The architecture, engineering, and construction (AEC) industry is plagued by two major problems: project delays and cost overruns. Of major projects, only 8.5% finish on time and on budget. Whether building a bridge or renovating a bathroom, talk to anyone and they have first-hand experience with these problems. It is so commonplace that it is presumed to be structural.
AEC is considered one market, but it’s actually a collection of disciplines, operating with different rules and different tech stacks. Overall, the software they use is big, clunky, and deeply entrenched. Their market dominance is based on their steep learning curves, and designers and engineers are expected to be fluent on day one. Incumbent software vendors have failed to deliver basic levels of integration and collaboration, with interoperable and disparate software systems unable to manage the fragmented workflows. This makes it difficult to do the most basic parts of project management, like easily finding relevant information and tracking incremental changes across working groups.
There are two lanes of innovation we’ll see as AI comes for AEC. The first is that LLMs will allow a new wave of vendors to run at the collaboration and interoperability problems costing time and money. AI has the opportunity to succeed where others have failed because it can operate as a layer on top of existing solutions, so no individual discipline would be required to change their tech stack in order for it to be effective.
The second is discipline-specific and purely additive: using generative AI to replace human labor supporting design, engineering, etc. The problems and tasks in AEC are a nice balance of constrictive and creative, with a ton of legacy data from previous projects, making the subdisciplines a good candidate for AI support.
The industry today
Any project starts with an owner, who may bring in financiers or business partners, early feasibility design vendors, as well as concepts to pre-plan before any real action is started. From there architects, engineers, and designers are brought in to create the plans and collaborate with government and permitting. A general contractor will take the design and bring in sub-contractors who themselves may bring in more sub-contractors, plus procurement and suppliers. While the process is described as linear, it is anything but, at each step reverting to stakeholders in earlier stages, with multiple interdependent project schedules running concurrently, and the end project being touched by dozens if not thousands of people depending on the size and scope. One result of this madness is that crisp budgets may not exist until years into any project when a general contractor gets firm bids. A side effect of this is that even before projects are started (pre-construction), projects are already delayed because stakeholders don’t have the right data.
In many other industries technology has brought together this fragmentation and delivered significant automation. But not in AEC, where each stakeholder may have their own technology stack. To highlight how software vendors are trying to handle this utter madness, Autodesk has 112 products in their current catalog. Some products like Oracle, Procore, Revit, and AutoCAD have reached a certain level of ubiquity, but nevertheless integrations and interoperability between systems is incredibly low. For example, permitting is generally done early but also ongoing throughout the process, but the system for managing permitting is disparate from the design system (e.g. CAD). The net result is that later in the project, design decisions are not understood and a contractor has to circle-back to an architect or other stakeholder to understand compliance, or confirm an unusual design element.
Catching errors early through collaboration & AI
One of the major drivers of the two big problems in AEC is that you can’t fix problems you don’t know about, and, without the ability to collaborate with other stakeholders, you can’t know about problems until you’re on the job site. The traditional ways of working require each team (mechanical, electrical, etc.) to work off a core plan using their disparate tech stacks, then bring them back together to diagnose problems and implement the plans.
There are (at least) two ways to solve for this. The first is a collaboration layer that improves transparency and closes the gap between plans and implementation. We’re seeing this done well by Scale portfolio company Dusty. Dusty is known for its cute and efficient layout robot, but customers rave about the value beyond the layout printing. By encouraging construction layout to be routed through a single vendor (Dusty), versus the individual trades, Dusty has created a collaboration layer where mechanical, electrical, plumbing, and dry wall trades all bring their plans back together to identify conflicts before any work is done.
The second is automating design review. Construction drawings can be thousands of pages and errors are a natural occurrence. LLMs are increasingly capable of understanding construction drawings and using that understanding to provide the professional a second set of eyes. Firmus is a startup that is tackling this problem, focused on being the pre-construction “spell & grammar check.”
Generative AI can create in AEC
Generative AI can provide an intuitive interface in contrast to traditional AEC systems like AutoCAD that are incredibly complex, requiring years of training to the point where they are taught in school. Therefore, experts are required in each organization to create and interpret data in these systems.
Generative interfaces bring these barriers down. Not only do they allow anyone to query the data in a natural human interface (just ask a question!) but we are also starting to see natural human language being used to create content in these systems, or for these systems. The result has the potential to become an incredible disruption in skilled human labor. Motif and Arcol are bringing the ease of concepting software with the depth of production software, all powered on the back of LLMs. Likewise, Augmenta takes simple inputs and creates a full 3D electrical schematic in hours that would have previously taken days if not weeks or months to produce. These are incredible strides in technology.
On the professional side, design teams can use generative AI to gain the efficiencies we’re seeing across markets from both generalist and purpose-built tools: research, drafting, brainstorming, pressure testing, etc. This could even extend to the consumer. A simple bathroom remodel requires an architect or designer to draft your plans. Generative AI tools are getting to a place where they can get prompted with basic dimensions accompanied by desired end goal descriptions in words or inspirational images. The systems can then return plans for you that would have taken an expert years of training to produce. It allows laymen to encroach on the work of the professional, and further, inquire about work from professionals by uploading plans and inquiring about those plans via a chat interface. Or the human or AI generated plans could get a permitting or safety review initiated by a non-expert. While this example is a residential example, it is no different than ChatGPT is changing the workflow of the non-professional writer and the professional writer, and further blurring those lines.
LLMs can also interpret
We know that LLMs are particularly good at interpreting large volumes of unstructured data, and reverting that into digestible, mostly accurate, but not perfectly precise data. This sounds a lot like estimating, which is plagued with data overload and human limitations on what can be processed and updated in a timely fashion, not to mention good old fashioned optimism (none of which machines are constrained by). I think companies like Ediphi and Xbuild will completely revolutionize this process as the gap between what actual cost is and estimate will tighten significantly as more and more customers incorporate these technologies.
The other part of that is efficient take-off solutions, which is a massive category where many customers complain about reliability. Not only are the estimating vendors above incorporating their own take-off, but Withrebar and Drawer.ai are focused on building their businesses focused on this task. The more accurately we can understand what the drawings say, the more accurately we predict the costs (and ultimately create agentic purchasing).
Can LLMs bring it all together?
APIs are the modern way to solve interconnectivity, but APIs fail when there is a long tail of new and old vendors because the volume and maintenance of these connections is not economically feasible. The outcome is that it’s hard to find relevant, accurate, and current information about a project. Our strong belief is that companies building on top of foundation models can bring together these “locked” datasets. Much like humans, they can train on large volumes of historical outputs, interpret these, and are not brittle as anything changes on the output or input of data. And, they would not require individual teams or stakeholders to make any changes to their techstack, making it the easiest solution on this list for contractors to implement.
Can you imagine asking anything about your project, and getting an immediate answer? TrunkTools and Gryps are making that a reality today, and the results are astounding. Pelles is focused on just sub-contractors, which is a juxtaposition against the current model of contractor-as-customer, as we will discover whether any company can span across these constituencies (our gut says yes).
The real change coming
Each of these pieces is incredibly exciting in themselves, but there is a much bigger vision that may evolve. A few executives at the companies mentioned have alluded to this, and I ultimately want to give credit to Dustin at Ediphi for his crispness: as these technologies roll out, how will the value chain evolve? That is, today there is a specific set of players for each task, and they interact with the other players in a predictable fashion. But there is no doubt that AI will make the tasks of some of these players redundant and furthermore will allow you to move around the value chain in unpredictable ways. This is because inherent in each of these roles is risk management, but as the roles blur and risk is moved around does that result in further collapsing, or aggregation of risk to specific stakeholders with the most powerful tools. The net result may be a complete reworking of a century-old industry, and that’s okay — if it decreases the hated time and cost overruns.
