Latest / Construction Technology Podcast / Construction Technology Podcast - Episode 008 - with Guido Maciocci from AECFoundry
Transcript
- Luigi La Corte: Good morning, everyone. Welcome to construction technology podcast. I'm here with Mike from build vision, Andrew from join myself, Luigi from provision, and we have a special guest. Welcome Guido from AC foundry. Good afternoon, Guido. I know you're based in Italy. How are you? What's going on? How was the week? Guido Maciocci (AECFoundry): Hey guys, thanks for having me. Great week. Happy it's Friday. Sitting in Rome, ⁓ Italy. weather outside, so definitely going to go and enjoy some of that after we finish this call off. Luigi La Corte: I bet. Rome shut down on Easter Friday, ⁓ Good Friday? Guido Maciocci (AECFoundry): Yeah, not today, no not on Friday, but Easter Sunday and Monday. Luigi La Corte: Well, appreciate you joining regardless. So Guido, maybe you can just tee up the audience about AEC Foundry, what you do. What you do that's a little bit differently than most technology companies in this space. And then we can just dive into it. Guido Maciocci (AECFoundry): Yeah, so AAC Foundry, partner with AAC Enterprises to, I guess tagline is to turn fragmented knowledge and data into scalable and we do this through platform driven data first and AI delivery model. what makes us different in this space. ⁓ think the things that we sort of highlight ⁓ is we're not a consultancy terms of, we don't just deliver strategy and slide decks. We don't do strategy without outcomes. We build custom but we're not a dev shop. So ⁓ we push a on requirements and we try to embed ourselves within the... our clients' teams and workflows as much as possible to understand how they're actually working. So we don't build software on request and we're not a sort of an agency type. we're really long-term tech partners. The we have with clients are, ⁓ know, to 24 typically, if not longer. They keep coming back, departments. So we try to really get in the organization and try and embed ourselves as much as possible and as close as possible the domain knowledge. Luigi La Corte: I see. Foundry. Did you draw any inspiration from Palantir's Foundry? Should we think about it the same way? It's like the ontology layer that you plug in. That's kind of like the technological layer. Tell us a little more about that. Guido Maciocci (AECFoundry): Yeah, no, there is. Yeah, so no, name, Palantir came up with Foundry, I think, a long time ago, maybe 2015. 2016, somewhere around there. Could be earlier, not sure. But there's no link to the name. There's no to the name with Microsoft Foundry either. Admittedly, they picked that after us. So if there is a sequence, it would be like Palantir, then us, then Microsoft, in terms of the naming. ⁓ But yes, are. ⁓ Andrew Zukoski: Hmm. Luigi La Corte: Microsoft Red Hat would do, obviously. Andrew Zukoski: They have deep pockets if you can get like a trademark going against them. Luigi La Corte: Yeah, they'll 50k a month for that. Okay. Fantastic. Yeah, exactly. Mike, I know you had a couple of questions regarding models and where the industry is currently, where it's going. pass it over to you. Guido Maciocci (AECFoundry): All right. Yeah. I can trade market in Australia or something. Mike Powers: What? Yeah, for sure. Thanks for taking the time to chat. How are you thinking about foundational models? Are you building on top of foundational models? You building your own models? We'd love to kind of have you unpack that. know, being an book, we're spending like $15,000 a month currently through Gemini. So obviously, that's a huge API cost. We'd love to get your thoughts on like, just how you're thinking about the cost per token and what technology you're leveraging. ⁓ Guido Maciocci (AECFoundry): Yeah, so we're completely agnostic in terms of how we integrate these models within our platform and then offer those capabilities to clients. We infrastructure for clients, so they own the infrastructure. And we purposefully designed our infrastructure to be vendor agnostic so that you can sort of ahead of the curve and keep pace with the technology that's changing pretty much every day. So in terms of In terms of like model selection, it's often comes down to, with enterprise clients, it often comes down to the IT contracts they already have. So, you know, if they're with Microsoft, they'll gravitate towards open AI or whatever models are available in ⁓ the catalog within, Foundry, ⁓ they're with Google, ⁓ of course they'll make use as much use as possible of the Gemini models. And the reason for that is that enterprise clients have contracts with these companies and model providers that give them millions of dollars in free credits. So it's in their best interest to then leverage all those services they pretty much get for free. Andrew Zukoski: Can I ask a follow up about, sorry, just like how that change cycle works? Sorry to interrupt, Like often ⁓ your partners and clients like foundational model that they're working with and like who drives that? Cause like there's a ⁓ new model that might be better at something every few weeks months it feels like. And I'm just curious. Yeah. Mike Powers: So when you say like, ⁓ yeah, go ahead, Andrew. Guido Maciocci (AECFoundry): Yeah. So the reality of AI in enterprise and in companies that do it well, like if we take FinTech or large banks or tech companies, if you go and look at the AI they've developed, they've obviously been on the bandwagon from earlier anyone else, right? And they've been building internal tools and systems. based on models they had access to at the time. And you'd be surprised how many of these systems today are still in production with models that everyone thinks, ⁓ they're obsolete. like embedding models that were the first embedding models or language models that had been surpassed by six versions, old ones are still in production and they're still doing what need to do. The difference is that companies that do this properly and how AI is implemented properly within AI itself. is that you always start with benchmarking, right? And this is probably where the conversation is going to go, but you always start with benchmarking and evaluations. So if you don't have like a way to... quantitatively or at least qualitatively measure the outputs of an agentic workflow against your data and what your expected outcomes are. You have absolutely no way to tell whether a smaller, cheaper model would perform equally well or a sort of counterintuitive way, if the newest version of the latest model performs worse. And that happens all the time. Because if you set up an agentic workflow and certain prompt Luigi La Corte: Hmm. Guido Maciocci (AECFoundry): engineering for a specific version of the model. Often, the way they train the models changes, they improve, but The prompting structure, for example, changes for new families of models. And then your whole pipeline, which was amazing, unlike GPT-5.2 out, all the benchmarks say this is the best model in the world. Everybody switches. They're like, ⁓ just switch it to 5.2. Just keep switching to the best. But the performance ⁓ isn't You might be degrading performance. And without benchmarking and and particularly when doing it within an industrial domain without expert evaluations and benchmarking that are driven by subject matter experts and domain experts, there's no real way to tell when you're to switch if it's better or worse. It just intuitive. It's like some people. Andrew Zukoski: Okay, the, like IT org is responsible for so like picking the foundational models they're work with, they're leaning on their existing contracts, often there's a financial reason there. In order to guide like decisions going forward, they will have to get good at running benchmarks. on the domain system. It's not just like, all right, we're going to upgrade Postgres from 15 to 17 because read latency goes down. They have to have domain benchmarks and the skills to be able to collect the data and run them and evaluate them. And they have to do this periodically and pay attention to it if they want to keep the model. OK, that's a lot of work. Guido Maciocci (AECFoundry): Yeah, it is. It is. And I think one of the key that we talk about and one of the key challenges in this industry is in ⁓ what you said, like ⁓ the IT team and it shouldn't be the IT team. It should be a dedicated AI team or external consultants like us, of course, that can help with expertise that's just missing in these organizations. But by treating AI simply another API that we can call in a software stack, They're kind of missing out on all of the complexities and nuances that IT teams just don't understand. Andrew Zukoski: How do these AE firms quality? ⁓ Like quality of goals like I'm my background software engineering. work mostly with general contractors. ⁓ Delivering software as you know right you you tend to have like someone there tends to be someone who is responsible or there's a responsibility sometimes it gets folded into under parts of the organization to keep an eye on what's going out and make sure that we don't accidentally break anything and that things work. There's all sorts of automation you apply around but there's like a unit that does that and is assessing our work product as close to what the customer will consume like how to Guido Maciocci (AECFoundry): Thank Andrew Zukoski: How do AE firms do that? is that the place that be best suited to judge the quality ⁓ these eugenic workflows and benchmark them as they iterate and evolve? Guido Maciocci (AECFoundry): I mean, it if you're talking specifically about quality within architecture and engineering is very much focused on the review of technical drawings. And that's typically just ⁓ funneled up Typically the way it works is that you have a bunch of juniors that don't really know what they're doing and they produce a very big volume of technical documentation for projects. All of this stuff gets funneled to a senior like engineer or a senior architect who does know what they're doing. And they just spend like most of their time just going through drawings and reviewing. And the thing the industry wants the most out of AI at the moment, or at least the request that we get all the time is please automate QAQC and compliance. Andrew Zukoski: It's interesting. So it's not automate the juniors, it's automate the QAQC. Guido Maciocci (AECFoundry): because of work. Yeah, we always flip that question around. We always say, instead of trying to build AI to try and the mistakes you're making, why don't you try and embed AI upstream so that you prevent mistakes altogether, or at least you reduce the number of mistakes. It's an interesting conversation. ⁓ I think they just so much turnover in the base of the company. They hire and fire people so much based on project volume that retention of those skills and working certain types of tools, that becomes hard in these companies because they're trying to teach juniors that leave all the time. ⁓ Luigi La Corte: Greedo, if I can just jump in, what do you think is happening in the industry? I see services being bifurcated, products, services. On services, we see people being very diligent, or at least want to understand accuracy. They want to understand F1 recall precision. On products, it seems like no one really cares, and it's mostly vibes. You know, we're guilty of this as well, but every product that we've put into the market, we asked the question on these calls, what is your accuracy? we'll walk them through the F1. We're gonna be a lot more ⁓ about that in the future. Mike's done a really good job being transparent about where they are with accuracy. But why do you think there's this bifurcation and when people start to care about accuracy? Guido Maciocci (AECFoundry): I think people start to care about accuracy now we've gone through this initial hype of like AI can do everything. It can solve anything, any problem. There's no problem that can't solve. And then people try to actually use it within ⁓ domain settings. And they, you know, was a lot of disappointment, the trough of disillusionment or whatever. part of the hype cycle, right? But I think now people are getting serious and they're like, okay, if you're saying AI can do this, I needed to do that within this threshold of accuracy. Otherwise there's no ROI for me. So for example, in QAQC, if the usual, the typical question is, the typical expectation is a hundred percent, right? It's like either it a hundred percent or it's not very useful. that's like a wrong framing for a number of reasons. But anyway, question usually is if it's not doing like at 90 % of the work and I can trust that it does 90 % of the work automatically, there's no point to any AI assisted QA, QC workflow because if it's below 90, you stop trusting any of the results and then you have to QA, QC everything anyway. And that's the biggest challenge, I think, now is building user experiences and interfaces and workflows that build trust in AI, not as a compromise to its inaccuracy, but... as part of workflows. I think it's an ingredient that just has to be part of it. The layer and understanding how the AI has come to a certain conclusion, where it went wrong, why it went wrong. think that all of layers are kind of in the back end and obscured at the moment in most products. I think exposing and giving people the opportunity to drill in and audit super important. Luigi La Corte: you're on mute. Mike Powers: Sorry about that. During construction in my office. So when you think about accuracy and benchmarking and in the QAQC example, the way that I would think about it is there isn't actually a real answer in QAQC because there's no such thing as 100 % accurate document set or construction set or even what's installed in the field compared to the drawing set. How do you think through that and just educate the customer on like... Hey, we're gonna do a hundred percent quote unquote whatever that means benchmarking QA, QC. How do you define what success looks like when, I mean, there's like, there's false positives. There's, you know, there's every kind of combination of things that can happen. Just curious, like get your thought process on that. Guido Maciocci (AECFoundry): So the way that we approach this, and maybe I'll give you another example from something we're working on now, which might be helpful. But the way that we approach this ⁓ we to create, we try to first narrow down the use case. So QC, QA, QC is not workflow. It's a whole series of processes with different requirements. And you can break that down in different layers. Like what are you QA, QCing? So for example, you might start at a very basic level. A typical QA, QC process might be checking the consistency of title block information across like thousands of drawings. That's something very achievable, very easy for AI to do. Proven pipelines, high confidence. Mike Powers: But does that even move the needle? not saving, that's not helping an engineering firm make more money. That's not helping a contractor reduce cogs. ⁓ Guido Maciocci (AECFoundry): It's not, but that's like just an example of where AI shines, right? It's really good, but it doesn't have a high ROI. And then on the other end of the spectrum, you have things like compliance, which is a super hard thing to do. But it would have an immense ROI because that's where a lot of the time is spent in cross-referencing drawings against specs, against code, against standards, and so on. So there's a whole spectrum of things that you can do. ⁓ And some are like low ROI, but they save time and cumulatively you might shave off like 50 % of a QA, QC workflow on the stuff that's high confidence. then you have all this other really hard stuff that's way difficult to manage. because assessment of the outputs is much more complicated. And the way that we approach that, at least in document pipelines, is by providing the initial work that do with clients is to provide them with the extraction pipeline. the human review interface and starting with a sample set of documents and the things they want to extract from those documents. We provide them the interface, they upload the documents, and the first cycle that they go through is to basically manually inspect every output for every document through the UI. Every edit they make gets recorded, so we know the edit distance between the generated prediction of the documents of the extraction and what they actually wanted. We understand where AI goes wrong, which fields, which parts of the thing is an understanding. And then through that process, they go through the UI, they correct all of the predictions or extractions, and then we get a golden data set. That becomes the... the baseline for the first initial benchmark for that particular data set and that particular workflow. And from there, it doesn't matter how bad it is initially. It could be 50 % accurate. It doesn't matter. That's just the baseline. And if it's 95 % accurate out of the box, awesome. Maybe they can use it straight away, and it gives them enough ROI as it is. If it's at 50, 60, 70, whatever, and they need it to be at 90, that's where the work starts to configure and fine tune. train models if necessary with increasing degrees of effort and complexity. Mike Powers: So how deep, yeah, that's a great lead into like, how deep do you guys go? Like, are you going to open source models and changing the weights? Are you doing reinforcement learning? Are you doing prompt engineering? would love like, what's your breakdown for a customer? Like if I have a contractor sitting on the other side of this podcast, like how do you think about it? And how do you say like, hey, like, you know, there's 5 % we're training our own models, 95 % is prompt engineering. What does that look like for you guys? Guido Maciocci (AECFoundry): So there's very few customers where they have enough of a budget and the ROI calculations make sense to do customer training of models. we have like language models, actual pre-training reinforcement learning. That's a thing that you maintain as an enterprise asset for years to come. The investment is not build me a model. The investment is, ready to spend a million dollars maintaining it every year? Right? Mike Powers: Mm-hmm. Guido Maciocci (AECFoundry): When it comes to things like computer vision models or sort of off the shelf models that you can fine tune or train with little data, we try to prefer those. ⁓ the lowest effort possible. But we start with prompt engineering and orchestration. That's our go to to begin with, with the best models possible. And then we work our way down to try and optimize if we have to optimize cost or latency. And if we have to train, we try and sort of that's the last resort. Because you get into training, then there's all of the maintenance surface for a company to that going in the future. Mike Powers: Yeah. Have you seen any, do you obviously not have like disclose names or processes, but have you seen any ROI value for someone like, or are you guys helping them train a foundation model, like retrain, I guess not retrain is right word, but like pre-train a model. No. Guido Maciocci (AECFoundry): Pre-trained, no. No, pre-trained, ⁓ no. We've done computer vision model training for specific. know, settings and workflows. And we fine tuned both vision language models and language models for domain specific applications, but that doesn't require much effort and it's pretty quick to manage and redo if you have to. But even fine tuning a language model that, you know, you can create synthetic data sets, you know, it's not hard to do. Mike Powers: Yeah. Yeah, yeah. Guido Maciocci (AECFoundry): But there's a lot of effort in maintaining it in the future and keeping it up to date. Mike Powers: next model comes out and ⁓ better than any of the fine tuning that you're able to do. And now you're back to zero or back to zero, but. Guido Maciocci (AECFoundry): Yeah. Yeah. And other thing is in this industry, there's little static data. So fine tuning a model. Andrew Zukoski: of Guido Maciocci (AECFoundry): If you have a product and you've got your own domain and you've collected all that data and you're offering it to other customers, there's an advantage. if you're a customer, if you're a client, like an enterprise client, there's very few pools of data that you can say, let's take all of this and then spend a bunch of money fine tuning a custom LLM. And then we don't have to do that anymore for like a long time, right? That data is just going to keep changing and it's going to be obsolete very quickly. ⁓ Andrew Zukoski: I mean, is that just at the lake, like the, like you take all your projects from the last year, you train on all the documents, all the QA issues that came out of it, but like next year you have a different project mix. And so like there are a bunch of implicit assumptions that no longer hold or could you go into, yeah. Guido Maciocci (AECFoundry): Yeah, wouldn't probably train a language model on project data. You would probably train it on things about the like standard operating procedures and domain organizational structures and things like that. More about the domain of the company stuff that doesn't change. Andrew Zukoski: Yeah, okay. Okay, Mike touch on ROI, I'm running RIPSHOD. When AE firms are deploying systems looking to automate these things, like where is the impact the business? Luigi La Corte: But we don't shoot. Sorry, Andrew. Andrew Zukoski: The classic thing that I've heard, it's been a while since I spent a lot of time inside of A &E firms, is well, they bill hourly. And so if you make them more efficient, they bill less. And so actually, it's worse economically for the decision makers. And so I'd love to hear the, is it top line? Is it bottom line? Is it risk? Where are firms seeing ROI? Guido Maciocci (AECFoundry): At the moment, it's just purely based on efficiency gains. It's, that's the easiest thing to measure because, they're starting the point of view of AI can ⁓ reduce number of people ⁓ and can do more with less. That's sort of what everyone is thinking. So they measure it in terms of if I needed X number of people to do this thing, if I can do it with one person and an AI agent. Awesome. The hard ROI, think, to measure, that's like sort of the immediate superficial ⁓ But the ROI that's very difficult to measure is sort of the compounding ⁓ of not building one use case. I'm not buying separate tools, but building them on a platform and getting them to sort of compound off of each other. That's ⁓ hard to put into numbers. Like when you start to combine, you ⁓ do three use cases and now you've combined three different data sources and knowledge bases and you've got different agents and. workflows and suddenly you can start to ask questions that you could never answer before. You know, what is the ROI of being able to do that? What is, kinds of insights are you able to now get and how does that impact the business? That's very hard to put a number on. Luigi La Corte: Something along the same lines, yesterday, the day before released Gemma 4, a one-byte open source model, not state of the art, but very close. It's made headlines because of how ⁓ it is relative to its intelligence. You could run it on your phone. And my CTO sent me a message yesterday saying, look how amazing this is. So I the trend there suggests that will be, as some people say, too cheap to meter. It'll be nearly free if it can run on. something as ubiquitous as a cell phone. This is a two-part question. Given that, what should contractors, AEC owners and professionals do to take advantage of this, let's say limitless intelligence, structurally? a data lake, AI committee, something. What specifically should they do? And two, who do think is gonna benefit the most from this? Small firms, big firms, medium-sized firms? What's your opinion on both of those things? Guido Maciocci (AECFoundry): I think anyone can benefit from this, but in different ways. So I think small and medium firms are going to benefit from like cowork and Claude in a very meaningful way if they learn how to use these tools. Like that's the other challenge. It's like nobody's actually using these things or, you they're to buy products, but without realizing that these tools will pretty much... build you whichever product you're trying to buy instead or will replace it in some way. think enterprises, large enterprises can't benefit from the same tools. They can't benefit from of generic platforms and they can't because they're just not suited to the kinds of workflows they need to perform. And they're not going to benefit from buying vertical tools for every workflow they're trying to solve, right? And vertical AI tools, because they're going to stay disconnected. They don't talk to each other anyway. There's already too many SaaS tools in every organization in the world. If anything, should be the goal of AI should be to reduce the number of products and software licenses a company is using. So I think enterprises that are going to benefit the most are ones that are doing two things. focusing on data, so making their data AI readable in whichever way possible. And two is ⁓ on infrastructure. So building out the piping and foundational infrastructure they need so that they can keep up with the technology as it changes. As models get smaller, they get faster, they're different. Rag is a perfect example. Like the entire world for nothing but two years, rag was the answer to absolutely every possible question anyone had. And now it's like, ⁓ yeah, it's super useful in very specific use cases, but you know, there's a whole other world to unlock actual useful and things like that. So I think those are the, there's two scales of. what you need to do to benefit and it depends on whether you're a very large enterprise or a very small company or in the middle. Andrew Zukoski: A &E firms gonna give all these gains back to owners or is it gonna change the operating profile of the business? So I'm thinking of Revit. Revit came out and it made creating stack of drawings a lot You needed a lot fewer architecture engineer hours to create some number of And so there was a very brief period of time ⁓ where profitability soared. And then the industry figured this out and reconfigured to the point that fees were commensurate with this in some way. People weren't fooled by the stack of drawings. I don't even know that it's a good thing that there were more drawings. I'm not arguing that. But the efficiency got better. before after, it's not clear the margins or profitability or stability of the businesses that deployed the efficiency technology were better. What happens with the? AI technologies. What are you seeing? Guido Maciocci (AECFoundry): Well, everyone's talking about outcome-based pricing as the next of evolution of the industry to be able to or deal with this. ⁓ know if that's going to happen because it's really about architects deciding to move an outcome-based pricing model. Maybe small companies do that, but large companies working on large projects, they're obligated by very, very complex contractual structures that go all the way back to government. So unless all of that stuff changes, unless the contractual nature of how supply chain interacts with an AC changes from top down. really don't I don't see this having any kind of. legs within the industry, like the idea of outcome-based pricing. So I think it will result in what you're saying. There's going to be a period, potentially very long in this industry, because it's very slow to adopt. There's going to be a period where people can capitalize on the asymmetry. Those that use AI and charge you the same number of hours that people that don't use AI, but they do it in like, they have 80 % margins instead of like 2 % margins. There'll be a transition period where that's gonna happen. I think that'll last a long time, many I ⁓ assume. because of how slow the industry, mean, if you think that there's like the majority of the industry doesn't use BIM, and we'd be talking about BIM for 20 years, you know, I think it's gonna be the same deal. Luigi La Corte: Specifically in AI adoption, Guido? Because it seems not to be the case. ⁓ in GC, at least from what I'm seeing, is tinkering around with Claude. And I know there's a difference between playing with those tools and meaningfully adopting AI. But it seems like those same curves of the past no longer apply. Andrew Zukoski: But at least the on-ramp for adopting Claude and other AI tools is way, way shallower. It's awesome. Guido Maciocci (AECFoundry): it Luigi La Corte: Yes. Yeah, there's a plot, there's a plethora of on-ramps and I don't think, you know, it's mostly permissionless because of how cheap these products are. Like Anthropic is subsidizing these plans. Yes, the compute is basically free. do really wonder if we can still rely on that old trope, construction moves really slowly. A16Z published an article recently. you know, asking for builders to improve space meaningfully. know Mike has some opinions on that. I'll pass it over to him and I think he might have some questions for you. ⁓ Mike Powers: Let's get Grito's first. Yeah, yeah, I love Grito's thoughts initially. I think one I would have on the outcome-based pricing, I have definitely outcome-based pricing piled. I think architects and engineers, it's hard, and it's always going to be hard because they bill hourly. Just like how hard for lawyers with Harvey and Legora be able to charge more for the service they provide now that AI can help them ⁓ for you contracts faster. the lawyer shift is hey, ⁓ And I would talk about this a lot as like due diligence took three months before we charged a million dollars in fees to do that. If I do that in one month, I'm going to charge you $4 million. That makes a lot of sense in legal because there's like beginning and an end. In AE, that's not the case. Like you're not going to be like, I'm not reducing the construction of the building by 90 % because I used AI to design the building. So think that for A, firms can be very, very hard for outcomes. think for contractors, subcontractors, vendors and suppliers, they will be able to charge for outcomes because they'll be to do things faster and respond faster and do X, Y and Z better. ⁓ Charge for outcome. ⁓ for outcome. Yeah. So it makes a lot of sense for them to tie to it. It's like, hey, we're going to perform X or Y and this is ⁓ return. ⁓ example, we tie to customer's gross margin. If increase their gross margin, we want a portion of that gross margin as Andrew Zukoski: they already, especially at the trade level, already price on outcome. That's the whole thing. Mike Powers: compensation for that. So that's like very like black and white from that position. And I think AI helped with those pieces. Then go to the A16Z piece. My initial gut reaction was it was a bunch venture capitalists trying to talk about a space they really didn't understand. I think even the definition of like $3 billion of Revit spend, to Guido's point, that doesn't expand to $30 billion of Revit spend because there's AI now, there's still a $3 billion bucket to replace. But Greta, we love your thoughts and opinions on that. Like the positioning, you're seeing in the market. I mean, you're very much in the weeds with customers right now. And we just of loved your understanding of what are moving the needle. It was a lot around, like, I'm sure you read it around, like, design, MEP What was your initial thoughts? Guido Maciocci (AECFoundry): Yeah. So, so what is the, this is the, this is the most interesting thing to me in this industry. It's, everybody's focused on trying to automate the stuff. And this comes back to benchmarking as well, but everybody's trying to automate or solve problems that is really, really bad at solving. And because Like the industry had like it's, it's awakening with chat GPT. It's like, intelligence exists and now we can use it. It's like, yeah, okay. But you, we've, there's a whole history of artificial intelligence research and it's been albeit less accessible. It's been usable decades, like simpler models and things that have been solved over and over again in different industries, all of machine learning and statistical analysis. The industry has never touched that stuff. Like we've always kind of stayed away from it. Maybe engineers a little bit through certain models simulations and so on. But generally there's no one's really tried to use, you know, the that is ⁓ mature. Everybody's trying chase the bleeding edge, but not understanding that if you do chase the bleeding edge, then you have to be ready to take risks and fail and like, you know, keep chasing it. But there's this weird thing in this industry that it's like, you don't, you, you realize you're at the bleeding edge, but you're still expecting bleeding edge technology to work like a robust piece of technology that's been around for, for a decade. And there is that technology, but the industry is completely blind to it. And computer vision models are an excellent example. Andrew Zukoski: a lot of pleading. Mike Powers: Why do think that is? Guido Maciocci (AECFoundry): just, I think, just general ignorance technology and AI, you can't ignore it now. It's become part of daily life, social life. it's like the internet, right? cannot ignore it. And so as an industry, everybody ⁓ talks about it, hears about it, reads about it, but they're missing ⁓ the whole history. think AI started three years ago with Luigi La Corte: But why is that relevant, Guido? I mean, it's not usable, ⁓ there's been computer vision models for a long time, but ⁓ could not expect someone in AEC to implement that on a data set and then solve the UI UX problem. or solve the accuracy problem. it seems to me like as the non call myself, let's say like the, I'll take the position of non-cynical to venture capital dollars. They now people in construction are aware that they can solve their problems with AI firms like us ⁓ this call have said, okay, I'll take those dollars and then I'll create an experience where people can benefit from that technology gain. And becomes ideally a self-reinforcing loop. People pay for those services. The market expands, people invest more money. What, is your, what's your take? ⁓ saying that, you know, venture capital is a bit late to the party? Or are you saying, yeah, it is actually a good thing because now people can benefit from those technology gains in a more distributed way. Guido Maciocci (AECFoundry): It's definitely a good thing. I think the industry is late to the party and it's, you know, there's a lot of problems in industry that can be solved through workflow automation without AI, with machine learning, with a of different stuff that is way more robust and proven. than the latest and greatest technology. I think it's just a mindset thing. You have to understand the context within which you're adopting the technology where that technology stands on its own development and where it is. That's what I'm concerned with is trying to the wrong tool every single job. Andrew Zukoski: Yeah. like feel like a of my recent conversations we've been looking at, the way some relationships are arranged in the industry, it feels like people are ready to question another level of assumptions in their businesses in a way that they weren't three years ago or five years ago. And it might be the technology that you would use to deliver on some new vision or some new way of running something, I mean, Case is a great example. Case, a small architecture firm, automated the heck out of interior office fit outs and offered outcome-based pricing to WeWork. They priced per square foot until they got bought. They it. It worked great. It allowed to out an incredible amount of office space incredibly quickly and Adam Newman to have interesting life. It feels like there is something just to this shift of like, we could do things differently, willingness to question and gauge more substantively than existed a few years ago. So it's interesting to think of that as the big unlock. Guido Maciocci (AECFoundry): Yeah, and I think that's key. you try to just bolt on AI to whatever you're already doing, it's going to work very well, I don't think. Re-engineering processes and operations around what actually offers, I think, ⁓ is the right Mike Powers: I that makes sense. So I think one of the reasons that we reached out or Louisa reached out was I think you're very prolific at publishing your receipts, which I love and appreciate. I think it's much, much needed in the space to drive trust to the industry. There's obviously been some news yesterday. DocuCrunch announced that they were acquired by Trimble. You saw the data grid acquisition. We'd love to get your overall thoughts of like AEC startups thriving and dying with this recent wave of M &A activity. It's going to be interesting. It's like, yeah, we just love to get your idea. I think you published an article about this or a report about this. I'd love to unpack that. Guido Maciocci (AECFoundry): Yeah. Yeah. Andrew Zukoski: Wait, you told, wait, where were each of us? Like were we on the die side or on the thrive side? ⁓ even worse. Luigi La Corte: I don't think we were in the, I don't think we were in the report. Okay. Guido, you have to on this call, you have to on this call say which one of us will die and which one will thrive. Provision, build vision, join, go. Mike Powers: ⁓ wait, we weren't on the reports. This is why we're not TVPN yet. This is why we're not getting open AI. Andrew Zukoski: Yeah. Mike Powers: Yes. Guido Maciocci (AECFoundry): Yeah, report sort of highlighted startups that made a for themselves within the AEC tech space over the last five years. And they're sort of the household names when you think of AEC tech startup. And now, of course, there's a whole new generation with AI and everything. It's completely exploded. at a time not so ago, It was literally like a handful of startups. Most of them were focused on and feasibility. And that report and article was more focused on how Autodesk and Forma play, which is now also playing out. Everything is Forma now. It's a brand and it's a Forma platform and everything is changing names to be more AI focused. And I think the set of report And my opinion still stands today, especially with the last week of &A announcement. I think it's the startups that focus on making AI readable, AEC specific data ⁓ AI readable. and the ones that focus on the knowledge layer to enable AI to navigate domain-specific knowledge. Anyone that's doing that, and it could be as a platform, it could be a vertical product, it doesn't matter, but I think those are the ingredients of successful startups in this space now. Mike Powers: I guess like the last two acquisitions in the space have been the opposite of that, been the opposite spectrum. Both Grid and DocuCrunch, I would say are taking unstructured data and structuring it and are the knowledge workflow layer. So like, as you think like the targets of acquisition, so it's Trimble, Procore, Autodesk. Having gone through an Autodesk acquisition, I think the knowledge layer piece that you mentioned is very, very hard for them. them plugging in point solutions, plugging in data layers. I would guarantee Autodesk is on the hunt for some sort of data layer, AI company right now, AI agent company, however want to position it. It just seems like the market isn't responding that way right now. And be curious to see how that changes. And why you think that changes, really. Because I don't see the Procore strimples or Autodesk of the world saying, hey, we're gonna replace Forma, we're gonna replace our plan acquisition with a new tool. It's gonna be these like plug and play, here's the node connector pieces. ⁓ Guido Maciocci (AECFoundry): Yeah, we'll see how that plays out. think as are all sort of incumbent legacy companies don't really have a choice. It's not like Autodesk can recreate entire ecosystem from the ground up around AI, but that would be the ideal situation, right? If we have what we wanted, then you would have Autodesk become more like Anthropic, right? and then they build the Autodesk ecosystem around the future vision of the industry and what that would be. That's never, it's never going to happen, of course. Mike Powers: But if you, if you, would push back on that because if you asked how they've structured for and why they have Forma, Forma is that. Forma is like, hey, we don't care about the brand point solution anymore. We pay, we care about the data infrastructure. We have all the tooling to do that. Um, I, if I were to take a guess, I would guess that they would buy a company that is saying that they're an LLM for construction and plug it on top of Forma and train on the proprietary data that they have that no other startup can compete against. Guido Maciocci (AECFoundry): They might be. Mike Powers: And that would be the bull case at their 90 % margins right now. Guido Maciocci (AECFoundry): Yeah, no, think they will do that. And I think all of that stuff will work within the confines of their own And that's it. ⁓ Mike Powers: I just like the reason I push back is like everyone thinks like auto like they've been crushing their earnings. They continue to get market share. everyone's like, ⁓ our auto desk is like this archaic organization. I push back like they're crushing it like they are one of the fastest growing vertical SaaS companies. They're reaccelerating. I think they have a lot of upside in this world where like workflows get automated So any point solution to zero like that we like to talk about and then Autodesk Procore become these massive infrastructure layers, on the construction side, one on the design side. And I feel it's a hard for any other tool to wrap around that. ⁓ Guido Maciocci (AECFoundry): Yeah, I mean, if that did happen, you would still have the problem of ⁓ enterprises having to use both those platforms. So even if they get to the point where they have infrastructure and the platform. Exactly. And that's the problem. ⁓ Mike Powers: They do now. They do today, It's not a problem, though. It's not. Guido Maciocci (AECFoundry): I mean, it is for the clients that we have. Like the main problem we solve is the fact that they can't manage data and knowledge across different systems. And they have workflows that span 10 departments and 25 different people. And the reason it takes them to perform that workflow is because of decentralization and they... they separate all of that data out and knowledge and expertise. Like most of the work we do, if we had to generalize, whether it's in generative design or AI or automation, it's really about taking fragmented that span multiple departments and data sources and experts and combining them into something holistic that. fewer people can run, either one person can execute multiple people can execute together collaboratively in a way that they can, you know, review the outputs and participate in the process. ⁓ Mike Powers: But isn't the like, is for an A firm, they're going to centralize back on Autodesk and for a construction firm, they're going to centralize back on a Procore. Guido Maciocci (AECFoundry): Yeah, but I think fine. I think they'll solve the problem for the firms that can run an entire business on Autodesk. The issue is that all of these platforms solve for domain specific workflows, but they don't solve for operational workflows. So Autodesk becomes a business operations platform, any architecture business, any engineering consultant, it doesn't matter how big or small they are, they're still a business. And all we talk about in this industry when it comes to AI is solving design and solving engineering, when really all of the value today, at least, is in solving business operations. Luigi La Corte: That's interesting, that's a different opinion. mean, that would bode well with the traction that teams like Agave are getting, even within the Procore ecosystem, because they are kind of last mile of connectivity between Procore, Sage, or ERP. I mean, that does make sense. I know though, ⁓ if where all the value is. At least I talk to teams, it's not like... I'm desperately trying to integrate my ERP into this. It's like I'm losing money or I want to win more work. Those things don't, I mean that doesn't conform to your worldview. Guido Maciocci (AECFoundry): It's, think there's a disconnect between what people want, clients want and what they actually need to achieve what they want. The focus is always on the core business workflow. So engineers will ask you to automate their structural calculations and their Tecla models. Architects will ask you to automate their QA, QC and check you know, all my dimensions are correct in the drawings. But. ⁓ Andrew Zukoski: Yeah. Guido Maciocci (AECFoundry): We never, when we speak, we eventually get to this, right? We eventually talk clients up to this, but we never have clients that come to us and tell us, we our system badly. Can you help us improve the way that we search for leads or how we identify opportunities? Can you help us accelerate the way we do financial modeling and financial estimation upfront before it even gets to become a project? Can you help us with setting up the team performance and task management across the team so that we can track and improve? All this stuff, AI is amazing at, but hardly any client ever comes in and asks AI to do one of those key business tasks. It's always like drawings and engineering counts. Luigi La Corte: Well, it kind of, seems like, I've been talking a lot, so I'll pass it over to Andrew and Mike, but I don't imagine, let's say before Harvey, lawyers were saying, I want to reduce my billable hours. It was once Harvey was part of the system where they created a state of disequilibrium. Okay, now there's game theory. I have to use them to be AI native to sell more product that people started adopting these tools. There was a lot of... Mike Powers: you Luigi La Corte: there was a lot of uncertainty in the legal market as to whether or not these tools were going to be adopted because of these very obvious incentives. I want to sell more hours because that's how I make my money. So I don't know, it seems like... These people who are maybe hiring you, they're not saying, ⁓ I want to automate my core competency because that's how I my hours. But that doesn't mean that that product can't be created and be very successful. It just needs to proliferate and it needs to be successful once before people will adopt it. So maybe, maybe they're not asking for that just because of their incentives. you see, do you see that being valid? Guido Maciocci (AECFoundry): Absolutely. just see there's sides to it. One is ⁓ really stuff to solve and that needs a lot of investment and it will have a high ROI for those that sort of are in it for the long run. But if you're trying to adopt AI and get as value of it as quickly as possible, then business operations is where you're gonna get the most value. And software engineering, where it's proven, that that's where all of these tools have ended up in. Marketing, HR, and software engineering, and customer support. If you can map those problems to domain-specific use cases in AEC, So for example, customer support in AEC is field data collection. It's site inspections. That's customer support, right? You got someone walking around taking pictures of defects. That's the same as customers reporting bugs in a software system. It's the same thing. Or like a customer leaving a one-star review in a shitty restaurant. It's exactly the same as a worker on site, know, leaving a complaint about ⁓ safety hazard they discovered on level three. So if you, if you start mapping. mapping in that way, like look at where AI has had the most impact outside of the AEC industry and map that back to the same sort of problems that we have in a more domain specific way. I think that's how you really start to extract value from AI. And yes, I believe that eventually we're going to have more automated design, more automated engineering, better compliance checks, more digitized workflows. Of course, we'll get But it's a question of like a really long strategy and high investment roadmap versus wins in proven Andrew Zukoski: Okay, I have a question to turn it back. What counts as thriving for an AEC startup? Guido Maciocci (AECFoundry): counts as thriving? That's a good question. Like an AI startup or like any AEC startup? Mike Powers: you Andrew Zukoski: Just any AAC startup, did AAC founder raise venture money? OK, good for you. Mike, did. Luigi, you did. Right? Joyne did. Right? I just like, I may have told you guys this story way when we had a meeting with Sequoia. We pitched Doug Leone, who was the plan grid investor. Right? he's the managing partner at Sequoia. He's like the ⁓ god of capital. And. Guido Maciocci (AECFoundry): No, we're we're- Mike Powers: Thank Andrew Zukoski: he was just politely finding a way to pass on us because he didn't like the way we smelled or whatever, which is fine. But it's like drilled into my brain. He's like, construction. Plan grid. I we are in that for like five years. Exited for 800 million. You know, it's okay. Don't get me wrong, but... The whole like, get like the two thriving startups in AEC technology are in the last 15 years are arguably Procore and Equipment Share because they IPO'd. And everything short of that is not. Luigi La Corte: If you're an investor, you would think that. Guido Maciocci (AECFoundry): Yeah. Andrew Zukoski: Yeah, that's true. Oh, but it's also like, it, it, um, I feel like it speaks to the, um, All right. Scale of the challenge. Like Mike building connected as a beloved product, like all of our customers using are never going to move away from it until out of the scratch price up too high. But, um, the, the model of like, you build these things and they get bought by these big public companies and then like rolled out with a larger sales force has not moved the needle on. like industry outcome. Like it hasn't It's been great for like, you the employees and owners of a number of these companies, but I just know that we have a lot of examples of truly thriving. Like. Guido Maciocci (AECFoundry): Yeah, if you base it on impact, then I would have to agree with you. The biggest companies in this space are arguably some of the worst products that I've ever seen personally, just in terms of like basic product design and user experience, know, simple standard stuff that you would expect to be of a higher quality. But thriving in the sense, I mean, in the AI space, one company that I love to bring up all the time is Enplan. And I think, you know, in this space, they, in my opinion, are arguably of the only, if the only true AI startup or AI company in AEC. Just for the fact that they've done it and they've built it from research, from scientific research, all the up to a product and a platform, maintaining a research lab. And they do it properly. Like if there's one company to pay attention to in terms of if you want to be a real AI company in this space, that is a really, really good one to look at as opposed to like an AI enabled product, like a true AI company. So I mean, thriving in that sense, you know, they have an IPO. They certainly don't the same. the same stature, but you know, they got their, they got their hundred plus millions of funding. they're doing amazing work. Luigi La Corte: The Enplan guys are very smart and they have a proprietary data set. Guido Maciocci (AECFoundry): They do, I don't honestly know how much that's worth anymore in day and age. Luigi La Corte: Well, I was going to say, mean, like, what's a real AI company these days? And does it really matter? Because it really is just about the output, right? Like if you can contort an LLM to perform magic, it doesn't matter if you're a rapper. really does. I don't know, for sure. Guido Maciocci (AECFoundry): I know, no, I have nothing against rappers. think rappers are essential because, you know, that's it's user experience. can't, if you can't make the technology usable, then it's useless. Luigi La Corte: Of course. It's really like the whole thing is kind of a false dichotomy, right? It doesn't matter. There is value as long as the customers can experience value. But I do do believe that the best solutions here will be ⁓ about their benchmarking and how they'll be improving. I think we do this quite well within provision. We don't do a great job marketing this, but everything has a benchmark. And then, yeah, after a certain point, like if you want to increase your margin, you go build your own models or you distill a model. You you can't do that, but you take ⁓ once you that data to give of the best outcome to the customer being reducing costs giving some magical experience. Guido Maciocci (AECFoundry): Well, here's a question. Do you think that Document Crunch had benchmarks and that's what they sold to Trim? Luigi La Corte: they have benchmarks? Well, never saw. So I'll say this. I have a horse in the race. We compete with document crunch on contract review. I don't believe contract review is an enduring. business. just too ubiquitous. Did they have a benchmark? I would imagine they did. I think the of Doc and Crunch Relief from what I understand is they started before LLMs and they worked with the third party to build a proprietary data set. I'm struggling to remember the name, but they were a classifier, like they built a classifier and their neural net would say this is an LD, this is a warranty, et cetera, et cetera, et cetera. Now, I don't know if that product was very successful, that was an old product, but must have had some benchmarks. Any company that is using LLMs worth its salt has know that if I'm putting this out, it's better, to a degree. If you want product velocity and you just know like, hey, I can plug in this open source model, Moonshot AI Minimax and give it the right harness, our customers need to experience that and we'll do the evals later. but I think it depends on the workflow. In their workflow, you're sitting across the table from business owner who has their life on the line. I think Josh had, I think they had benchmarks. I don't think that's why Trimble purchased them though. It wasn't for the benchmarks. Guido Maciocci (AECFoundry): No, no, no, no, of course. Luigi La Corte: Andrew, Michael, what do you guys think? Mike Powers: if DocuCrunch had benchmarks or what they got bought by, what the question? ⁓ Luigi La Corte: Yeah. Why do you think they, why do you they purchased them? Well, the question I think from Guina was like, did they have benchmarks? But you know, think the, maybe the question behind the question is why were they purchased? Mike Powers: their purchase, this their AI play. It's like a pretty like, we have a bunch of data structuring. we use the pipeline that DocuCrunch has created to do extraction to supercharge rest of our workflows? They didn't get a stock bump from the acquisition news. ⁓ They did not ⁓ get a stock bump. was trading under after hours after announcements, but it needs to be Luigi La Corte: Say it Mike Powers: I know Trimble's interesting because Trimble has like 9,000 dip. That's over exaggeration, but like 9,000, they're like product lines and they've GM to the product lines where like the Autodesk Pro Core is more of like full businesses around these. So think it's interesting. Like this is going to be like they're in, I did they change? Andrew Zukoski: that still true? Do they still have two members? I think that they've been, like since Rob Painter came in, they've been, like, there's a lot of interesting change that's happened inside of Turnbull, it's like three years, but yeah, go ahead. Historically, 100%. Mike Powers: Yeah, it's like the structure like this. They like the structure acquisition now is like the materials kind of business unit. Again, that's that's it. It's more like acquisitions. I be curious to see how that plays out. But this would be their intelligence play. It makes a lot of sense for them. Go build this on to every customer that they have. Charge them 10,000, 100,000 hours a year. Say, hey, here's your AI play. It's construction specific. It's not copilot for Microsoft. I think it's interesting. think it was a great opportunity for exit. I've that been going up against Harvey and Legora, which I couldn't imagine competing against companies that have raised hundreds of millions of dollars and have gone from one to a hundred dollars in revenue in 18 months. Just talent of acquisitions of employees is challenge in that world. So a one. It's a tough position. I think Josh is great. think they, I'm super excited for them. It's been like a great opportunity. And yeah, I'm super excited they were able to go down that path. Andrew Zukoski: Great note to end on. Luigi La Corte: Absolutely. Guido, very grateful for your time. Thanks for joining us on this beautiful day in Rome. won't be the last time we have you on, but hopefully you had some fun as well. Andrew Zukoski: Guido, thank you. Guido Maciocci (AECFoundry): Thank you guys. Yeah, it was totally fun. Let's it again sometime. ⁓ Luigi La Corte: Okay, I'll end the stream now.