Latest / Construction Technology Podcast / MCP vs. APIs: Harnesses, Headless PMIS & AEC Compilers (with Clifton Harness)
Transcript
- Luigi La Corte: What's up, everyone? Good morning. Welcome to Construction Technology Podcast. I'm one of your co-hosts, Luigi, at Provision. I have my two fabulous co-hosts with me, Mike from Build Vision, who's in an airport in Minneapolis. And we got Andrew from Join in his garage. We also have Clifton from Tesfit. What's up, Clifton? How are Clifton Harness: Good morning. Nice to be on with you guys. Luigi La Corte: Yeah, thanks for thanks for joining us. So Andrew Andrew you teed this up. Give us the give us the inspiration up. Andrew Zukoski: Yeah, so we were gonna talk about MCP, the model context protocol. Model context protocol servers and clients, and thought we could start with this slide. because there's a lot of terminology here and I thought we should at least level set amongst us, and there might be some listeners who aren't familiar with all of it. And so this is a slide, it's very joint centric, but that we used with a workshop we ran with a bunch of our customers a few weeks ago just laying out all the different parts and pieces of this. And boy are there a lot. You've got like a system over on the left, in this case join, that exposes an MCP server. And what that does is allow MCP clients to connect to it and talk to it. And so typical MCP clients will be all that AI tools you know and love. Clot and codex and copilot and so forth. And an MCP server is a lot like an API, which you might use to build integrations or to pull data into some sort of like business intelligence system, except it makes it really easy for AI systems to interact with, or it's like the you know, best known easiest way to get that going. around the harness, you have the underlying model that you're dealing with, right? You have the the harness is the program and then the model is the smart thing or dumb thing, depending on the model. and then you may have on Clifton Harness: Lots of lots of dumb harnesses out there, Andrew Zukoski: Alright, is that a joke about your family members, Clifton? Harness? Sorry, it's anyway. you got it. There we go. Yes, his last name is Harness. Clifton Harness: Ha ha ha ha. Andrew Zukoski: yeah, so the like MCB servers the API, it exposes tools, tools like do a thing. They might search for data, they might modify something, they might run a calculation or something like that. and so this is the like constellation of mm concepts that like we're working inside of. You have the server, the client, the harness, the skill, the model and tools. Is this are we level set? Is this what everyone else is seeing? Luigi La Corte: Why Clifton Harness: Yep. Luigi La Corte: why should people care about MCP or a harness if the people listening aren't super familiar? Like what's the what's the so what? Andrew Zukoski: and so the so what is that you're able to connect your other systems into you know into the AI tools that you know and love, right? And so one of the things that we found that these AI systems are very good at is is coding, right? Just writing code. And so with them, with the data from one of your systems available to one of these things. You're able to run analyses or run other calculations based on it. Like the the AI can look at the data, write some code to process it, generate an output. Like an example of that that we've seen was folks pulling the wrist log from join and then running a Monte Carlo analysis of it. And like Claude can just write the code to do that. Right? It used to be you had to find a vendor and pay them, or like write the code yourself, and now you can be like Claude. I want a Monte Carlo analysis of this. It can pull the data from join. It can run the Monte Carlo analysis. And then the other thing that we've seen that these AI tools are really good at is generating artifacts, right? Like in olden days you had to like use WordPress's what you see is what you get editor to make a web page. And now you just ask Claude to like write HTML. Right? Or if you want a polished visual, you can work with these tools to create polished visuals very easily. If you want custom graphics or custom visualizations, they're really good at doing that. A lot of that's because of that rests now on just like writing code automatically. and then I think the third thing that we don't see as much, like join is largely a data source, right? Like people also push data back into Join through these tools, but like we're kind of a data source. I know Clifton's been doing work with MCPs as well. Test fit is a calculation engine, right? It's able to lay out buildings, lay out parking, lay out circulation, given certain constraints. And now the LLM and the harness can drive that and incorporate that into other workflows. And so you can basically say, like, okay, Mr. Model, like try different combinations of offsets and setbacks until you find something that hits like you know given IRR. for this site that we're looking at. Clifton Harness: Go fail overnight for eight hours and try to find some data that makes this work well. Yeah. It it it th I agree that this is the rough architecture everybody's dealing with, but I also might add that Web pages is a human MCP server. Like you go to somebody's website and you're reading it, it's like if you want to have a analogy to what a human would have is literally this visual web page or the visual interface. Those so those are the normal people protocol server, right? and and the value of MCP is for people that are like in a gateway, you know, into the API world or gateway into structured data, the MCP server will help. get, you know, a normal person over that over that hump. I do think that because the models are getting so powerful in the harnesses, they'll probably abandon trying to use the MCP and they'll just write code locally to manipulate APIs. But I could I could be way off base there. Luigi La Corte: That's an interesting take. And you think it's because well you think there's some fundamental limit with MCP that just would make the manipulating of APIs better for the model? It's like very easy for for someone like me to just plug in even just like direct codecs to the page and say, Go for setup MCP, you know, it's pretty performant. Clifton Harness: agreed, but we're trying to do, you know, fifteen thousand calculations a second or something, and the amount of time that an LLM waits in the MCP world is significantly longer than batching through an API, you know, all of those requests and then waiting for it come back. so it's just a difference in latency and MCP is really good for products that are incredibly slow, possibly single-threaded, that take, you know, ten seconds to update, that maybe shouldn't be an API, because they really were built in an era when it was waiting for, you know, the human to keep up with it. So that's where I see the the strategic difference is like I I think Revit will probably never get a great API because it's too slow to update, but it'll probably be perfect MCP target. Luigi La Corte: I see. Andrew Zukoski: So I would take the other side of that actually. I think that that APIs aren't gonna go anywhere and that if you're if you're building something, you probably wanna build on an API. But if you're doing something, like you're creating a report or you're running analysis or you wanna check something out or you wanna ask a question or something, I think there are a lot of benefits to MCP around Ease of distribution and access. Like you can add these, everybody, modulo, non-trivial IT barriers can add these. Like, like Luigi Mike, you guys must have a bunch of connectors inside of all of the tools that you're using, right? Like you're able to find these things, you're able to add them, and then you're able to like govern who's able to do what with them. Right? Like you're there's there's many more controls built in there of like, okay, I want to I want to turn off every write tool. I don't want I I don't wanna be writing to provision at all. I only wanna read information from it. I wanna be extremely confident that the LMM isn't gonna go off the rails and like overwrite everything that we've already sent to the client or something like that. Like I I get nervous around this. I ran like Gong charges you for the MCP server and they don't charge for API tokens. and I wanted to run an analysis of a bunch of our call transcripts, so I used an API token, but like boy, did it make me nervous because you can't scope the permissions For the API. And so that token had write access and delete access. Which is, you know, gotta go like audit that code before you run it, but like who knows what the who knows what the LLM actually makes the thing do. Whereas with an MCP server, you're always like, okay, I can I can govern this. If I'm in an IT organization, I can govern it for all the other people inside the organization. Like you get this ability for governance that I think is I think is important. Luigi La Corte: Sure. I I think less than one percent of let's say like the construction technology user population is gonna be doing what Clifton suggested. I think Clifton, your you will probably do what you suggested in like, you know, when you when you want it. But I can't imagine a you know, anyone using codex and trying to manipulate the API themselves will just, you know, try and make it work the MCP or just give up. but I do think like the power users will do what you're suggesting. I also I I also think that that's where the whole industry is going. A lot of people right now, we we we just signed a pilot with a customer yesterday, very large, Boston based contractor. they're getting an enterprise license with Claude. and they said, like, hey, I'd love to see your product. I really want to use it via MCP. That's the way we're using almost all of our tools. The interesting side effect of that is I think like there's a great opportunity to displace these kind of legacy platforms, like Autodesk, I think, is is gonna have a very interesting time. They were these power brokers where you would integrate in and now you could benefit from the ecosystem. The whole product surface is changing now, right? I don't need to have a thousand integrations if I'm a startup and I need to you know Andrew Zukoski: Yeah. Luigi La Corte: benefit from a large incumbent. I just need to route my way through codecs. So like you know take the scopes out of provision or take the RFIs out of provision, you know, Andrew Zukoski: Yeah. Luigi La Corte: MCP somewhere else. And great I don't have to work with Autodesk. Now we we do want to work with Autodesk, but you get the idea, right? Like You don't have this kind of integration hegemony anymore. It's it's it's gonna be interesting. I'm not really sure what the implications are, but Clifton Harness: So I I think on the on the C P side of things, I think it's just the the the gateway into into how SAS companies are gonna fully integrate their own models, their own agents, and then cut off access to to your harness. I believe that will probably happen and then Why do I think that? all right, so we go through releasing our MCP in August. we get I don't know, two or three users that put up more usage in a single month than the rest of the entire ecosystem combined. so those people I would be like, all right. Maybe they're like SDK level customers and maybe we need to have an SDK for for those kind of people. and then for everybody else, they maybe are not even on Claude or on Codex or or using any of that stuff. this is why I I think if you look at Procore and their agents, it's like homebrew, they built their own stuff, it's running inside their own thing. I think we'll see a lot more of that because the the people, you know, that sign up for for these kinds of tools are most of them not going to be bringing their own model or their own harness or their own skill files. and I think in the long run for us Mike Powers: We're see but are you seeing all your all your cust every one of our customers we talk to have a Collater Open AI subscription? There's there's nobody that doesn't. Clifton Harness: Yeah. But I've got Luigi La Corte: I see that too. Clifton Harness: s seven hundred and eighty customers in the long tail, and I've got like this many of them that are, you know, big enterprise groups that are are ac actively thinking through this kind of stuff. So I'm more responding to how the rest of humanity, you know how they're actually using AI. Like I don't it's it's at two percent market penetration to people that are in it ten hours a week. So the the product market fit for MCP doesn't need to be from Claude or Codex. It could be from a Quinn model that you put into your own harness inside your your website, you know, and then that's the new table stakes problem that we have to solve is you know, in addition to having the cloud systems and role based access control, it's cloud systems with the AI tooling and role based automation control. But that I Mike Powers: I just don't think that's the case. I think the play out is either you have a claw or opening eye subscription or using a tool like Inst i like Instinct as a user, you don't even use no I no a customer is not gonna spit up a quin instance if they don't wanna use a clawed instance. Clifton Harness: No, you are. You have to spin it up. You have to serve you have to serve the LM. Andrew Zukoski: software vendor would. This is the ProCore model, you spin up MagPy. Mike Powers: No. Clifton Harness: Yeah. Mike Powers: Just the the the for the vendor, for the software provider, they should use the frontier models all the time and expand and be able to sell it multiple times. Sell the extraction. And by using an MCP, you're pushing the cost off to the customer. Like that's the biggest value of an MCP. Hey, use this MCP and use your claw tokens, not my claw tokens to do X, Y, and Z. Luigi La Corte: I think that's I think that's great too. if y you saw I think at the recent Andrew Zukoski: They also get to control what happens to their data then. Yeah. Luigi La Corte: Well, the the the benefit Mike Powers: Mm. Luigi La Corte: from a go-to-market or even an economics perspective, if you can now sign in via Chat GPT to, let's say, provision, well great. maybe that's a free product that I thought was too expensive previously because the inference cost too much money, but you know, there's not much value, but maybe there's a distribution play there. I that it hasn't really played out in the market yet, but like yeah, I don't think we've seen an example in construction tech where bring your own bring your own inference has actually happened. but I'm excited by it. I I think Clifton Harness: Procore? Mike Powers: It's even Clifton Harness: Doesn't Procore do all that? Mike Powers: it's even one further than that, like for provision, like if someone says I need to do a scope analysis, how is provision the first tool that pops up in Claude or OpenA? Luigi La Corte: Well through AEO. Mike Powers: Right, but that's that's so we're talking to manufacturers and we're helping manufacturers become the first company to show up when someone asks for, Hey, how do I I need to select a chiller. How do I make sure I go to train select assist as opposed to carrier s selection software? And that's all in Claude because an engineering firm is using Claude or OpenAI to do all this. No one's going to Google anymore. So that that interesting perspective is it's a fear for those customers. So we're gonna help them build out MCPs so that Anthropic can recognize here's how you look at trained selectacist. Luigi La Corte: Right. Clifton, a question Andrew Zukoski: It's interesting. Luigi La Corte: for you on MCP. I'm sorry, Andrew. Clifton Harness: Yeah. Luigi La Corte: So I do agree with Mike. I I do think the models are somewhat gonna be the kingmakers and it'll be not just like find test fit, but like let's go implement test bit tet test fit and like download the or you know, hook up the MCP. Do you play around with the idea of like a freemium plan where the model can first like instantiate the MCP, get a taste and okay, now the user has discovered you or is that just not the motion because there's a very defined group of enterprise buyers and that's not not a focus? Clifton Harness: I think it's two things at the same time is the the buyers don't I mean I like the buyers are not all sitting in the AI harness as much as we are. and they're you know, people that are Genuinely trying to figure out w where where they put their their businesses, you know, resources invest in the future. you know, so I'm not talking about, you know, billion dollar construction companies. I'm talking about 250k a year architecture firm. For them it's it's I don't I'm not ever gonna put any of my data into into the chat 'cause it's gonna steal all my ideas. the level of suspicion is very high. and especially in real estate, where your pipeline and your track record and all that stuff is like is your your IP. you know they're less likely to to really wanna put all of their information into anyone's LLM model. So it's I'd say there's a normal distribution and on the side that is, you know, anti-innovation, which is eighty percent of I know I said normal distribution. Eighty percent of people, not eighty you know, not normal distribution of of decision makers, they don't really want the world to change. and you know, please start to use AI, okay. you know, they'll ask Google, you know, that this is why Google Search, I think, taking the the chance that they did is so huge is like that's now a lot of people's first interaction with an AI model. so you know MCP for us is two things. I have 21 employees here and you go look at the Claude usage every year and you can see you can see growth. But when we released MCP, the amount of internal usage of TesFit went from, you know, 1X to 30X. It enables your internal business operations people to actually understand the entire surface area of your app in a way that they didn't know before. Which like you have mixed panel, you have all these things wired in, you're like, well, how do we not know all these things? Well, because we were selling the software and running the business and supporting it and We weren't always doing all of those things at the same time in one, you know, in one environment. So I'm of several minds. yes, it's absolutely a gateway into thinking about APIs and structured data, and that's for the one percent of hardcore builders. For the next ten percent, it's a way for them to learn how to do long running tasks and to treat work like baking it instead of you know hunting the animal and killing it. Like, here's all the work, you give it to the thing, and then you go to sleep, and then you wake up and here's all the work. For the rest of humanity that isn't in this this group that I would, you know, term builders and and tinkerers, I think the custom LOM sitting inside your own, you know, your own server, that'll be where everybody's trying to get to, you know, by the end of the year next year. Andrew Zukoski: You custom LM sitting inside of the software vendors service. Yeah. Okay. Clifton Harness: Yeah, yeah. Andrew Zukoski: This is what this is I mean, we can quiz Jeff at some point, but like this is I believe this is Procore's strategy as well. They're saying like, look, they're all the adoption hurdles, people want us to do it for them. They don't want to do it. And I was I was honestly very skeptical of this, but that I don't think, at least in construction technology. MCP adoption is very widespread on a user basis. Like I and I we're mostly in the enterprise space. I would say on a logo basis, there's some builders and tinkerers at every logo of everything installed. But the like most places that have enterprise cloud subscriptions, at least in our experience, Mike would be interested in yours, Luigi would be interested in yours, have have it locked down where you can't just add connectors to it. And so there are these like IT barriers to it. It's made easier if you have public listings for these things, but very, very, very few construction technology companies have public listings for their MCP servers. There's a few in Anthropic. I don't know if there are any in Codex or Copilot. Copilot is like the highest, highest bar of validation. Claude just like waves you through without validating anything. no offense to Claude. And Clifton Harness: Right. Well well th think of the opposite ends of the spectrum where those two products started. One Andrew Zukoski: Yeah. Yeah, that's fair. Clifton Harness: started at at like high end business enterprise and other one was software developers, right? Andrew Zukoski: Yeah. The yeah, so I don't like the the we s we see substantial adoption friction. it feels like it feels almost like a like a web app or something, where it's like a new like category of product surface area that like you need to have and then there's everything associated with that in terms of discoverability and installation Luigi La Corte: Right. Andrew Zukoski: and and so forth. But there's there's real friction there. Luigi La Corte: I yeah, I'm I'm of two minds on how this is gonna play out. One way Procore can become the agentic work surface, and then people will want to MCP into Procore. Alternatively, they cannot fight the rise of co pilot, codecs, Claude within construction, and they're gonna Andrew Zukoski: So much money there. Yeah. Luigi La Corte: be subjugated by that work surface. And then but they're not offering an MCP. So it's either it's actually quite a bold strategy, right? Because they're not MCPing it, will people decide to go to a PMIS that does have MTP as a result? Maybe. Maybe. Clifton Harness: ProCore's core user is like someone in the field reviewing a document. You know, it's like the quickest way to get them the information they need is not with any more than one system. And I think that's what they're they're optimizing towards is is that. I I can see a world where the MCP does enable the back office to to connect in and if they don't already have people demanding that, they will. And they'll have to build that. But they're they're optimizing for their you know, the volume of users that they have. Why? 'Cause they have to support all of those users continuously and that's great for CAC, not great for for your metrics. Is it magpie? It's you it's it's a it's like a Gemini-based model, I think. And if that can carry a significant amount of that customer success load without needing to retrain the workforce, which you can't do it in construction. Fundamentally you can. I think there's a a massive market for them to to just earn, value on the delta between what it costs them to run the model and what what they, you know, get get to bill out. and they're they're Luigi La Corte: Clifton, but Clifton Harness: go ahead. Luigi La Corte: sorry, the only point the only thing I want to interject is they're not even letting people drive this. It's like, but it's behind a paywall. You cannot get the value from all their agents without forking over a lot of money. So I I it's a risk, right? Like people have to first say, Yes, I'm gonna opt in, I'm gonna spend thirty percent more. Great. Now I can get it. But if you're not Clifton Harness: Though but they're in their first year of this modernization, right? So they're like you don't even know if your product market fit on stuff until two or three years in, maybe. Like it it's a they have to make a bet in one spot so then it can then move to where where we need it to be in construction tech, right? Luigi La Corte: My bet is that there will be some users who are already saying, okay, I'm ready to be spending too much on Procore. AI has promised me that I will get, cheaper software. they keep increasing rates, and I don't even have MCP here. I'm gonna go to another PMIS. And then I've already purchased an enterprise subscription to Claude. Great, now Claude is my work service. I'm sure that will happen. I just don't know how widespread that is. So I think everyone's saying there's a bull case for MCP. at least I like I I believe it. Like I I don't know if it makes sense to go up against codecs. Andrew Zukoski: Mike, you seem to think so, right? Mike Powers: Think C P is a stepping stone that that you're not in two to five years no one's gonna be using C P at all. Andrew Zukoski: Why? Mike Powers: You're gonna you you're you're either in the construction world, you're either gonna be using an instinct type tool where you're just texting something and it just does all the work in the background and you don't have to worry about it because the bar barrier of entry is so low. Or there's gonna be super users that are using Palantir style style things like McCarthy does. And it's gonna bifurcate into two markets. Cliftons are exact like in Clifton's like Andrew Zukoski: How will I w is it is it the in sorry go ahead. Mike Powers: Clifton's position is so right, the guy the tr the guy that's driving the truck that's doing earth moving work on a job site, why can't he just text something? And it just does all the work and he doesn't have to worry about he doesn't care. If it's a Claude, if it's an MCP, if it's test fit, if it's provision, he just wants the answer. I think that's one of the areas that Procore is gonna struggle with is because they have so much revenue built into SAS, for them to do something like that would drastically reduce that revenue. It would be very hard. And then how do you monetize on it? instinct just raised at a fit thirty billion dollar valuation without any revenue. so there's some there's something there. And that's that that's the maybe the biggest even even what Meta's doing, there's no reason why a company can't just go do that for construction and just make it all Andrew Zukoski: Won't is it that if I have Muse or Instinct I don't need Procore? Or is it that if I have Muse or Instinct the work surface shifts to this but it's still connected? Mike Powers: I don't there's two type two different Luigi La Corte: couldn't Mike Powers: types or there's two bell bar bell Luigi La Corte: you couldn't use ProCorp. Mike Powers: customers. There's the customers Clifton Harness: Work service. Mike Powers: that don't that need Procore, like McCarthy, and there's the guy at the the the c the guy that's laying carpet to on the job site today does not need to access Procore. Why can't he just text? Andrew Zukoski: But they don't have Procore today. Mike Powers: That's what I mean? So that's those are the two bifurcations. And then that world, no one cares. It's just it's you're using whatever is obfuscated away, and then the other world is they're using they're gonna build out their own rails because they're gonna have to do stuff at a higher at a higher clip. Clifton Harness: Why can't job site guys text Procore? You know, like you could whitelist their phone number and then that's like that's honestly the perfect interface. Luigi La Corte: Yeah, they probably could. I don't think I don't know if they offer it, but they have the they have an app that's pretty capable. I'm sure they'll go in that direction. you have to imagine in their roadmap, they're thinking, okay, we've done rag. We hired Datagrid. We helped, they help with good agenc rag. Great. Now, persistent, long running agents that do things for you always, right? So like open AI's dot. They have to be thinking about that if they're trying to compete with the kind of harness. and they will, right? text your AI site super. I just don't know if it makes sense to paywall it. maybe give it to them for free. But Well index then then no Mike Powers: They can't though, they can't make money off it though. They need to make money. They need to they need to find another billion dollars in revenue. Luigi La Corte: Build them build on the usage. let the let the usage go parabolic and then build on the usage. Mike Powers: How you can't bill you can't bill a general contractor in usage. They don't have control over the capital to pay for it. Luigi La Corte: The general contractors are paying for Claude, right? And and as you saw in their S one Mike Powers: Right now. Luigi La Corte: or Anthropics S one, eighty percent of the revenue is coming from consumption, not subscription. Sorry? Clifton Harness: Pro probably not the API version, Are they paying for the API version or like this is the funny thing with with construction guys, is like, yeah, our our our company is on clot. And you go in I gave a presentation in one of these these big firms and and you know, come I come to find out at the end of the end of the presentation, it's it's actually one person at the at the business that has you know, a five X whatever account that that data mines, emails one time and then it it created so much value for for them 'cause it's the first time they ever did that. Like they they're not thinking in terms of systems. most businesses do not think in terms of systems 'cause they think in terms of of the people and the people are the system. And that really is the the major barrier I think to AI winning in the next five years is like how quickly can it's why like the C R the AI based CRM is such a great idea. Like it's the same exact tech that HubSpot has with an AI on it, right? Okay, well you gotta now bring that to market and you gotta change actual people in and how they how they work, right? Like I feel like entrepreneurs are very quick on our feet to change. and and it's much slower, you know, everybody else behind us. And simultaneously, I have experienced more change in business in 2026 than the previous like nine years before. So we don't even know where things are gonna be even six months from now, right? Like It could be that they've the open AI figures out chat in a way where it ta is texting people. They've already done the phone call thing. Why? Because that's for an older demographic. I'm sorry, I just get o get lit up by some of these insights and y'all are corrupting my thinking. It's great. Andrew Zukoski: Wasn't like we pick up we're picking on Procore a lot, but the PMIS systems are all implemented in the twenty teens. Wasn't the reason they worked that the organization wants a central view of say their risk exposure and change orders? Like they want a p a database they can go to that's like, here's all your stuff. Like and then through mobile phones you had the ability to collect information from the field, right? Like Mike Powers: Yeah, the the the Procore pitch was hey we're gonna I two you literally put internet on a job site as Andrew Zukoski: Yeah. Mike Powers: the first pitch, right? And then it was like we manage F RFIs, change orders and submittals. Andrew Zukoski: But connect the field to the office, right? Mike Powers: Plan grid's pitch was, hey, now that you have all these documents, instead of bringing out a roll of forty two inch drawings, you're gonna have it on your iPad. and then Autodesk obviously is designed and then they've like all merged into, hey, we need to become manager of submittals, change orders and RFIs they based on document changes is really what the tools do now. Andrew Zukoski: And a key thing is that you have a have an office like model of where the project is and then the ability to push updates out and like collect information back via mobile devices. Right. Mike Powers: Yeah, I would say Procord even today still doesn't do that successfully 'cause I there are instances of projects are not always matched. And what that means is like I'm a subcontractor that has a Pro Core subscription. I have a Pro Core instance. Andrew Zukoski: Yeah, sorry, okay, so they have a they have a view of a for a customer, I a hundred percent agree on the yeah, just that subcontractor's view or just that, but it's like the subcontractor's Mike Powers: for just that subcontractor and then it doesn't it's funky, they're messy. Andrew Zukoski: office theoretically can understand what's going on for them on the project site, right? Isn't texting or calling just another way to feed that information model? Mike Powers: I just don't think you need a UI of a yeah, I don't think you need the actual PM UI in the field. Andrew Zukoski: But doesn't why would the office not need a UI? the UI has been Mike Powers: The office. Hey, you could you could ask Claude or you could ask Instinct what are the open change orders I have right now? Andrew Zukoski: And it's gonna show me a list. Mike Powers: What what did what correct in a in in a format that you w 'cause the swing back is gonna be back from ten, twenty years ago we were trying to get people off email and Excel and we're gonna swing back to email and Excel. Andrew Zukoski: Cause I feel like the consistent UI, what's Mike Powers: Why wouldn't you? Andrew Zukoski: beautiful about it, is that multiple people understand it the same way. Clifton Harness: 'Cause you're y you're moving towards execution. Andrew Zukoski: it takes time for people to understand a presentation the same way. Mike Powers: But that's that's the b that's the beauty but that's the beauty of natural language. you can you can read the same data however you want, or you can read it in Chinese, Arabic, or the US. It doesn't matter. Andrew Zukoski: I don't like that though. I don't I don't want to Mike Powers: Okay. Andrew Zukoski: have me okay, show me show me a list of at-risk accounts, like and have it do whatever, you know, analysis based on my history and Claude and show me something that it's crafted to me and then have like our head of sales or something say like, okay, show me a list of at-risk accounts and have them be different. I want them to be the same list. I Mike Powers: But that's Andrew Zukoski: want it Mike Powers: but that's diff that's different. that you should it depending on how the harness that you've set up, it would report the same clients, same list. If you're just g get Luigi La Corte: A agents still need structured data. it's not like they can just deal with I I I do believe that they need struct they don't need it, need it. It's just gonna be very inefficient. You want the same consistency of reporting. Mike Powers: Right, but you don't need a s you don't need a UI for it. You don't need a hu you don't need a human U UI, you don't need to go on a website to do it. for our we're using Monaco for our CRM right now. No one signs into Luigi La Corte: Well I Yeah, is it good? Clifton Harness: Yeah. Mike Powers: Monaco. Yeah, I don't I don't I don't ever sign in, I just ask Claude. Luigi La Corte: Well, is it good does it give does it result in deals? Close deals. Are you getting close one? Mike Powers: result in deals always close. I'm is Andrew Zukoski: The deals getting closed, yeah. Mike Powers: a CRM has a CRM ever helped close a deal? No. so does it get the job done so I don't have to go into another application? Yes. Luigi La Corte: Okay. I'm still bullish on a PMIS. I don't think PMIS is go away. I don't think a good a good G C one that's growing and has let's say over a hundred million in annual revenue is gonna rely on Google Drive with C P to Muse or to instinct. I think they're still gonna benefit from the the structure of data that you get from a PMIS, whether they use it or not, maybe not. Mike Powers: But in that world, in that world though, McCart I that's why McCarthy has spun up its own instance of this. They have their own dashboard. They don't they don't need a massive Procore dash. They have their own dashboard. They made it. They made however they Luigi La Corte: I still use Procore. Mike Powers: want it to be. Andrew Zukoski: Procor McCarthy has some pretty chunky technology investments in accounting and financial and PMIS systems. Like Mike Powers: I know. I know, but yeah, but because it's not gonna be a switch, it's not just gonna be a turn off. But I think there's gonna be general contractors that are scaling right now that are gonna be able to take advantage of being headless and they're gonna be able to execute faster because of it. Luigi La Corte: It can be headless, but that doesn't mean that they don't need the structured data from a PMIS. I think those those aren't mutually exclusive. Mike Powers: I I yeah, I but again you always have structured data. I'm just saying you don't need a actual UI for a person to engage in the application. Luigi La Corte: You don't need a UI but you still need a product. I still I th I still think you're gonna buy a constructible or Procore or, whatever the competitors are. Andrew Zukoski: I don't know you don't end up needing to go in and interact with the data in some tuned UI. I think of Carta, we use Carta to manage our cap table, and literally years will It doesn't. Mostly I mostly I don't log in. Luigi La Corte: What is it what does a cap table need managing? Mike Powers: Have you used the MCP? Have you used the c have you have you used the car to MCP, by the way? Clifton Harness: Ha ha ha ha. Andrew Zukoski: I have not. but years go by where it's doing its thing and I don't have to log in, and then periodically I'm like, okay, I need to like like look at something and figure something out and explore something. And maybe I can do it through the MCP, maybe not. Mike Powers: Just just ask Claude. Just ask Claude. Claude's better. Cla Claude's better than the Carter UI. With the with with the Claude MCP, it's better. Andrew Zukoski: god. I the Carta UI, I don't know. Yeah. Okay, I might be digging a hole here. Mike Powers: That's what I'm telling you. you're like, where the hell is literally just connect to connect the MCP and just ask Andrew Zukoski: Yeah. Mike Powers: it to build out things that you want it to build out? It's gonna be ten times better than Cardo. Guaranteed. Andrew Zukoski: But it okay, take another one, HubSpot. Mostly I don't interact with HubSpot except for when I do. Like, and then I need to go in and like engage in the data in detail and I wanna be able engage with I wanna go look at Mike Powers: Like what? What does that mean? What what what do you g Andrew Zukoski: every interaction with a particular person over the course of years. I like I'm confused about like why we're assigning something to a particular stage and I wanna go step through like everything that happened and I wanna happen. a shared understanding with everyone around me. and yeah, I could have Claude build me a bespoke UI for it, but then I'm gonna have to train everyone who I want to understand it to use that bespoke UI. Or I could just use the HubSpot Mike Powers: In that example it's just like you Andrew Zukoski: UI that for all of its faults has like been tuned over years. Mike Powers: But no one's updating data, it's like all it's all reactive. I d I there's there's like so many like those static data it's a static database. It's a snapshot in time. Luigi La Corte: Let's let's bring it back to MCP really quickly with the Clifton Harness: Isn't every database static until you update it? Luigi La Corte: Sorry, Clifton, I was gonna ask you a question. So do you do you see a world where people are not interacting with their drawings, their 3D models, because it's been abstracted away by like an agent? Or are you bullish on no you gotta look at the you gotta look at the model, like you gotta look at the test fit 3D rendering because this is like your bread and butter? Clifton Harness: that is a really good question. Drawings I think are the reason why we're having a lot of this this discussion right now. if it wasn't for these shitty rasterized five hundred sheets of paper that are you know, were structured information and then turned into not structured information and now we gotta back and turning it back into structured information. we would like I think we'd be in a better spot, as an industry. Mike said that he died. Hope he's okay. Luigi La Corte: No, it's Andrew Zukoski: Yeah. Yeah. I'll I'll take that as a win in our argument for whether or not Luigi La Corte: it's it's it's not the bike. Andrew Zukoski: you need a UI. Looks thanks, Mike. Thanks for seeding the feel. Clifton Harness: okay, if if the construction documents were structured and on a you know, a persistently known graph for how to traverse a building and and its, you know, documents and specs as opposed to five hundred raster ed sheets of dumb information, then you could text the graph and you could say, Hey, what is my work that I need to do today? I'm you know, b I'm doing the framing on Phase one, part two, this part of the site, and then it could s send you the information that you need as a as a carpenter or whatever. like I I I think the we will get to a place where intent is minified, so that so that you can we like we over document buildings in in Rabbit simply because because we can. and that's a waste of everybody's time and yeah, like getting back to the minimal amount of documented design intent and then ideally delegated design for all of the other stuff, you know, you're no longer dealing with waiting for an you know, for the architect to to push the information in. And I think creating I don't know, living, breathing specs, you know, we allow these kinds of things on a range, n don't even need to RFI, but the specs need to have dependencies and they don't because they're not structured, and they're not tied into the drawings. Like, it's all a mess, because we're all looking at PDFs. and that's you know, that's probably the hairiest problem that someone should should go solve. Luigi La Corte: So you're long term bullish on people reviewing the drawings because that's just a constraint of the industry, if I understand what you're saying. Clifton Harness: I don't think like what's the difference between a compiler and an industry that doesn't have a compiler? Like Luigi La Corte: That's a huge difference. Clifton Harness: that's a huge difference. And if we don't create, you know, AEC compilers, then we will never get to a place where we're not looking at the drawings. Luigi La Corte: Will we ever? I don't I'll take I'll I'll say no. I'll say we'll never get there. It's a risk transfer problem. We've seen this over and over and over again. At least in lump sum. Maybe in CMAR or Andrew Zukoski: No, 'cause you there's too much that's unknown. you have the intent, but it like compiling it to a building, you the the the intent is under specified. there's so much that has to be worked out, like based on information that is only found out after the intent has been specified. Like market conditions change in a given material system spikes in lead time or price. Like you have to adapt to that. Like Clifton Harness: Yeah, your spec book would be a proprietary, performance based specs with dependencies and a range on what you allow based off of what's available in market. Instead of specifying one single thing that, you know, you gotta go buy and put in the product or in the project, like everything can be on a range with dependency graph. can't get these kind of windows, okay to hit lead we gotta juice the wall assembly by you know, this amount R value. Andrew Zukoski: Mm. Luigi La Corte: Maybe. in a way, GCs are just spending money now to compile it themselves. Like all there's like a plethora of like RFI generators who are trying to find issues or conflicts. And that's like a symptom of the problem that you're describing. Will an owner say, Okay, this new autodesk product, it helps make sure that the specs are perfect and the drawings are perfect? Yeah, I can see them doing that. Andrew Zukoski: the the space of possibilities Mike Powers: There's not perfect. There's no such thing as perfect. Andrew Zukoski: is just so large of things you have to react to. Luigi La Corte: What do you mean? You can have a million lines of code that are perfect that are com that compile. Andrew Zukoski: But the combinatorial explosion of a particular project and the weird shit that can happen on a project, like think of all the wacky shit that can happen from neighbors, like all the weird local politics, like all the weird local supply chain things. Like like every project is a snowflake, like r like really. Like how are you gonna Luigi La Corte: Not everything interacts though. Not everything interacts. Th there's a lot of interaction on site, obviously, but Clifton Harness: Ugh. Luigi La Corte: a conflict can be resolved at one point in three dimensions. And yeah, there could be a spillover point if it's like depending on the trade. we're at like if we're if we're kind of i ideating what could exist in the future, I would imagine like a building is completely designed in Revit. The spec is reconciled. There's checks against the, you know, if these products exist, if they're still current. there's all these RFI checkers, kind of like virus checkers in the background that are operating on your computer. Like I could see I could see that happening, it doesn't exist. Clifton Harness: Well, everything is a static model created at a static point in time. Just like Mike is saying, you know, Hubspot's a dumb database, it's static. Like, okay, Netflix, I think they have adversarial system that goes and kills parts of their stack randomly so that they know that it self heals itself back into back into production. you know, why can't we demand the same thing out of our our BIM software? you know hey you you moved the whole mass out of the way and now my energy calcs are different and now I need a different mechanical assistant. Like it's just we're we're manually encoding intent with hundreds of humans in every single way and that's gonna be the the thing that I'm hopeful AI wakes people up the fact AEC is an insurance product. It's not actually how an industry should work. Like, these are all different professions because one's got the design risk, one's got the execution risk, and one's got the finance risk. And that's why you end up with the documents the way they are, and that's why you end up with the adversarial culture that we have. And if you get the model to a place where it can perform for those three people to mitigate risk, then you're gonna have more profit margin and maybe AEC stops being an insurance product. Luigi La Corte: I think if you can have there was this news that came out very recently where Claude Sonnet 5, with access to the right tools, performed brain surgery on a mouse to treat its epilepsy. I have a hard time believing that it won't be possible to find issues on construction project if that's possible. I do believe Clifton Harness: Yeah. Luigi La Corte: in Dario's data center full of geniuses. it's not really a thesis, I guess it's just like a way of thinking about LLMs and If you have infinite smart people working on a building that costs you fifty bucks, I see no reason why you can't build a perfect building. Or at least design a perfect building. Mike Powers: I think the ki there's not I don't think there's such thing as a perfect building and and there's no right answer to any of it. That's the biggest challenge in construction. So for example So for exam Luigi La Corte: But you have an imperfect building. You have an imperfect building. Mike Powers: for example, we did the renovations in the Eagle Stadium and the reason we won the renovations is because we built the initial project. We were drilling micropiles in the parking lot at two in the morning because we didn't know where bedrock was. I had to scan all of the core concourses to figure out where the drainage was for the t for the rain drains. We still pierced every single drainage system. that th there's no reason it I I just having a perfect project doesn't make I don't think that it solves anything. It doesn't like like it Clifton Harness: ha having a having a data model mic that can continuously update itself even to the second as that core sample is being drilled is is kinda my point. Like As new information came in, okay, our footings need to be 50 feet deeper, and the hard cost pricing shoots up. The amount of time to complete the project extends significantly, and the first thing that you need to do is go figure out your financing, right? Like that's the dependency graph immediately, instead of okay, we gotta like figure out if we really can't find the bedrock for the next three months. getting the CRM idea you're talking about into construction is what we should be doing. Mike Powers: in that example, it did we already owned the pro we already owned the execution that we knew where Bedrock was. So it didn't matter if we couldn't find it. We were at risk of absorbing all that cost. Like the project had to open because there was a Taylor Swift concert coming July of twenty whatever it was, twenty fifteen. and the liquidity damages on that were like wild for whatever it is for a Taylor Swift concert. Luigi La Corte: Mm. Mike Powers: And as you like go through those pieces Andrew Zukoski: You just shake it off. Mike Powers: like as we go through those pieces Clifton Harness: Dude, that's that's millions of dollars to small business around. they did some some research on her impact on local markets. Mike Powers: It's insane. Yeah. But the so that first shot for me was like I read the spec book and read the drawings. And if I had an AI agent that just understood the spec and drawings, I wouldn't understand the project. And replicating that brain inside the person that's in the field is I think is a huge barrier that we're starting to lose by like these tools just read specs and drawings. and then the opposite side of that is like there's not because there's not a right answer. It is like a collaborative effort. the the architect wanted to put wings in the concourse that would like sway with the wind. And it just like did they designed it, sure. It just didn't make any sense. Like this steel the she the steel guy was like, What are you talking about? Like this the steel weighs this much. It doesn't it won't move in the wind. And like nothing like those types of things, like that's where like pain points come out in projects. It's not like, hey, like Yeah, I don't I don't know, it's it's interesting that's where I struggle because every project is so in product data centers. you're seeing Jacobs already do this. Jacobs is like we make cookie cutter data centers now. And in that world, I think it makes a lot of sense, right? I don't really care about how the building is designed. I just need the output of the building. Luigi La Corte: Mm. Mike Powers: that's modularization, that whole world, like that should be solved, like that is gonna be a solved problem. I just think of this other world where there's taste of a person that doesn't actually know like the owner building a hospital actually doesn't really know what the engineering architectural decisions are. So they hire an engineer architect. Those architects and engineers have different opinions amongst their team than the owner does. And then they're arguing over fascia and light blue paint on the wall. Clifton Harness: So aren't those people just gonna use cloud then? And that we said at the beginning they're just gonna hop into the AI and read all of the stuff they need to be successful? Mike Powers: I think in that example, right, you would have access to all the data like you'd have access through Claude to be able to go through that as you're looking at the wall, looking at the sample, which color do you want? Clifton Harness: I mean, look, I I I feel like half my mission here is to inform architects about the income in the building and the expense of the building so that they don't get laughed out of the investment committee room with a design that, you know, doesn't match either one of those things. and we're only gonna continue to make TestFit smarter and more deterministic around pricing and around income so that if they are using an AI to help, they're informed of all these information ahead of time. And you know, then you don't end up with, metal wings or something, right? Like shouldn't there have been a classifier inside Revit saying, Hey, this is a really stupid idea? Luigi La Corte: Not even a classifier. An agent is gonna be like, okay, well let me review this. Hold on, this metal detail is a bit odd. We shouldn't do this. Guys, let's just redesign this. Clifton Harness: This is a terrible idea. Mike Powers: The the owner the owner wanted Eagle Wings in the concourse. I don't know the Luigi La Corte: Ha ha ha. Mike Powers: owner, what do you want me to tell like? I don't know what to tell you. Right? Andrew Zukoski: I don't know what to tell you, yeah. Clifton Harness: So we gotta give the owner the design tools. Andrew Zukoski: Yeah. Luigi La Corte: Alright, let's let's call it there. Clifton, this has been awesome. Thanks so much for joining us. I think in summary MCP's MCP's Clifton Harness: hopefully I Luigi La Corte: great and you're gonna see it everywhere for the rest of your life. Clifton Harness: until the C L I comes out, which will be y'all should we should talk C L I's the next time. Luigi La Corte: That's right. CLI edition next. All right. Thanks, Clifton. Thanks, gentlemen. Talk to you guys soon. Yeah. Bye.