safety. It is among the most rules-heavy and labour-intensive parts of the chain.

A building moves through several stages, from a client brief to a finished, occupied building. The design phase includes the brief, concept and developed design, and the technical engineering and coordination that resolve how structure, MEP and envelope fit together. Afterwards, a project passes through compliance and permitting into procurement and pre-construction, then construction itself, and handover and operation. This sequencing matters because design is a sliver of where the money sits. Fees for design run to a single-digit percentage of total project cost, while pre-construction and construction consume the rest.

Each of these layers is owned by an incumbent, and mostly by one company. AutoCAD is the standard for CAD drafting, its DWG format the common currency consultants pass work in. Revit dominates BIM and much of MEP, with ArchiCAD the main alternative. Autodesk, which owns both AutoCAD and Revit, sits across most of the stack. These are embedded systems of work, each with its own file formats, standards, firm habits, training, and liability, and a single building passes through several of them, handed between firms as files. Rebuilding CAD is therefore not about a better drawing tool. A new entrant has to take one slice of this stack outright while reading and writing cleanly to every incumbent beside it.

The comparison cuts both ways. No one has rebuilt the Adobe, Office or Atlassian suites from scratch, but Figma, Notion, and Linear each took a single surface off their legacy competitors, did that one thing far better than the incumbent, and built a large business on it. Taking one slice of the stack is exactly the path open here, and the prize is large. These are gigantic value pools sitting on tools whose daily experience now lags what people expect from software with Autodesk alone making $7.2bn in revenue in FY2026.

AI applies across all of these stages to varying degrees, but the opportunity is not evenly distributed.

The ‘pain’ is concentrated in spatial coordination, which is where the design takes shape and becomes real. It starts with the plan: floorplate options, circulation, area schedules, daylight. Most of that is still done by hand, and it is where AI has historically failed, because you may, for instance, find yourself asking for a lighter facade and see the structural grid drift with it.

That failure is the one editable-geometry models solve, because they can generate and compare options while the hard constraints hold. Move the core, add daylight, hold the floor area, compare three circulation strategies, export the plan. From there the design is coordinated: walls take on thickness, material, and fire rating, building systems go in, and cross-discipline coordination begins. The heavy files and virtual-desktop workarounds slow this stage to a crawl, a model shared across numerous consultants is brittle, and collaboration is really file-syncing and annotation rather than anything like multiplayer.

What a serious tool has to deliver here is version control: select a wall and change its thickness, lock the stair core, tag a fire door, export the schedule, see what changed between versions. That is hard and unglamorous, and it is what the AI-native BIM platforms are trying to own.

Then comes the part of the chain with the clearest return: technical design. Structural engineers size members, MEP engineers route ducts, power, drainage, and fire alarm, clashes get resolved, and everything is checked against code. It is repetitive, rules-bound, and measured in hundreds of hours. Running alongside all of this, from spatial coordination onwards, is compliance and permitting. Every drawing is checked against jurisdiction-specific codes, permits are filed with local authorities, and the compliance documents that govern the project are reviewed by hand. Today this runs on manual code-checking and external consultants. The AI openings are narrow but real: automated code and drawing review, permit preparation, and extracting structured data from dense regulatory and contract documents. None of it requires replacing the modelling tool, and the buyer is often the consultant or contractor rather than the architect, so the economics differ.

The first problem is consistency.

For architecture, inconsistency is fatal. If a user asks for a lighter facade, the floorplate cannot change. If they adjust the lobby, the structural grid cannot drift. If they regenerate the interior, the windows cannot move. Backgrounds, context, room proportions, site boundaries, neighbouring buildings, and design intent need to persist. This is the spatial version of hallucination. In legal AI, one hallucinated citation can sink trust. In CAD, one hallucinated dimension can make the output unusable. But internal consistency is not enough. Any product in this space needs to understand the rules across codes, clearances, adjacencies, structural and fire constraints, and maintain perfect adherence across every iteration.

“Every prompt cannot be a new universe. If I ask for a material change, the model cannot quietly move the window, change the proportions, or redraw the background. In architecture, that is not a variation. That is a broken file.”

- Architect at Tier 1 architecture firm

World models are interesting because they explicitly move towards persistence, consistency, and controllability. But professional design requires consistency at a much harsher standard than interactive exploration. Architects do not need a model that remembers a room for a few minutes. They need a system that remembers a project for months.

The second problem is editability.

A generated world can look spatially coherent and still be useless for a professional. Architects need to select the wall, change its thickness, lock the stair core, tag the fire door, export the schedule, and know what has changed between versions. Engineers need structure. Contractors need quantities. Clients need options they can understand. Regulators need documentation. This is where many AI-native tools fall down. They generate pixels, meshes, splats, or videos, but not editable systems of work.

The frontier is not prompt-to-render. It is prompt-to-editable geometry, prompt-to-parametric model, prompt-to-BIM object, prompt-to-simulation, and eventually prompt-to-procurement. The most interesting companies may therefore look less flashy than the best demos. They will spend time on boring things: layers, symbols, blocks, annotations, constraints, file imports, comments, permissions, export formats, and latency.

The third problem is collaboration.

Architecture is an intensely multiplayer workflow. A single project can involve architects, interior designers, structural engineers, MEP engineers, contractors, planning consultants, clients, quantity surveyors, landlords, suppliers, and local authorities. Many of these stakeholders sit in different companies, use different tools, and join the process at different levels of technical sophistication. And yet the files are still hard to share.

Real-time collaboration in CAD and BIM exists, but it does not feel like Figma or Google Docs. Changes often sync locally and then update across machines. Elements may need to be locked to avoid clashes. Large files create delays. Consultants link models across different software environments. PDFs still become the lowest common denominator.

This is not a marginal inconvenience. It shapes the entire workflow. If a designer cannot easily share AI-generated work with a colleague, consultant, or client, the work remains personal inspiration rather than organisational output. If a model is too heavy to open on a laptop, the browser-native dream breaks. If collaboration only happens through annotations, the tool is not truly multiplayer. The Covid-era shift to Miro, Figma, Notion, and Google Docs changed user expectations. Architects now live in a modern collaboration stack outside their design tools. They know what good feels like:CAD does not yet feel like that.

“We believe AEC design software will be rebuilt around two foundational shifts: cloud-native multiplayer collaboration and agentic AI. Multiplayer collaboration solves today's coordination problem; agentic AI solves tomorrow's productivity problem. But these shifts are inseparable. As AI agents become active participants in the design process, collaboration will evolve from human-to-human to human-to-agent and eventually agent-to-agent. The design platforms that succeed will therefore be those built for collaboration first, because collaboration is the substrate on which agentic workflows will emerge.”

- Bastien Dolla, CEO & Co-Founder, Rayon

The fourth problem is performance.

For years, architects tolerated slow software because the work was heavy and the alternatives were worse. That tolerance is collapsing. Rendering that once took an hour can now happen in seconds. Consumer AI tools respond instantly. Designers expect to iterate, compare, and undo in real time. In spatial design, latency is not just a technical metric. It changes the creative process.

This is why real-time generation, real-time collaboration, and real-time simulation are so important. The killer feature for architects is not just that AI can generate a building. It is that the architect can stand inside the design conversation while it changes.

So, can AI-native AEC software finally be rebuilt?

Yes, but not because AI can draw a beautiful building.

“Technology for the built world was always constrained by deterministic software trying to handle a messy, edge-case-heavy vertical. The result was tools that worked for simple cases and broke for everything else, or solutions that served one of the several dozen stakeholders across the building lifecycle, but never the whole thing. AI changes that. For the first time, you can build technology that handles the full complexity of the physical world. Once you realise that everything outside of the building is simply a document, it unlocks extremely ambitious companies that could never have existed even two to three years ago.”

- James Stirrat-Ellis, Co-founder & CEO, Build.inc

Software here can be rebuilt because the interface to spatial work is changing. For the first time, software may be able to move from representing what a designer manually encodes to actively participating in the design process itself. It can generate options, understand context, maintain a spatial memory, reason across constraints, and collaborate with humans inside the model.

But the difference between a demo and a company will be precision. The model cannot forget the building. The geometry cannot be decorative. The workflow cannot stop at inspiration. The output has to be good enough to share, good enough to edit, and eventually good enough to build.

In legal AI, nobody wants an AI lawyer that makes mistakes. In spatial AI, nobody wants an AI architect that moves the walls when you change the lighting.

Autodesk will not be rebuilt by a prettier rendering tool. It will be challenged by a new design environment where the world itself becomes the interface, and where intent can finally become something editable, collaborative, and buildable.

If you’re building in AI-native CAD, spatial design, MEP and systems automation, world models, the infrastructure that makes physical design programmable or systems that supercharge the process steps in the AEC value chain, we’d love to hear from you.

You can find us at shamillah@dawncapital.com, alexandra@dawncapital.com, and atiyu@dawncapital.com.