Audio article narrated by OpenAI
The AEC industry will pilot almost anything. The problem is what happens after the pilot. 80 per cent of the challenge, Andriy Mulyar argues, is not building a product that works; it is teaching the buyer what they just bought. That ratio, and the strategic choice it forces, runs through everything Nomic is doing: why it stays a platform rather than becoming another point solution, why drawing review is an entry door rather than the destination, and why the industry’s instinct to slow down has stopped being protective. Andriy came to this market not through years of proximity to construction but through deliberate elimination: a two-axis investigation into where agents currently fail and which large industry is worst served by that failure. The answer pointed at the built environment. He followed it.
About Andriy Mulyar
Andriy Mulyar is the founder and CEO of Nomic AI, a New York-based AI agent platform for the built environment. He has been building in AI/ML for nearly a decade, starting in NLP in 2016–17, before transformers existed, when state of the art still meant throwing LSTMs and CNNs at character-tokenised text — and was an early engineer at Rad AI, training billion-parameter language models on 100B-token datasets for radiology a year before GPT-3 shipped. He left an NYU PhD program in 2021 to start Nomic, raised a Series A from Coatue, and shipped GPT4All, a proof-of-concept infrastructure capability that reached a quarter-million monthly users, before sharpening the company’s focus on AEC in late 2024. Born in Ukraine and raised in Virginia an hour south of Washington DC, Andriy was his family’s first generation in the US. His father, a small-business contractor, was the family’s entry into the American economy through construction, and Andriy still ends his Nomic days helping his dad review contracts on a notepad and a phone app.
—
“We didn’t pivot. We sharpened our focus. We stopped trying to boil the entire ocean and found one sea and focused on it like crazy.” Andriy Mulyar
—
There’s a tidy version of the Nomic story that reads as inevitability: a horizontal AI infrastructure company raises a Series A, surveys the landscape, and identifies AEC as a sleeping market just waiting to be picked up. The version Andriy actually describes is messier and more interesting. It’s a story about an outsider deciding to act on something an industry of insiders had not yet articulated, then building the muscle to operate inside it. The decision wasn’t a pivot in the usual VC sense; it was a sharpening: a deliberate narrowing of the company’s surface area until one set of problems could be solved in depth.
The two-axis investigation
Nomic was founded in 2022, nine months before ChatGPT existed. The founding team had built billion-parameter language models in radiology at Rad AI on 100B-token datasets in 2020, before GPT-3. The original product was tooling for ML and infrastructure teams to make use of pre-API foundation models, back when you had to spin up your own GPU stack. One viral artifact, GPT4All, hit a quarter million monthly active users and generated zero revenue.
A little over a year ago, Andriy ran two questions in parallel. Where do agentic systems currently fail to deliver value? Long-horizon multimodal tasks. What large industries are stuck because of that failure? The built environment. The only industry with lower AI adoption than construction at the time, in his reading, was food services.
“Every single piece of data that flows between organisations references a 400-sheet multimodal document that language models basically just fall flat on when they touch.”
The choice that followed was the part most founders avoid making cleanly: not to widen the aperture but to close it. Hire architects and engineers. Iterate deeply before any customer touches the product. Build for the realities of 40,000–50,000-person organisations. This is the same posture Will Meurer described as building complete floors rather than slivers. The difference from most founders who fall in love with an industry through years of proximity is that Andriy arrived at it through deliberate elimination.
The outsider’s path: candour as go-to-market
The objection to a non-industry founder taking on engineering is obvious, and Andriy says it himself before anyone else does. When a random guy with no engineering background shows up to do AI for engineering, engineers will show him the door. So he did the opposite of what most outsider founders do. He booked tickets to a Texas engineering conference, cold-emailed every attendee, and stacked three days of meetings with leadership of engineering firms. He told them exactly what Nomic was, exactly where they were, and exactly what they needed from design partners. Many of those firms are now customers.
“Engineering is an industry of trust. When you have candour, people trust you. That’s what I learned early on.”
The second move was equally deliberate. He hired a structural engineer to sit with him every single day, on calls, in the office, translating vocabulary and use cases and the actual texture of how a point solution gets evaluated.
This is the same outsider posture Raymond Zhao described at Structured AI, where he, Brandon and Issy sat at a vendor roundtable as vendors and asked the room to tell them what not to do. The two stories are siblings rather than copies. Raymond’s company earned its credibility through six years of part-time contracting by his co-founder; Andriy’s was earned by compressing that proximity into a deliberate operational structure: direct contact with leadership, an embedded domain expert, candour as default. The structural condition is the same: the people approving purchases haven’t used the technology, and the only way through is to make verification cheap.
Platform, not point solution
Andriy used estimation as the counterexample. Pre-construction teams hire mountains of estimators for quantity takeoffs, and there is a mountain of estimation software, old and new. At an industry conference last week, all the estimation vendors were lined up in the same row: the old guard next to twenty-two-year-olds who had just raised $15M, all smiling at each other across the aisle. His critique was direct.
“I just think that’s fundamentally the wrong way to think about delivering value in a world where you can meter intelligence behind any API.”
His framework distinguishes two classes of software. There is software that maintains a system of record (Autodesk, Bentley, Aconex). And there is software that does the work for you: it takes action and moves data from one state to another, where a human used to sit in the middle. The first class isn’t going anywhere. The second class is about to enter distribution for a foundation model in six to twelve months.
The chokepoint is structural: when a tool’s performance determines the overall performance of the final product, the tool provider gains structural power over the solution provider, and as the tool improves through learning, the tool provider moves directly into the solution provider’s business. Most AEC point solutions are about to be re-bundled by the agent layer underneath them.
Nomic’s response is to cleanly separate the two layers. The agent orchestration capability (kick off an agent, track costs, talk to every system and integration, and keep jurisdictional data governance straight) is independent of what the agent is actually doing. Drawing review, submittal review, drawing revision comparison, RFI rule-set generation are all the same agent with different prompts, sitting on top of the same orchestration infrastructure. Internally, the team shipped an initial high-quality drawing review within three to four weeks because the abstractions were already in place.
“We’re selling both. When you buy the tool, the tool costs you the same as buying the entire platform. And then you realise the day after that there are 20,000 other tools available, and you can build your own.”
Sequoia’s “Services: The New Software” essay frames this as the copilot-to-autopilot shift: the platform sells through the tool because the tool is the artefact a buyer can hold; the platform is what makes the next twenty tools shippable in a quarter.
Drawing review as the entry door
I asked about the graph representation Nomic generates from drawings. Andriy was straight to the point. The graph isn’t a novel concept; a BIM model exposes a graph through the Autodesk API. The real point is that Nomic’s ML models reconstruct that graph from the drawings themselves, without needing the structured BIM model upstream. It’s an approximation, but it’s good enough for many useful tasks.
What matters more is the ownership posture around it. Every markdown file generated, every markup made, every output that comes back through the API belongs to the customer, not to Nomic.
This inverts the common dynamic in AEC vertical SaaS, where the vendor accumulates the customer’s domain signal and the customer rents access to a derivative product. Nomic is making the explicit bet that the orchestration layer, not the firm’s data, is the durable asset. Whether that bet holds depends on whether foundation models continue to commoditise from underneath, but it is at least a clean architectural answer to the question of where value should sit.
The deeper observation is where innovation is happening. Across the industry, the firms moving fastest aren’t always the largest. Some of the most technically ambitious work Nomic sees is coming from 30-, 40-, 50-person architecture and engineering firms ripping together their own integration stacks, writing lightweight internal automation, and asking Nomic specifically to make their drawings parseable by systems they built themselves.
The reality of the industry creates uneven conditions for speed. Larger firms have more layers of governance, carry more integration complexity, and more systems to coordinate before a new system can be deployed at scale. Large firms are ambitiously working across multiple geographies, platforms, and compliance environments simultaneously, meaning firms with fewer dependencies can move from an idea to a working prototype faster. This speed translates into a competitive advantage.
A localhost MVP built with Claude Code does not operationalise at scale. The orchestration layer is what keeps the system from breaking as it scales beyond a single laptop.
The 80 per cent that isn’t building
“20% of the challenge is building a working product that knocks the socks out and delivers massive amounts of value. 80% of the challenge is the change management, the education.”
What an agent is. How to use it correctly. How to verify the output. How to scale agent usage across a firm without ending up in cost overruns or silly behaviours. These are the educational problems Nomic solves alongside its technical product, and they are why Andriy has spent the last two and a half weeks personally training the people who train the customers. The team is hiring more, not fewer, engineers because of coding agents. The cotton-loom dynamic runs the same way: the loom didn’t reduce demand for clothes; it expanded the market for them. Agents don’t reduce demand for built assets; they expand what a smaller team can deliver.
This dynamic surfaces consistently across AEC product conversations. Onur Ekinci at CalcTree named the two-pronged adoption motion: top-down decision-makers and bottom-up engineers, both required, both fragile. Andriy’s diagnosis adds a third layer to that picture. He runs an implementation team because some customers cannot reach value without one. He also argues, harder, that the limiting factor is not the implementation team but the people above it.
“You can tell who has gotten their hands on the keys and tried some of the tech in the last six months, and who is just repeating what they heard at some conference by some thought leader who doesn’t go past Outlook on their computer.”
The argument is that the industry runs on received opinion at the top: decision-makers calling the shots on five-year technology bets are listening to vendor marketing and conference panels rather than sitting down for 40 minutes with the actual problem. It has never been easier to go from a clearly defined problem to a working prototype. The firms that don’t try are making a strategic choice, whether they call it that or not.
I’ve written about this dynamic as one of the four hidden barriers to innovation in AEC: boards that empower innovation teams to work within the current delivery system but never on the system itself. Andriy is naming the upstream version of the same dynamic.
The cost story is upside down
The recurring fear that surfaces in the conversation is cost. I fed it five words and got billed $40. It’s a legitimate concern: Give agents access to critical decisions and budgets run away fast.
Andriy’s framing on this is direct. Cost only matters relative to the value of a human doing the same task. Nomic spends $1,500 per engineer per month on coding agents; substantially more than the ~$200 per engineer their fastest-adopting AE customers currently spend on AI usage with Nomic. The value-to-cost ratio favours the customer, dramatically.
The bigger structural argument runs further. In five years, consultancies will need to adopt outcome-based pricing to keep growing at current rates. Agents collapse wall time; tasks that took humans days now take 30 minutes, plus 30 minutes for verification. End customers will know this and will squeeze consultancies for it. The operating model has to change.
“Right now, we’re at the beginning of the curve. The technology will only get faster, more capable, and more cost-effective from here.”
This is the same shift I’ve been sketching in my own thinking on the move from time-and-materials to outcome: per-seat collapses the moment headcount isn’t the constraint anymore, and time-based billing collapses the moment two firms doing the same work at different speeds are both pricing the hour. The conversation with clients about value versus time hasn’t kept pace with what technology now makes possible. The firms with the most to lose from outcome-based pricing are also the firms with the most power to delay it, which is why the transition will be uneven and politically messy rather than clean.
The industry that doesn’t internalise the pace of change will be hit harder. Not because the technology is hostile. Because the gap between what’s available and what gets approved is widening every quarter, and at some point the delta becomes structural rather than procedural.
AEC-Bench: disruption on purpose
Open-sourcing the benchmark was a deliberate move on two axes. The first is reputational. Being spiky, weird and non-standard is how a company gets noticed in a slow industry. It caught Andriy’s attention while he was mapping the market; releasing AEC-Bench is the same outward-facing move.
The second is structural. AEC AI is about to attract enormous capital and building effort, most of it uncoordinated. Without a shared benchmark, there is no common standard for what “better” means; without that, the industry will build the same capabilities twenty times in parallel and measure nothing. The version Nomic released is a month and a half old; the company has internal benchmarks ten times larger, currently focused on measurements as a domain. Before release, one of Andriy’s product people argued they shouldn’t ship it because OpenAI would pay someone to generate a 100x version and bake it into a training set, eroding Nomic’s parsing edge. Andriy disagreed.
The willingness to give away something that isn’t critical IP, something that also doubles as a marketing budget, reflects a logic that is now spreading beyond startups. Google and OpenAI have both recently moved toward deploying professional services arms alongside their model offerings; the underlying bet is the same one Andriy is making. The durable advantage lies in the customer relationship and the platform beneath it, not in the artefact that a competitor can replicate. For Nomic, that durable layer is the orchestration platform and the domain expertise it earns through deployment; not the benchmark, and not the parsing models that score against it.
The Palantir reality and the customer where they are
Some sub-verticals of the built environment will only buy professional service contracts. $2M total, $100–200K platform fee, $1.8M of which is five engineers per year doing implementation. Palantir is, as Andriy puts it, the repurposed management consultant for the world of technology. It works because forward-deployed engineers do configuration work, not vibe-coding. An ERP is an ERP is an ERP with slight modifications depending on which professional services industry you’re selling to.
Nomic does some of this because some customers won’t buy any other way, and without it Nomic can’t reach its mission. The rule he comes back to is operational, not philosophical.
“You need to meet the customer where they are at. When you build an agent for a customer, the agent starts where the human starts work and ends where the human ends work. If it doesn’t, they won’t use it every day. They won’t accept it.”
The workflow boundary is the adoption boundary: if the agent doesn’t start where the human starts and end where the human ends, it doesn’t get used every day. This is the same lesson Joe Patrois described from the defence-into-construction direction: outsider rigour only translates into adoption when it meets the operator at the actual boundary of their day. The forward-deployed posture is not nostalgia for management consulting. It is the recognition that some customers cannot reach the product’s value without a human carrying the configuration weight, and that refusing to do that work is the same as refusing the customer.
The champion-in-the-corner dynamic is real across every AEC product conversation, but louder champions are not what’s missing. What’s missing is leadership that has actually used the technology in the last six months. Until that gap closes, every champion is just a translator for people who don’t read the original.
The structural drag runs deeper than organisational culture. Every US government project’s RFP response is owned by the government, which currently prevents Nomic and others from training and operating on that data. The engineers and policymakers at the peak of innovation in government understand this; the rule-execution machine in the middle does not. Worse: project owners who said no AI five years ago lock entire assets out of AI for the next four years.
“The technology works, it would make us money, it would save us money — but the contract said no AI five years ago and there are five other things happening. That whole thing just feels so upside down to me. And I see that every day.”
The asset gets no AI. Not because the technology fails. Not because the GC doesn’t want it. Because a clause written before this generation of models existed is sitting in a contract that nobody wants to renegotiate.
When unbundling becomes the operating model
Andriy still ends his Nomic days helping his father review contracts on a notepad and a phone app. That detail is easy to read as colour, but it contains the whole argument in compressed form. The person who wrote the contract, the person living inside it, and the technology that could change what’s possible between them are all present in the same moment. The gap between what could be done and what the contract permits is the same gap the industry faces at every level: from the government clause written before this generation of models existed to the leadership team that hasn’t spent 40 minutes with the technology.
The story Andriy is telling is not about a vertical bet. It is about a structural moment in which the tool layer is consolidating value away from the solution layer, and the industry’s reflex to slow down is no longer protective. It is exposure. The agents are already pricing faster, reviewing faster, and learning faster than the procurement cycles that govern them. The firms that internalise the pace of change get to participate in rebundling: they decide which workflows are automated, which data stays theirs, and what the next bundle looks like. The firms that wait discover those decisions have already been made for them. This is the closed loop: wait, and the choice is made for you. Act, and you choose the terms of your own disruption.