Arc 2: Agentic AI in AEC | The Next AEC Tech Stack
Arc 2 of The Next AEC Tech Stack: Agentic AI in AEC. In a real firm, today, what can it actually do, who signs off when it's wrong, and what it costs you in who you still have to hire.
This series brought to you by
Graitec provides AEC software, training, and expert support to help architects, engineers, and construction teams design, build, and manage projects more effectively.
Arc 2: Agentic AI in AEC
"AI" has stopped meaning anything useful in a vendor's marketing, so this arc asks the question straight, to the people building it and the people using it: in a real firm, today, what can it actually do, and what can't it? Who signs off when the software says a design passes. What still needs a human to verify. And what it costs you, not in licensing dollars, but in who you still have to hire.
Three episodes, working through it with the people building these tools and the people using them every day.
With support from
AVAIL's content management system (CMS) helps architecture and engineering teams find, manage, and reuse their Revit, AutoCAD, Civil 3D, and other technical content in one place.
Episodes in this Arc
Coming soon
4: From Autocomplete to Agent: Where AEC Actually Is
Coming soon
5: Who Signs When the AI Is Wrong?
Coming soon
6: The AI Easy Button Isn't Free Anymore
Featured voices in this arc
Theodoros Galanos
Generative AI Leader
Aurecon
Oliver Goldman
VP, Software Architecture, AEC Solutions
Autodesk
Randall Stevens
CEO & Founder
AVAIL
Greg Demchak
VP, Emerging Technologies
Bentley
Ariel Cooke-Zamora
Product Operations Lead
Bimbeats
Matt Wash
CEO
Bimbeats
Campbell Yule
AEC Tech Enthusiast
campbellyule.com
Roderick Bates
Senior Director, Product Operations
Chaos
Mike Soldan
SVP, Global Technology Solutions
Egnyte
Jesper Wallgren
CPO & Co-Founder
Finch
Holger Kreienbrink
Director, Product Intelligence & Partnerships
Graphisoft
Marty Rozmanith
CTO & Co-Founder
Skema
Gavin Nicholls
BIM Educator
The Aussie BIM Guru
Prasanta Bose
CEO & CTO
TwinMaster
Michael Jansen
Chief Business Officer
TwinMaster
Scott Reynolds
CEO & Co-Founder
UpCodes
In their words
Moments worth reading twice, pulled straight from the tape.
“What you're seeing in the rollout is less about the order in which the ideas occurred to us and more about the depth and the complexity of the work that needs to be done to bring those experiences to market.”
Oliver Goldman
Autodesk
“Ultimately, I'm responsible for the products that I deliver. I'm accountable for those. And it doesn't matter what tools I use or how good they are or they aren't, I still carry that accountability.”
Oliver Goldman
Autodesk
“Building Layout Explorer is a generative AI model. It actually uses a large language model under the covers. The user interface doesn't have you chatting with it.”
Oliver Goldman
Autodesk
“Part of the MCP server vision is not that you'll ignore Autodesk Assistant, but that you'll use Autodesk Assistant in combination with some of your own agents — maybe an agent trained on your data or your ways of working.”
Oliver Goldman
Autodesk
“You could literally spend all your day just watching the stream of announcements and trying to download the latest models and never get anything else done. But at the other extreme, you ignore it entirely and the world's gonna move on and you're gonna miss out.”
Oliver Goldman
Autodesk
“We've been using AI and machine learning for the better part of five, six years for use cases like image classification... to find things like defects, exposed rebar, rust, cracks in the road. This is something we had in the market before the whole LLM explosion took place.”
Greg Demchak
Bentley
“There's no agent that's actually signing off on the structural analytics or the final documentation. What the agent does is basically radically accelerate your time to that deliverable. It's still the human in the loop to check and verify all that.”
Greg Demchak
Bentley
“Do you want to put your work on autopilot and be able to totally trust it, or do you want to be involved in it? And if you want to be involved in it, then you've got to be involved in it.”
Greg Demchak
Bentley
“You do try to remove some bias in these systems by creating multiple agents, multiple personalities that will cross-check, test, criticize, try and tear down your argument — you create a counterpoint, so you don't get into this sycophant type loop.”
Greg Demchak
Bentley
“Show me this complete building model, isolate all the concrete, and colour code the concrete based on the quantity of embodied carbon... That is something you can do today.”
Greg Demchak
Bentley
“The model context protocol is really just a way for AI to understand the APIs that already exist in the software. These APIs have a long history of being used and they're quite trustworthy.”
Greg Demchak
Bentley
“The more information you embed into the geometry itself, the better results you're going to get... So it's not making as many assumptions.”
Greg Demchak
Bentley
“It's kind of like a real intern — you're letting an intern come into your office and mess with your BIM model... there is a period of time over which you build trust in the system.”
Greg Demchak
Bentley
“It's easy to draw a parallel to visual scripting in Grasshopper, where each step is a node. But now some of these nodes are essentially magic — I input something, some magic happens, out comes the output.”
Ariel Cooke-Zamora
Bimbeats
“At the end of the day it's going to be someone's professional stamp on the piece of paper. It's going to be the person with the stamp that's held liable.”
Ariel Cooke-Zamora
Bimbeats
“That's caused an awareness that the tech is moving faster than the laws we have in place. My answer was about a stamp on a piece of paper — that's an ancient methodology for validating who's in charge.”
Ariel Cooke-Zamora
Bimbeats
“If you'd asked me a few years ago I was all in on honing a niche. I think things are moving so quickly now that the opposite might be true — you need people completely comfortable throwing out the way they've done things for the last five years.”
Ariel Cooke-Zamora
Bimbeats
“There is no way that the AI was ever going to tell me what it felt like to be in that space.”
Matt Wash
Bimbeats
“If you're training it on the knowledge of your people, that's where it changes — you'll still employ the same people, but how they get experience and knowledge will be different.”
Matt Wash
Bimbeats
“Right now AI is the drug, and everybody's being given this drug and lapping it up. At some point there's obviously going to be a cost to that.”
Matt Wash
Bimbeats
“Before, they knew it was a hacky Python script, so they'd use it with caution. Now the application looks professionally coded, like it took hundreds of hours. That's the bit I think is tricky.”
Matt Wash
Bimbeats
“A lot of projects are going to pencil out that didn't pencil out in the past, and I'm hoping for a boom in construction because of that.”
Ariel Cooke-Zamora
Bimbeats
“Just building a building, I don't care whether you're licensed or not, it's still hard to do. That is not an easy thing to replicate.”
Roderick Bates
Chaos
“There's no doubt that there's going to be people losing their jobs because of AI, full stop. AI doesn't care. It may seem very human, but it does not care.”
Roderick Bates
Chaos
“They know that when we say we're not going to train on their data, we're not going to train on their data. Whereas a smaller-scale startup — they might just find out that, oh yeah, actually we were training on the data. I'm sorry. And, you know, sue us. We're not worth anything.”
Roderick Bates
Chaos
“We saw one of our competitors basically offering free AI visualization. We're like, wow, how are they doing this? Well, it turns out they were just burning investor money.”
Roderick Bates
Chaos
“At one point people complained that everyone who used Enscape had a similar-looking output. Now they're telling me they really like the way Enscape treated light and materials. Eventually you stay around long enough and people appreciate the aspects that were once a liability.”
Roderick Bates
Chaos
“The AI easy button, back when everything was free, is no longer there. Before, you just couldn't model that much, so you didn't render that much. Now you just don't want to pay that much, so you don't want to render that much.”
Roderick Bates
Chaos
“There's a good chance learning will not be as much of a part of the production process in architecture going forward.”
Roderick Bates
Chaos
“The number of people really willing to put in the effort to do these types of things has always been small. The people that could do a great hand sketch at a firm have always been in the minority.”
Roderick Bates
Chaos
“They're more willing to get feedback from AI than they are from people. If AI tells you it's ugly, you're like — oh, okay. Whereas if someone told you that, it really would hurt.”
Roderick Bates
Chaos
“Instead of just bolting AI onto a process, they're looking at it and saying: how do I rearchitect how we look at that?”
Mike Soldan
Egnyte
“AEC content is hyper complex. It's got nested tables, stamps, handwriting, versioning — all things that are very, very painful for a modern LLM.”
Mike Soldan
Egnyte
“They deploy Claude across the entire firm and all of a sudden they're getting hundred, hundred-fifty thousand dollar token bills in a month, going: wow, this math is starting to not math.”
Mike Soldan
Egnyte
“It's not even really hallucination anymore. The bigger risk is that my reference content is five years old instead of two years old, not that it fabricated something.”
Mike Soldan
Egnyte
“It's a low-margin business, so every mistake, every little thing is just eating another point away at your four percent. There's a bit of the devil you know: I'm going to do four percent again.”
Mike Soldan
Egnyte
“AI is great at a lot of things, especially going from zero to one, but it's not the hammer you should use on everything, everywhere.”
Jesper Wallgren
Finch
“I heard one firm that had written a document that checked a thousand things in the Revit model. And there were so many things going wrong that you just stop listening.”
Jesper Wallgren
Finch
“I think for sure it will be possible to prompt a building in the future. The more interesting question is: do we want to prompt a building?”
Jesper Wallgren
Finch
“I don't think you should design anything that should last for 100 years in 100 seconds.”
Jesper Wallgren
Finch
“At the end of the day, you are the architect, it's your building. Finch is just a tool. It's a great tool, but at the end of the day, it's your building.”
Jesper Wallgren
Finch
“Where we want to see AI is on the tedious tasks — placing dimension lines, cleaning up plan files, placing labels, doing the checks. This is what no one wants to do.”
Holger Kreienbrink
Graphisoft
“I see AI as a very well-equipped and educated draftsperson — I tell this draftsperson I want the design like this, and then it will do it. That's the North Star.”
Holger Kreienbrink
Graphisoft
“Token spending will be huge if we are not cautious. It becomes a second licence cost, and it can be even more expensive than the licence cost if you overuse it.”
Holger Kreienbrink
Graphisoft
“Everyone is expecting that AI will reduce costs. Maybe it will not. It will make us more efficient, but still we need to pay a price for that efficiency.”
Holger Kreienbrink
Graphisoft
“Never have I seen so much progress in the whole BIM industry as in the last two years. For the last twenty years everything was just evolution — and now there is a lot of disruption.”
Holger Kreienbrink
Graphisoft
“You don't have to be the first. You can be the second or the third — but then do it properly.”
Holger Kreienbrink
Graphisoft
“There will be a lot of individual solutions solving tiny little problems, and they cannot live on their own. They need a connecting data layer.”
Holger Kreienbrink
Graphisoft
“Deterministic things — scripts, Grasshopper, Revit journal files, Dynamo — you put in the same inputs, you get the same outputs every time. Versus AI, which is probabilistic.”
Marty Rozmanith
Skema
“The AI itself writes a script, which is deterministic, and then runs it — because if you ask it to do it again, it will produce the same result. It's the combination of using AI to interpret your goal, driving deterministic things, that yields the greatest results.”
Marty Rozmanith
Skema
“When technology gets in the way and imposes a tax, we should eliminate it and reduce that friction so the tax isn't there anymore. I call that democratizing technology.”
Marty Rozmanith
Skema
“You actually need to know a lot more about putting a building together in order to draw a building in Revit than you had to in AutoCAD. Back in the day, it was a raising of the bar. So if there's a lowering of the bar now, maybe it's about time.”
Marty Rozmanith
Skema
“What we don't want is an AI to be a black box, where you feed it twenty years of firm history and it starts spitting out designs and you don't know why.”
Marty Rozmanith
Skema
“A human's going to get annoyed at having to remember which of the five hundred inputs it is. An LLM doesn't.”
Marty Rozmanith
Skema
“In copilot, the human in the loop is ever-present. In agentic, the human only reviews the end thing when the AI says I've completed the goal. That doesn't absolve you of responsibility for it.”
Marty Rozmanith
Skema
“What the agentic stuff opens up is for you to make custom tooling that in the past would have been way too expensive or unapproachable — that for your firm's workflow might be very, very useful.”
Marty Rozmanith
Skema
“Revit is — and people might bristle when I say this — a data entry tool. People sit in front of it for eight hours and put in walls, doors, windows and staircases.”
Marty Rozmanith
Skema
“When production's free — just like renderings became free — it forces the question of what being an architect even is, because I think we lost sight of it.”
Marty Rozmanith
Skema
“To get to a certain standard of care required a certain number of hours. I think that equation is now going to be fundamentally broken.”
Marty Rozmanith
Skema
“Autodesk has to have an AI story. They're a publicly traded software company, so they had to do something. Whether that something is useful remains to be seen.”
Marty Rozmanith
Skema
“Building APIs on systems that are file-based and architected thirty years ago probably isn't a good strategy.”
Marty Rozmanith
Skema
“A lot of things you see in the market today are single-purpose — AI for drawing, AI for energy consumption. Arch-e works across the entire lifecycle, across disciplines, across KPIs, and across software platforms.”
Michael Jansen
TwinMaster
“Ninety-five percent of the tools calling themselves AI tools are actually LLM wrappers — a bit of code wrapped around an existing LLM. Otherwise we're just going to go from thirty algorithmic tools to thirty AI tools, and it's no better.”
Michael Jansen
TwinMaster
“You must be a good writer, a good speaker — because that's what's going to help you make a good design, a good system.”
Prasanta Bose
TwinMaster
“You have to get out of this thinking of living and dying by the fee. A fee is time-bound — job's over, your fee is gone, next job. That's the business model.”
Michael Jansen
TwinMaster
“The art of the impossible becomes now the art of the possible.”
Prasanta Bose
TwinMaster
“When you talk to your lawyer or your doctor, you divulge a tremendous amount of private information. Thanks to privilege and the Hippocratic Oath, they're not supposed to divulge it. Same thing with your AI service provider if you're in AEC.”
Michael Jansen
TwinMaster
“I don't fault AEC professionals for their risk aversion. I understand it. It's really the job of the vendor to prove it.”
Michael Jansen
TwinMaster
“I would not rely on the incumbents to be spearheading all the innovation in the industry. I think that's foolish.”
Michael Jansen
TwinMaster
“We're gonna queue it up for you, and we're gonna show those mistakes, and we're just gonna let you decision on those.”
Scott Reynolds
UpCodes
“We process 60,000 changes per month across the United States, across all our jurisdictions.”
Scott Reynolds
UpCodes
“The juniors don't know what they don't know. The seniors don't know what they do know.”
Scott Reynolds
UpCodes
Arc 2 partners
A series like this doesn't happen without companies willing to sit down and answer hard questions on the record, without approval over how the answers get used. Special thanks to the partners below for backing Arc 2 and for showing up to a conversation they didn't control. Please visit and support them.
Autodesk makes software that helps people design, build, and operate the buildings, infrastructure, products, and experiences that shape our world.
Bentley Systems develops infrastructure engineering software that infrastructure professionals use to design, build, and operate better, more resilient infrastructure.
Bimbeats is the tool that shows you how the rest of your tech stack is working.
Chaos is a world leader in design and visualization technology, providing an ecosystem of accessible tools that support every stage of the design and creation process.
Egnyte's AI-powered platform helps organizations unlock the full potential of their content, efficiently, securely, and responsibly.
Finch is an AI-native building design platform for architects.
Graphisoft empowers teams to design great buildings through AI-assisted software solutions, learning programs, and professional services for the AEC industry.
Skema turns your hard-won design knowledge into a reusable asset, so every new project builds on what you've already proven.
Arch-e, from TwinMaster, is the AI decision assistant for architecture, engineering, and construction, helping teams anticipate consequences and optimize decisions across the building lifecycle.
UpCodes is the AI-native QA/QC platform that helps architects, engineers, and contractors research, coordinate, and verify building code requirements, reducing risk and keeping projects on track.