Luigi La Corte: But everyone, welcome to Construction Technology Podcast. I'm here with my co-host Andrew from Join and Mike from Build Vision. Gentlemen, how are you? Happy almost, Fourth of July. What's going on?
Mike Powers: Happy almost fourth July.
Luigi La Corte: So we are going to Fable a spin now that it's back in everyone's pocket. we thought it'd be cool to benchmark Fable ⁓ against projects that have already been estimated and answer or continue to answer the question of whether these models out of the box can be fairly proficient at quantity takeoffs and whether they can do scoping. So for the agenda today, both Mike and I have processed a project where we have a ground truth. ⁓ we're gonna pull it up on screen and we're gonna walk through the results. And so I think Mike has a project. I have a project ready to go. then Andrew is I just gonna ask questions about the process. ⁓ So let's started. So I'm happy to I'm happy to go first. let me let me share my screen and let's get going. So the TL the TODR with this analysis was Fable. let me just tee it up for the for the team. So I uploaded a project fairly complicated.
Andrew Zukoski: Thank
Luigi La Corte: ⁓ manufacturing plant retrofit. this project was ⁓ I around three hundred million capex. and Fable actually did a good job. It did a good job on scopes. It did a poor job on quantities. And I'm gonna I'm gonna increase the size here and we'll be able to see what what they did here. So Let me just decrease this so we can kind of review this a bit. Okay, great. So this is ⁓ like I said, industrial manufacturing plant. So the headline is on scope versus ground truth, we have an F1, which is ⁓ like a data science term which balances ⁓ positives and negatives, false and true positives and false and true negatives against the ground truth. ⁓ and it scored ⁓ almost 90.90 on the scope identification. But a materially poor quantity takeoff at one. and I wanna walk through these obviously and go through go through ⁓ some of the specifics. But I will say out of the box, Fable is good at scoping. ⁓ the cross-document scope analysis is strong and it's only getting better. When we reviewed 4.7 against the same project, it was it was scoring 0.69. ⁓ so I think in general, scoping. is just gonna get easier with these models. Mike, I don't know if you saw anything similar.
Mike Powers: Yeah, and I think like like the zoom out of this is like like construction documents already are not a hundred percent accurate. S there's always gonna be scope gaps, and obviously there's an entire industry around doing scope gap analysis, there's an entire industry around doing document analysis, there's an entire industry around doing takeoff analysis. ⁓ but I don't even think being a hundred percent is the end goal. I think it's all about like helping with the efficiency of like getting you to review as an estimator or as a pre-con manager the pieces that need your review. Because in construction there's not a right answer. This isn't like, hey, ⁓ I'm going to buy like, you know, e-commerce widgets and I need to buy a widget and I know exactly the dimensions, the exact the output that I need. Like there's this flux in construction, and that's why general contractors are where they are. So I think the point and my position around this is like it's never gonna be a hundred percent, but doesn't need to be a hundred percent. ⁓ relied too much on construction documents already, does these tools actually push you more upstream to actually be a part of the design process more effectively to have these conversations earlier? And I think Fable does a very, very good job of getting you to that earlier conversation with with the engineer and architecture firm.
Luigi La Corte: Mm. I I'll disagree. I do think when you say we're overly reliant on drawings, I mean like that's just the pro that's just the process. Right? Like people have
Mike Powers: Doesn't need be the process though, because if you look at like the EU, it's not the process. There's quantity takeoff like at like the on a document set in the EU, they would have the quantities. Like why I is like I'm like befuddled that like an engineer or an architect puts, you know, 17 doors on floor one of a building, compresses it into a two D position, and then makes it very, very hard for someone to count those doors when in reality that's not the first principle problem.
Luigi La Corte: Or what what's the process?
Andrew Zukoski: Thank
Mike Powers: The owner should be pushing off against the architect and engineer saying, hey, instead of paying, you know, $10,000 for a takeoff or scope gap analysis, like this should be almost like an open source data structure that you can read and engage with however you want. And then by forcing it off of documents into an actual computer-red language, Fable.
Luigi La Corte: Mm.
Mike Powers: OpenAI, it's gonna be able to read those a lot more effectively because it doesn't need to be a human reading those those endpoints. Like why like this again, this is why like I'd be bullish on Autodesk. Like an autodesk can open up those nodes, ⁓ and you can engage on a front end with a chat interface. ⁓ it solves a lot of these problems that were that don't necess necessarily need to be solved the way they're being solved now. And that's like reading documents isn't the solution.
Luigi La Corte: I I also wish that were that I also wish that were true. I mean, selfishly, I'm happy it's not true because then it it allows me to build a business. ⁓ so it is somewhat selfish for me to say that. ⁓ but it would eliminate a lot of waste. And ⁓ it probably would bring the owner and the GC ⁓ in a in alignment more than they are now, or versus it's like it's very adversarial. But irrespective, I mean like the the industry just doesn't it doesn't work that way here for various reasons. And like there's no own there's no owner incentive to share that information. Because they know it's just gonna be used against them. And your point, the design is just wrong. so the more they share, I think the worse it'd be for them when they get caught in like a change order. Now I was like, I'm curious to get Andrew's perspective on this as somebody who helps manage VE's.
Andrew Zukoski: Yeah. and all of pre-con. ⁓ I don't know that if you look at Europe, builders substantially engage earlier. think on many projects they do, but you have this huge, like the consultant layer in the value stack is thicker in Europe. Like a lot of quantity surveyors work for these, are the people administering these quantities on the project, Like work for... consulting firms who are hired by the owners and owners are out there for Turner and Townsend or Mace or whoever, right? And the builder and the trades get brought in, I think there's design, bid, build work there actually for the true prime contractor. And they actually get brought in later. because the consultants and cost consultants handle a lot of the like early engagement and construction feedback during the design process. So it's sort of interesting. I don't know if it brings the GCs in earlier. Having it be a point of fact would be good. It seems like there should be some sort of open source repository where it's like, okay, what's in this drawing set? Like I got it from a plan room. What have other people found in it? It seems like this doesn't need to be carefully guarded secret knowledge anymore. ⁓ But Luigi.
Luigi La Corte: Yeah, I I agree, but Yeah, go ahead.
Andrew Zukoski: I was going change topic. Go ahead.
Luigi La Corte: I was just gonna ask, like, I think there is a a ⁓ a process, like there is val it's value add when a general contractor goes through a set of documents and decides how to scope a project. Because I think it does raise a lot of questions that could lead to to V E's ⁓ or to design changes or like variation notices that they that they think will benefit the product the the project. So just to say, hey, we could just solve this whole process and just give you the answer.
Andrew Zukoski: ⁓ yeah.
Luigi La Corte: ⁓ I do like I think there's a lot of value when when the G C goes through that process now and scopes the job themselves. That's why I also am very bearish on like one click scoping.
Andrew Zukoski: Yeah, I would agree on that. I feel like many of our most pre-con forward, not many, but several of our most pre-con forward partners, it's like who emphasize early engagement, who are doing complex work, where they're hand in hand with the design team. They talk about how at its root, the highest value add that pre-con offers is getting... the field team, the operations team spun up and engaged earlier on the project. They provide a mechanism for experienced superintendents and project managers who are going to run the job to engage early and start to understand it earlier and start to provide their own feedback earlier. And that it's all about bringing site knowledge early. And I feel like that's one of the things that would fall out of a scope analysis is like hey, this is how our team builds, or if it's like the actual team providing a little bit of input, like this is how this is gonna get built. And that can bring immense value to the owner and to the design team.
Luigi La Corte: Yeah, I I agree. ⁓ okay, let's let's continue going through this and ⁓ yeah, go ahead. Yeah.
Andrew Zukoski: Hey, Keith, can you scroll up, actually? So the little note there, the read this first thing, ⁓ is this true? Like this thing where it says that the asymmetry drives nearly all of the quantification gap below it? Like it looks like somehow some PDS didn't make it into the analysis. Like what happened with this?
Luigi La Corte: Sorry. Drawing PDFs were absent for the folder and the flag rather than guessing. It's should you know the quantification gap below barely affects scope identification. ⁓ let me check the folder and let's see. Copy remote IFC architectural. So the documents that I gave ⁓ that I gave Claude were the same documents that we gave to the third party for estimating. So it may it may be true that it thinks that that's the case, but we still got quantities from the other documents that we used.
Andrew Zukoski: Okay. Okay, so this is, this is, it is incorrectly identifying the cause of quantification data rise.
Luigi La Corte: Yeah, the the no the the this is the same the same set. So when we build our benchmarks, we take we take one set, we give it to human estimators, either in-house or ⁓ out out house, ⁓ additional additional resources. ⁓ and then we'll do the same thing with, you know, set like our our model, and then we'll do the the same thing with like claude LMs just to build our benchmark. So it's saying that
Andrew Zukoski: I gotta say though, we gotta mark the day on the calendar. Like today's the day that an AI, you know, blamed a poor document set for like not being able to quantify a project.
Mike Powers: Ha ha ha.
Luigi La Corte: Yeah, that is that is true. ⁓ okay, let's continue here. So let's go through some of these items here. Okay, so F1 by trade. ⁓ so concrete, so like this is one of the items here that I think is interesting. ⁓ the F1 is for concrete is zero, but there's you have so you can do this takeoff from an architectural set. And the architectural set is provided. It's a fairly large set of documents, like 250 pages here. ⁓ so like the thing I I had prompted it afterwards to ask it why it got these things wrong. And obviously, like it's what it's what you would understand. It's hey, I'm reading the quantities from tables. ⁓ I don't have ⁓ calibrated area takeoff tools, I can't do lines or you know linear lengths, and I I certainly can't do volumes, which is what is what concrete is. ⁓ so it's really guessing at these, but at the same time, it also got some of them right. ⁓ it got them right, obviously, because If you go to overhead and special special doors, like it's a table and I can sum those tables, it's not gonna give me perfect doors. So I'm I'm surprised that it got me an F one of one, which is not easy.
Andrew Zukoski: Hey Luigi, for folks who are listening who aren't familiar with F1, you know, is this McLaren versus Ferrari or is F1 something else?
Luigi La Corte: Think of it as as accuracy. Consider it as as like a 100% accurate, consider it would be it would be like an F1 of one. ⁓ and then an F1 of 50, it's like, you know, it's a it's a coin toss. ⁓ but you use F1 when you have an unbalanced data set. And so ⁓ you know, think of it this way. If I were analyzing a contract, ⁓ and there's one, there's one liquidated damage clause, but 99 non-liquidated damages clause, ⁓ could ask me what's your accuracy and I'd say well I found all the the non-liquidated damages clause except for that one. So I have a 99% accuracy when really ⁓ you failed. You you missed you missed the the thing that you were supposed to find, which was the LD clause. So when you have those unbalanced data same thing like in an MRI, like did are we finding ⁓ are we finding a like a cancer nodule or not? ⁓ even though 99% of the images are are not cancer nodules. So you you're using F1 when you have an unbalanced data set. And you normalize ⁓ by the ⁓ the expected amount of false positives ⁓ or sorry true positives. So ⁓ it just helps you understand what the accuracy is ⁓ without skewing that would be favorable towards you know somebody who has a bias. It tries to give you like, hey, zero percent F one means like we basically found nothing. ⁓ get like relative to that example that I gave you on ⁓ on on liquidated damages. I mean like the point, the point. Yeah, go ahead, Mike.
Mike Powers: So like the but I'm gonna push back like concrete like Like you're like you're not that there's not a there's not a defined quantity of concrete that's needed on a job. Like you're delivering trucks that are coming to fulfill the project need. And and you're like you're within a range. Like you're within a range. You're not like, ⁓ I have ten thousand ⁓ you know, vol like cubic feet of concrete for this project and it needs to be a ten thousand or nothing. Like there's there's slop, there's things that fall off the truck. Like I just I don't think being a hundred percent on concrete even matters. Like the concrete bitter is probably gonna know based on the size of the project how much concrete they need. Just by this linear square foot. They don't need do a full concrete takeoff. Cause my bet would be that the architect did not even put all the details to include the things that need concrete, like a curb for an air conditioning unit on the roof. So I think like
Andrew Zukoski: But for the big poor, like, you're gonna get...
Mike Powers: But that's like linear let's that that's square feet on a project. You're like, hey, the project's a hundred feet by a hundred feet and it's a hundred stories. You're not going down to like the minuscule square foot where you need a hundred percent take off of concrete ever.
Andrew Zukoski: That's true. But that is like the first quantity that our customers bring up when you talk about bin based takeoff. And so I think that that's, I think that cubic yards of concrete is like a, a key number. ⁓ it is cubic yards. I'm like, yeah, you, you don't buy it that way. You buy, you buy a certain number of trucks and you keep them coming until the thing is full. Like.
Mike Powers: But I I just I just don't think I would push back that. And I guarantee if you pulled ten general contractors, they're not like relying on an accuracy count of quantity from their concrete sub to within like five percent to say, ⁓ that's a that's the exact amount of concrete that
Luigi La Corte: I think I think they are. I mean like w
Andrew Zukoski: I think they are, and they're often self-performing concrete, at least are. Go ahead, Luigi.
Luigi La Corte: Well, I was gonna well, I I was only gonna agree with you. I mean, they they do this exercise now because they wanna see if their subs are in the ballpark or if they're missing something. So I think G Cs have a lower level of resolution that they're targeting with with quantities. But
Mike Powers: But that but that's the point I I would say like the that like I would bet, I would assume that saying, hey, give me if you asked Claude, give me the estimated cubic yards of concrete needed for this project, it will be within ten percent. And that's all that matters. It doesn't need to be within ninety-nine percent for concrete.
Andrew Zukoski: Hmm.
Luigi La Corte: Yeah, I would agree. I I don't think for the G C it needs to be within ninety nine percent. I agree.
Andrew Zukoski: ⁓ my god.
Mike Powers: So then so then it doesn't mat like why even like spin up why even have a specific takeoff tool that does concrete to that nine nine percent when the nine nine percent doesn't
Luigi La Corte: I don't think I don't think you need a specific tool for for quantities. But I do think the like the biggest drawback that you have right now, the reason why you don't see people doing it, at least at large firms, is ⁓ there's the there's almost no auditability. There's no way of knowing whether it's right. Versus like if the human goes to draw all the polygons, you know, they can refer back to those polygons. And ⁓ Claw just doesn't have the tool use to to ⁓ to draw those polygons yet. Like maybe that'll change. So I think the point here is like it actually does a decent job. Like it obviously does a good job on scoping. ⁓ and it I I I mean, what what we what were your results, Mike, for for quantities?
Mike Powers: So for quantities it it's like again it's within it's the it's it's within plus or minus five percent.
Luigi La Corte: Across every trade.
Mike Powers: Across the trades I think you need to do quantity takeoffs on. Doors, windows, mechanical equipment, electrical equipment, things like carpet, things like concrete don't matter. Things like the amount of paint that goes on a wall. Like that that painting sub is cat like they're not saying, ⁓ we're gonna buy exactly ten hundred gallon tubs of paint. They're buying whatever 'cause there's gonna be need to be touch ups and like when the guy scrapes the wall, like out of the same paint, like
Luigi La Corte: Like
Mike Powers: I just like don't like so there's there's this like bifurcation of trades that matter and things that you have to like order and you have to you can't get from Home Depot or your local distributor need to have accurate accounts. So doors framed and hardware need to have accurate accounts because you can't just go get the 20, the 36 inch wide door down at Home Depot the day of the project. Things like an air handling unit need to have accurate accounts because you can't get because there's all D time on it. But these things that like are commoditized doesn't matter.
Andrew Zukoski: Is that because you bid it out to the trades and the trades are going to do a detailed takeoff?
Mike Powers: Trades own it. The trades own it. Like the like the like I'm like I I just like I think you get in the ballpark with Claude, you can level out the quotes and you can compare your subcontract. You get three concrete quotes from your three concrete subcontractors if you're self-performing and you're in the ballpark.
Luigi La Corte: How how can you bid level without knowing what the subcontractor amount should be?
Mike Powers: God directors do all the time.
Andrew Zukoski: Don't you always, you always, you have four, if have three bids in your own number, the assumption, like you're gonna, your number is missing things too. Like, look at this, like apparently, apparently you missed an elevator. right, that happened, that just, I mean, that's what we hear is like when you're leveling, it's like, we got four data points and I don't know that. I probably trust any one trade number more than I trust my own number.
Luigi La Corte: But then they make mistakes.
Mike Powers: Yeah. Forget the curve. Forgot the curve.
Luigi La Corte: Well right.
Mike Powers: I just I just don't think you can tr like I think the whole point is that you can't trust the drawings enough, you can't trust the documents enough to make a hundred percent assessment on a takeoff. Ever.
Luigi La Corte: But I I think that's your only recourse, right? Because once you enter into the contract, I'm I'm building per the drawings. Like that's that's a it's an exhibit in your in your contract. And if it deviates from that, like I can you know, I'm gonna go after a variation notice.
Mike Powers: You have intent though. But it's you still have intent as a contractor. But you stole an intent as a contract. For concrete, for a concrete pour, you're owned, you own intent. You own that you're probably gonna do engineering slabs, like you're gonna own the engineer system realmant. It's the same with structural steel. You own the structural steel. The majority of the time, the structural steel provider is owning the shop drawings. They're owning like like the thickness, the bull like the anchors, the welding, like they're owning all those pieces. So I don't think that I don't think the I don't think the structural engineer, the structural engineer is gonna sign off. Right, after a submittal, after shop drawings are submitted through the out through the process. But that document that comes out that you're doing a takeoff on.
Luigi La Corte: Mm.
Mike Powers: Like the there's there's so many mis there's so many missing pieces.
Luigi La Corte: But not in a ⁓ in in a design build, yes. But if I'm supposed to just be building per the drawings, the owner and the designer own the quality of those documents.
Mike Powers: Not all the time. Like so I was at the the first job I was on was in the Eagle Stadium, Philadelphia. And the architect wanted to hang ⁓ wings that moved in the concourse. That was the the the contractually on the drawings, there was these artificial wings that were in that were in the concourse that would move with the wind. ⁓ that was spit out to the structural steel subcontractor, W and W steel.
Luigi La Corte: Okay. Yeah.
Mike Powers: ⁓ they could not produce those wings. Because it was technologic it was engineeringly impossible to be it we able to withstand the engineering requirements of being outside plus moving in the Like that that that that there's nothing on the drawings that's gonna tell you differently of that working. And the structural steel guy it wa it was on the drawings, but it wasn't incorrect it was incorrect on the drawings. It was it was technologically impossible to make the way that the drawings were laid out. So
Luigi La Corte: It was on the drawing it was on the drawings though. It was on the drawings. But that's different than saying, Okay, the intent was I had three elevators. You ought to have known that G C versus like, Hey, I only see one elevator bank here. That's different.
Mike Powers: I that's a that's a hole. That's a that's just like the whoever laid out the elevator tell the laid out the elevators wrong. Like I'm gonna say I think that there's so much fuzziness in this world unless it's things that you need to that that you have to like, I need a widget. I need a I need a I need something that pushes air, I need a fan in in the ceiling. That's a quantity takeoff that can be done. And Claude does a very, very good job of saying there's a fan in the ceiling here on the drawing. As soon as you get to conduit wiring.
Luigi La Corte: Mm. Yes, I agree with you.
Mike Powers: Like that's never gonna be accurate. You can never you're never gonna accurately get something off of a two D document that's built in the 3D world. Ever. Never.
Andrew Zukoski: How, how, like, so take the GC, let's accept for a second, I'm trying to understand this, like, except that the GC says, okay, hey, quantity is never gonna be more than plus minus 5%, whatever, can't trust the quantities. If I'm a trade, like, I have to bid the project, I have more exposure to self-perform groups inside of GCs than I do to like strict trade contractors. But I know that they are obsessed with a level of precision that is much higher than GCs have, both for quantities that you would order, but also for quantities that you produce, right? Like they're tracking labor rates. They like close the loop between their production labor rates and their estimating labor rates. If you're installing a, if a crew can install a certain number of units of something per day, like you're doing your resource planning based on how long they're going to be on site. Like, and there's a lot of fuzziness there, but like at some point you do have to be on site or not. Are you planning for 10 weeks or are you planning for 12 weeks? Cause like the job will work at 10, but it won't work at 12. Like what's like, like trade, trade margins are higher. than GCs and it varies, like, generally substantially higher, but they're not like that much higher. You have to.
Luigi La Corte: I think they're gonna wanna be as correct as possible. But I mean if the if the question is can clod do quantities, the answer is yes, it can. You just you you'd have to have confidence that it's gonna be like I think for your quantity type, it's probably more doable. It depending on the quantity type, it's more doable or or less doable. ⁓ like maybe not for concrete volumes, but definitely for doors. Great. You can like you can use clod. In which case, like the there's no real there's no real debate. Like it it can do it. But There's different question is whether or not it works within your organization.
Mike Powers: Yeah, and just like and going back to the subcon tracking position, like mean this is going back to like that clawed article of like, hey, there in commercial construction, there's like 20% upside for AI LLMs because that's like the manual process of things that they're doing on a computer. But to back to Andrew's point of like, hey, I'm running ⁓ I can run hundred linear feet of conduit conduit a day per team. That has nothing to do with anything in LLM world. How do you do 200 linear feet? ⁓ I have to have two teams. there's not a robot that's like like laying conduit. It's labor ⁓ or prevailing wage labor. Like that's that that's a mechanism that nothing around AI is gonna touch for the foreseeable future because LLMs aren't gonna solve physical world problems like some of the physical world models that are coming out. So I think like that point of like, hey, the twenty percent of my work as a subcontractor needs to be done, can live with that being or minus ten percent accurate on quantities, the whatever you want to do? I think the answer is yes. If that helps you get a a response faster, it's helpful. Or if it helps you like find the gaps. Like it has like an estimating assistant that's sitting there, like, hey, I'm gonna go through all the documents as this estimator because I own the documents as a mechanical estimator. All right, I'm gonna have a bot or an agent go through and like also do the same kind of process that I'm doing. as a mechanical estimator that's been in the space for 20 years knows this mechanical engineer never puts fire on the drawings through the penetrations. All right, so every time I have a slab penetration, there's there's mechanical, there's conduit that's running through the floor. I need to add ⁓ dollar fifty per penetration. wait, but the drawings show just line of this conduit that's running through for the electrical or for the mechanical piping, the sprinkler lines. I know there needs to be three. Why is there three? Well, I knew like here's the reason why there needs to be three because of this code piece. And I think like Until those pieces, until that mechanical subcontractor is like embedded with that mechanical engineer and they're both incentive the both incentives are aligned to make the document as accurate as possible, doesn't there's no reason to solve for the document issues at Because they're gonna be solved. There's no incentive for mechanical engineer to put fire stopping on every penetration through the slab. ⁓
Andrew Zukoski: I mean, it like it can be straightforwardly inferred.
Mike Powers: Yeah, the yeah, like the per the the mechanical the mechanical subcontractor is inferring that they need to have that there. if you're reading the documentation, the documentation's not a hundred percent accurate about that. Or the spec says that you need fire stopping through the penetration, but the drawings are never gonna show that. ⁓ then the then the engineer has no incentive at all because they're billing by the hour. They're already squeezed by margin to sure that's accurate.
Andrew Zukoski: ⁓ okay, so you're Okay. So you're saying a literalist interpretation drawings, like, well, not that miles to go before that. can literally count the stuff on the drawing. Okay. ⁓ That's I feel like that's different from saying you can take a drawing set and use given just a drawing set, come up with.
Mike Powers: Correct. matter. Doesn't matter.
Andrew Zukoski: accurate quantities for what's necessary by applying knowledge and understanding.
Mike Powers: For the thing for the things that are accurately that need to be quantified, like fans or windows, where there's an actual like someone's actually doing this for like for windows, you're doing lighting, whatever you like that is yes, that's solved. For the things that don't matter, you don't need to have ⁓ windows on the wall or three light fixtures in the ceiling here. So paint, drywall, concrete, that doesn't So it's it's it's not a it's there's no reason to be a hundred percent accurate on those pieces. So you're you're bifurcating the the processes and you're like verticalizing a supply chain where things can be verticalized. the things that can be verticalized are things that are engineered, things that you need to like have a system design around, things that need to be like produced in Mexico and shipped to the United States because you gotta ⁓ chain manage that. You gotta make sure that that light fixture is there. But it's drywall or paint or concrete, like being within 10% doesn't matter. Like you're it's it's not gonna make or break the project.
Luigi La Corte: To who? Doesn't matter to who. Not to the the sub that's doing the partitioning.
Mike Powers: To anybody. To anybody. Like the painting the painting is not making or breaking the project because they had to buy ten percent extra paint.
Luigi La Corte: It's not like they have it's not like they have like twenty five percent margin.
Andrew Zukoski: Might be if they have to install 10 % more paint. Like what's a typical painting, painting, I mean guess you can recoup it through like through a change order to a certain extent like hey we.
Mike Powers: It's brut it's like brutal. Is that that painting subcontractor though is just betting on the fact that they can paint faster and better than their competitor.
Luigi La Corte: Definitely do it definitely does matter.
Mike Powers: And it's the manpower piece, it's not the takeoff piece.
Luigi La Corte: Well, I mean like do end up multiplying the square f like the square footage by a labor productivity rate. I just to think to think that to say it doesn't matter is like what I what I disagree with because they spend a lot of time making sure it's as accurate as possible. And if it didn't matter, why not just take the area kinda grow like grow square feet and apply a number and just be like, we'll see what it is. It doesn't really matter. Whereas reality, like they spend a lot of time going through to make sure it's as accurate as possible. So if it doesn't matter, why do they do that?
Mike Powers: I don't think painting subcontractors do.
Luigi La Corte: Say it again.
Mike Powers: I don't think painting subcontractors do.
Andrew Zukoski: Drywall, drywall on concrete definitely do.
Mike Powers: I just think there's so many there's so many dependencies on a project like that you're like already putting 10% extra drywall on a project because the comes sweeps by and and ⁓ and half of it. And that's not a change order. That's just kind of fixed on a punch list. I I feel like it's like it's like you're you're working toward it's not like self-driving where you need to be 99.99% accurate to save somebody's life. So if you can get to 90% accurate, all my position is if you can get to 90% accuracy by launching it through Fable, is good enough for the market. Because you don't need to be 99.99% and those additional nines that you're trying to drive for are just increasing the cost to provide the ROI that's needed.
Andrew Zukoski: That's interesting, right, Lake?
Mike Powers: Like if you're a painting subcontract paying a hundred thousand dollars a year on a takeoff tool, doesn't make sense.
Luigi La Corte: If it if if you spend ten last point on that, if you spend ten grand, if you spend ten grand for a takeoff tool, okay, it saves you from having to overspend relative to your budget on projects. Let's say your average contract value is a million dollars. That means you're spending an extra hundred thousand dollars ⁓ ⁓ on material, it more than pays for itself on one.
Mike Powers: I would say the controlling of your items on a project your process for QAQC is of a variable than the takeoff that you've done and the portion.
Andrew Zukoski: like if the infield QAQC. for some trade share.
Luigi La Corte: Help me ⁓ help help me help me understand that. So you're saying what actually happens on site is a variable that I can't control. And so it doesn't matter whether I hit but the thing is like it could hit me twice.
Mike Powers: No, it the the QAQC is what you can control on a job site by how you manage your team. We were gonna stick with a painting concrete contractor as an example. are you managing that process on a job site, which is the majority of where they're gonna lose money as to saying like, hey, ⁓ I you know, I bought a hundred of paint as opposed to a hundred and ten gallons of paint. So that quality, QAQC process is more important to that subcontractor than the initial estimate takeoff.
Luigi La Corte: But then I get hit both ways. I'm both short on quantities and I'm low on then I'm also hit on QAQC versus like I could only be low on QAQC unless I'm missing something.
Mike Powers: I just think I don't I I they so again, like, hey, we're doing the estimate, we're doing a takeoff. ninety percent accurate on my takeoff for paint by using whoever whatever. That's checkbox. Being ninety-nine percent accurate on the takeoff at that point is not balancing the profitability of you once you do the installation as a contractor.
Luigi La Corte: So what do you think is if so let's let's say that were true. what do you think is stopping subs from just doing their quantities in in cloud right now?
Mike Powers: I think it's education. I think some are. Some c some contractors are. think it's just like why like there there's a fear of like, hey, what actually happens when I upload my doll documents to Claude? Like who owns data? getting approval on twenty dollar Claude subscription is actually a hard thing for subcontractors right now, unless you have someone in IT that's like educated in that model. ⁓ we've talked to contractors still have on-premise email servers. Like and and the idea of like buying an AI that they don't really understand is is a huge hurdle. So that's my my entire pitch is that you should always you should like you should try it in Claude. Whatever you're trying to do, set up a skill, try it in Claude. you get ninety percent of the the accuracy that you need, I that's good enough for a
Luigi La Corte: You it'll
Andrew Zukoski: It's, it's this like 90 % number, like I'm very sympathetic to the argument that you don't need perfect accuracy, especially in QTO. that makes a ton of sense to me, but the 90 % feels low, man. Like if it's, if the arrow bar is greater than your median, like gross margin on a project, like something about that gives me the heebie.
Mike Powers: W well you're not just launching it. You're not just launching it. You're not saying, hey, send it to Claude and then just send the quote back to the general contractor. still having human in a loop, so you're supplementing it. How is your estimator? Is your estimator over 90% accurate right now?
Luigi La Corte: I would say most estimators at subs are probably at around like ninety-eight percent accuracy. That's like my vibe.
Andrew Zukoski: greatest. Yeah. They only do one scope, they do it day in, day out. They've got tight connections to people performing a work.
Mike Powers: Ninety percent accuracy on a document set that's not accurate though.
Andrew Zukoski: Yeah, because they've got all the rules of inference that you described earlier. They live that all day. I mean, it's interesting to like, you guys remember when Zillow did auto buying of homes, right? They built a model that could automatically price a home in a given thing. it like, they lost a ton of money. And the estimates were probably actually pretty good.
Luigi La Corte: But it doesn't like the accuracy component you can carve out.
Mike Powers: Yeah. There's
Luigi La Corte: They lost so much money.
Andrew Zukoski: The trick was they set themselves up for like rampant adverse selection where everyone who was in the house knew more about the house than Zillow ever could. And so if they thought they could do better than Zillow, they didn't sell on Zillow. And if they thought Zillow was better and they were going to get in the market because there was disgusting black mold infestation that was going to come out immediately when someone walked in the door, like they sold to Zillow and Zillow got fucked. Right. And so you like, you like you get If you drive 90 % accuracy and you're 10 % high, you don't win those jobs. They don't contribute to your margin. You only win the ones where you come in low. There's this selection that means that the error is accentuated in projects that actually win in a way that is not to your benefit. And that makes me wonder if that 90 % is like a man that feels low.
Mike Powers: But you but you're not you're not just like right like so hey all right I'm going back to the painting example. getting a budget number to my general contractor using Claude. Hey, I'm gonna launch this document send it to Claude, ask it to do a painting takeoff. All right, here you go. Here's my here's my 90% budget number. you as an estimator are still going through process to like invalidate those pieces. Claude does a great job of giving you the steps, giving you the data, giving you, like you can set up a skill to do this and you can find like you're exploring the gaps. Like you're looking at, you're having this like a sit again, it's like this assistant that's helping you for 20 bucks a month. And like that, like it's never gonna be you're you know there's no reason to be 100% because you're never gonna like there's no reason to be 100% because the documents that are never 100% accurate.
Andrew Zukoski: Yeah.
Luigi La Corte: That but that argument though doesn't hold in my opinion doesn't hold water because if I this is a lump sum contract and the owner has responsibility for design errors, I'm accepting those documents as is. If there's an error with the documents, I have recourse. could sue you. If you're not gonna give me substantial performance because you I I have to put in this elevator bank that was never accounted for, that's a that's a problem. So yes, the like the drawings are never completely accurate, but it's still like that's my contractual basis. If this were to go to court, they'd say, okay, it's an app it's exhibit A in the contract. There's clearly 200 200,000 square feet. The owner's asking for another 40,000 square feet that's never shown on the drawing. It's like I actually have legal standing to sue you for it. So I think that's what they end up relying on. But if we were gonna say Claude is gonna be an additional component, like a check, okay, well then then what we're not saying is it's gonna replace estimators or it's gonna replace estimating tools because estimators still need to go perform that takeoff themselves.
Mike Powers: 100% you own intent. You're never gonna offload that unless that unless the subc unless the software provider is absorbing all liabilities, you be a subcontractor and say, I'm just gonna like rely on the quantities of the software, which means you're gonna have to human verify everything that you're doing on a takeoff. You're gonna have to trust but verify the application. So that delta of ninety percent accuracy to 99% accuracy. Doesn't actually matter. And there's like the there's the psychological trick. Like when you go to an Apple store, for example, they have their computers at whatever degree, so that you go grab the computer and bring it up 'cause you feel ownership. Is it actually better? for a user to have a 90% accurate tool because then you're feeling ownership of the documentation by going through and validating all the different nodes that were inaccurate being 99% accurate because that 1% is going to screw you over because you're not trusting and verifying the process or the ⁓ the the process.
Luigi La Corte: Then what we're describing is a tool that's not as good as their current their their current practice. Versus like, okay, let's let's take let's take like a like a like a company that I think is like is is is really good. So like Bob Yard, for example. They started with like a single trade landscaping. when do takeoffs, you know, don't just do like counts, linear area, they break down into different types. ⁓
Mike Powers: Is
Luigi La Corte: So there's symbols, there's hatched areas, there's like r irregular polygons, et cetera, right? And so like they'll put that on the screen for you and you can just click. My so people are people are using it because it's very fast, but because it it maps to their current process. And at the end of the day, the estimator is still responsible. If you look at what Claude can do relative to that, I would say it fails in situ because it doesn't meet them where they are, but it's a good check. So but like, okay, so then if the question is like, can can Claude do quantities? The answer is yes. Can it do quantities the way subs currently do them? I would say no. It doesn't, it doesn't pass that test. So I like I'm st I'm still long, I'm still long quantity software because still need a human to verify, but unless Claude can just spin that up and like make a model on the fly that allows you to verify these accuracies. Sorry, ⁓ ver verify the accuracy. Like I I don't think that's gonna change.
Andrew Zukoski: Gents, I want to take a sideways for one minute. You talked about how education was a big barrier to trying this out. Can we get a tour? Do you guys want to do a tour of the setups that EG to drive quantities and scopes from Fable? What it like? What's the skill word? What's the...
Luigi La Corte: Yeah. Mike, why don't you go? I can show mine right after.
Mike Powers: Yeah. So anytime you do anything in Claude that you're doing a repetitive task. So there's two different systems in Claude that are powerful. One is a project. A project has an end date, so it closes out. So it's something that's like short term. The other tool is what's called a skill creator. And anytime you do anything in Claude that you're trying to do a repetitive task, for example, like a quantity takeoff, you want to ask the skill creator, so if he's backslash skill creator, ⁓ to build a skill, like say build a skill to do a quantity takeoff of a mechanical set of drawings. So this is going to take time to do, ⁓ but I'm running this in Fable Max. ⁓ and then you can call that skill to do a task. ⁓ so if I go over to this tab, it's actually processing this document. It's going through and checking all the different drawings. I hope that probably doesn't help. ⁓ so this is an example of what, oops, if I can zoom in. ⁓ what Claude is doing in the background.
Andrew Zukoski: I don't see any tab.
Mike Powers: So this is a quantity takeoff. It's going in and actually boxing out all the mechanical equipment. It's putting founding boxes, actually highlighting the different sections of the drawing. ⁓ So now this skill is repeatable and reusable on any of the documents that I want to upload in the application. And you can tweak and process that skill. So to go back to this position of like, hey, it's not in my workflow. Well, every estimator is doing a little bit different. You as an estimator can build a skill that follows your process exactly. And then it's repeated over and over again in the application to catch all these different things. ⁓ and you can just continuously upload document sets to do that. So ⁓ this is how Fable does it. again, like this is the idea that it gets you 90 to 95% of the way there. It can create different reports through different Excel exports, ⁓ different things that help you drive, like, hey, this is what's valuable to me as a user. ⁓ to get you to that point where now, like, I go look at the challenge areas, I go look at the penetrations through the slab, or whatever those things are that you're historically, you know, there's scope gaps around. ⁓ So again, this is this is out-of-the-box quad. So building that skill, ⁓ building out that process, updating the skill and continuing to evolve it around how you want to use the application ⁓ does a really, really good job for 20 bucks a month. And it can and it ties directly to the app the the way that you do it as an estimator and not kind of pigeonholing you into the process of how a software wants you to do it.
Luigi La Corte: So the does Claude actually draw the bounty boxes around the PMGs? What does it look like?
Mike Powers: Yeah. Yep. so these were bound ⁓ I lost you guys. So example ⁓ it put bounding boxes around ⁓ the pieces of quick here. ⁓ These lines. ⁓ I so I don't know why it's so blurry, but ⁓ yeah, ⁓ the bounty boxes. that bounding is getting better and better. I think Gemini does a better job with bounty boxes, but that just ⁓ has a better visual model, but yeah, they'll do bounty boxes around specific sections. you run into things with like context windows. So it's like, hey, let me just upload the plan set for the mechanical, the drawings you have, and it's gonna get you ninety percent of the there. Out ⁓ of box. ⁓
Luigi La Corte: Okay, so let's get okay. Let's let's ⁓ let's go round table with our with our conclusions then. So three questions. One, can Claude do quantities? Two, should subs Claude to do quantities? and would be an interesting third question? will they will they be using Claude for quantities? So those are the three questions. Can it do quantities? Two, should subs use it, three, will they?
Mike Powers: Yeah.
Luigi La Corte: Mike, what are your answers to all three?
Mike Powers: Yes, yes, and maybe.
Luigi La Corte: Andrew.
Mike Powers: So as a software pro again, at the whole point I'll say as a software provider, I think it's very important to educate your customers on how how the how these systems work. Because that's the value prop. Like you want to be consultant to them and that then teach them how to use these models. and that's that's the value prop you can provide as a software provider because they're not gonna have enough time to know, hey, do I use Fable, do I use Opus or do I use Gemini 3.5. ⁓
Luigi La Corte: Yeah, I agree with I agree with that. I agree with that. Bandre.
Andrew Zukoski: Yeah, I'll go yes yes no.
Luigi La Corte: Yes, yes, no. Okay. And I'll say I'll say yes, no, no. But I mean it's undeniable that they can do quantities. I think anyone who says at this point that they don't is like either they don't know or they don't want to know, but it like they can. What I just don't understand about the harness though is like how how how is it making an inference on area if it's if it doesn't have measurement tools? Is it just taking
Andrew Zukoski: Why wouldn't you give it measuring tools?
Mike Powers: Is use the tools.
Luigi La Corte: Well the the harness the harnesses don't the harnesses don't come equipped with like a scale measurement tool like a polygon trace.
Andrew Zukoski: Yeah, we got code for, there's Python libraries for that and Python libraries for opening PDFs. Have either of you guys tried out Bluebeam's MCP server? Have you used it? It's hilarious.
Mike Powers: And yeah, bluebeam I gonna say blueb. Mm-hmm.
Luigi La Corte: Hilarious how.
Andrew Zukoski: It's a local MCP server, when Claude is doing something, ⁓ UI has to be open and goes unresponsive and says, performing MCP operation. And I think that's, I think it's funny.
Luigi La Corte: But I don't think that's that's not what Cloud's doing when it's when it's generating an area. I don't think it's calling the Bluebeam MCP. It's probably using like a Python library, sure, but like those are not it's not like it's not accurate, right? It's like what is it doing? Like open C V and then it's just and then and then what counting the pixels, like maybe.
Mike Powers: So
Andrew Zukoski: maybe it should. Yeah. I mean, not a bad start. like, guys, like, you gotta, like, how do you not know this? Like, are we putting forward and we don't understand the, like, chain of technology that connected the drawing to the number? Like, there's a little bit of, like, wait, what? Like...
Luigi La Corte: Well, I don't know. I don't know what it's specifically doing. It just comes up with the number. Like with the quantities, it makes sense. Like I could take an image, and like the the the the image models are good enough to say, yes, there is one AHU And like AHUs are e are I'd say they're easy because they all have text tags and you can just OCR it. But if I have a floor plate, again it like it I don't think it's drawing the polygon. Well, how does it know?
Andrew Zukoski: Thank Yeah. Yeah. Yeah, I don't think anybody should use that then.
Luigi La Corte: Andrew's changed. He's a yeah, he's an he's a no no no.
Andrew Zukoski: Yeah, I've changed. I changed my mind, but like, yeah, I don't think we're gonna have this like unknowable black box. It's like, like it could be magic.
Luigi La Corte: Well yeah, that's why it's it's not useful.
Mike Powers: Never but never it should be in a noble black box.
Andrew Zukoski: I'm Zach Glantz.
Luigi La Corte: It is a bla it is a black box. Like it's good as a check. I I will agree. Like it'll say, Hey, like I think you've missed this elevator. That's fantastic. That's really useful. All right, we'll cut it there, but you heard it here first. yes, cloud can do quads. All right, everyone. I'll stop the stream.
Andrew Zukoski: Yeah. All right.
Mike Powers: When spent fifty dollars on Fable tokens and
Mike Powers: Happy almost fourth July.
Luigi La Corte: So we are going to Fable a spin now that it's back in everyone's pocket. we thought it'd be cool to benchmark Fable ⁓ against projects that have already been estimated and answer or continue to answer the question of whether these models out of the box can be fairly proficient at quantity takeoffs and whether they can do scoping. So for the agenda today, both Mike and I have processed a project where we have a ground truth. ⁓ we're gonna pull it up on screen and we're gonna walk through the results. And so I think Mike has a project. I have a project ready to go. then Andrew is I just gonna ask questions about the process. ⁓ So let's started. So I'm happy to I'm happy to go first. let me let me share my screen and let's get going. So the TL the TODR with this analysis was Fable. let me just tee it up for the for the team. So I uploaded a project fairly complicated.
Andrew Zukoski: Thank
Luigi La Corte: ⁓ manufacturing plant retrofit. this project was ⁓ I around three hundred million capex. and Fable actually did a good job. It did a good job on scopes. It did a poor job on quantities. And I'm gonna I'm gonna increase the size here and we'll be able to see what what they did here. So Let me just decrease this so we can kind of review this a bit. Okay, great. So this is ⁓ like I said, industrial manufacturing plant. So the headline is on scope versus ground truth, we have an F1, which is ⁓ like a data science term which balances ⁓ positives and negatives, false and true positives and false and true negatives against the ground truth. ⁓ and it scored ⁓ almost 90.90 on the scope identification. But a materially poor quantity takeoff at one. and I wanna walk through these obviously and go through go through ⁓ some of the specifics. But I will say out of the box, Fable is good at scoping. ⁓ the cross-document scope analysis is strong and it's only getting better. When we reviewed 4.7 against the same project, it was it was scoring 0.69. ⁓ so I think in general, scoping. is just gonna get easier with these models. Mike, I don't know if you saw anything similar.
Mike Powers: Yeah, and I think like like the zoom out of this is like like construction documents already are not a hundred percent accurate. S there's always gonna be scope gaps, and obviously there's an entire industry around doing scope gap analysis, there's an entire industry around doing document analysis, there's an entire industry around doing takeoff analysis. ⁓ but I don't even think being a hundred percent is the end goal. I think it's all about like helping with the efficiency of like getting you to review as an estimator or as a pre-con manager the pieces that need your review. Because in construction there's not a right answer. This isn't like, hey, ⁓ I'm going to buy like, you know, e-commerce widgets and I need to buy a widget and I know exactly the dimensions, the exact the output that I need. Like there's this flux in construction, and that's why general contractors are where they are. So I think the point and my position around this is like it's never gonna be a hundred percent, but doesn't need to be a hundred percent. ⁓ relied too much on construction documents already, does these tools actually push you more upstream to actually be a part of the design process more effectively to have these conversations earlier? And I think Fable does a very, very good job of getting you to that earlier conversation with with the engineer and architecture firm.
Luigi La Corte: Mm. I I'll disagree. I do think when you say we're overly reliant on drawings, I mean like that's just the pro that's just the process. Right? Like people have
Mike Powers: Doesn't need be the process though, because if you look at like the EU, it's not the process. There's quantity takeoff like at like the on a document set in the EU, they would have the quantities. Like why I is like I'm like befuddled that like an engineer or an architect puts, you know, 17 doors on floor one of a building, compresses it into a two D position, and then makes it very, very hard for someone to count those doors when in reality that's not the first principle problem.
Luigi La Corte: Or what what's the process?
Andrew Zukoski: Thank
Mike Powers: The owner should be pushing off against the architect and engineer saying, hey, instead of paying, you know, $10,000 for a takeoff or scope gap analysis, like this should be almost like an open source data structure that you can read and engage with however you want. And then by forcing it off of documents into an actual computer-red language, Fable.
Luigi La Corte: Mm.
Mike Powers: OpenAI, it's gonna be able to read those a lot more effectively because it doesn't need to be a human reading those those endpoints. Like why like this again, this is why like I'd be bullish on Autodesk. Like an autodesk can open up those nodes, ⁓ and you can engage on a front end with a chat interface. ⁓ it solves a lot of these problems that were that don't necess necessarily need to be solved the way they're being solved now. And that's like reading documents isn't the solution.
Luigi La Corte: I I also wish that were that I also wish that were true. I mean, selfishly, I'm happy it's not true because then it it allows me to build a business. ⁓ so it is somewhat selfish for me to say that. ⁓ but it would eliminate a lot of waste. And ⁓ it probably would bring the owner and the GC ⁓ in a in alignment more than they are now, or versus it's like it's very adversarial. But irrespective, I mean like the the industry just doesn't it doesn't work that way here for various reasons. And like there's no own there's no owner incentive to share that information. Because they know it's just gonna be used against them. And your point, the design is just wrong. so the more they share, I think the worse it'd be for them when they get caught in like a change order. Now I was like, I'm curious to get Andrew's perspective on this as somebody who helps manage VE's.
Andrew Zukoski: Yeah. and all of pre-con. ⁓ I don't know that if you look at Europe, builders substantially engage earlier. think on many projects they do, but you have this huge, like the consultant layer in the value stack is thicker in Europe. Like a lot of quantity surveyors work for these, are the people administering these quantities on the project, Like work for... consulting firms who are hired by the owners and owners are out there for Turner and Townsend or Mace or whoever, right? And the builder and the trades get brought in, I think there's design, bid, build work there actually for the true prime contractor. And they actually get brought in later. because the consultants and cost consultants handle a lot of the like early engagement and construction feedback during the design process. So it's sort of interesting. I don't know if it brings the GCs in earlier. Having it be a point of fact would be good. It seems like there should be some sort of open source repository where it's like, okay, what's in this drawing set? Like I got it from a plan room. What have other people found in it? It seems like this doesn't need to be carefully guarded secret knowledge anymore. ⁓ But Luigi.
Luigi La Corte: Yeah, I I agree, but Yeah, go ahead.
Andrew Zukoski: I was going change topic. Go ahead.
Luigi La Corte: I was just gonna ask, like, I think there is a a ⁓ a process, like there is val it's value add when a general contractor goes through a set of documents and decides how to scope a project. Because I think it does raise a lot of questions that could lead to to V E's ⁓ or to design changes or like variation notices that they that they think will benefit the product the the project. So just to say, hey, we could just solve this whole process and just give you the answer.
Andrew Zukoski: ⁓ yeah.
Luigi La Corte: ⁓ I do like I think there's a lot of value when when the G C goes through that process now and scopes the job themselves. That's why I also am very bearish on like one click scoping.
Andrew Zukoski: Yeah, I would agree on that. I feel like many of our most pre-con forward, not many, but several of our most pre-con forward partners, it's like who emphasize early engagement, who are doing complex work, where they're hand in hand with the design team. They talk about how at its root, the highest value add that pre-con offers is getting... the field team, the operations team spun up and engaged earlier on the project. They provide a mechanism for experienced superintendents and project managers who are going to run the job to engage early and start to understand it earlier and start to provide their own feedback earlier. And that it's all about bringing site knowledge early. And I feel like that's one of the things that would fall out of a scope analysis is like hey, this is how our team builds, or if it's like the actual team providing a little bit of input, like this is how this is gonna get built. And that can bring immense value to the owner and to the design team.
Luigi La Corte: Yeah, I I agree. ⁓ okay, let's let's continue going through this and ⁓ yeah, go ahead. Yeah.
Andrew Zukoski: Hey, Keith, can you scroll up, actually? So the little note there, the read this first thing, ⁓ is this true? Like this thing where it says that the asymmetry drives nearly all of the quantification gap below it? Like it looks like somehow some PDS didn't make it into the analysis. Like what happened with this?
Luigi La Corte: Sorry. Drawing PDFs were absent for the folder and the flag rather than guessing. It's should you know the quantification gap below barely affects scope identification. ⁓ let me check the folder and let's see. Copy remote IFC architectural. So the documents that I gave ⁓ that I gave Claude were the same documents that we gave to the third party for estimating. So it may it may be true that it thinks that that's the case, but we still got quantities from the other documents that we used.
Andrew Zukoski: Okay. Okay, so this is, this is, it is incorrectly identifying the cause of quantification data rise.
Luigi La Corte: Yeah, the the no the the this is the same the same set. So when we build our benchmarks, we take we take one set, we give it to human estimators, either in-house or ⁓ out out house, ⁓ additional additional resources. ⁓ and then we'll do the same thing with, you know, set like our our model, and then we'll do the the same thing with like claude LMs just to build our benchmark. So it's saying that
Andrew Zukoski: I gotta say though, we gotta mark the day on the calendar. Like today's the day that an AI, you know, blamed a poor document set for like not being able to quantify a project.
Mike Powers: Ha ha ha.
Luigi La Corte: Yeah, that is that is true. ⁓ okay, let's continue here. So let's go through some of these items here. Okay, so F1 by trade. ⁓ so concrete, so like this is one of the items here that I think is interesting. ⁓ the F1 is for concrete is zero, but there's you have so you can do this takeoff from an architectural set. And the architectural set is provided. It's a fairly large set of documents, like 250 pages here. ⁓ so like the thing I I had prompted it afterwards to ask it why it got these things wrong. And obviously, like it's what it's what you would understand. It's hey, I'm reading the quantities from tables. ⁓ I don't have ⁓ calibrated area takeoff tools, I can't do lines or you know linear lengths, and I I certainly can't do volumes, which is what is what concrete is. ⁓ so it's really guessing at these, but at the same time, it also got some of them right. ⁓ it got them right, obviously, because If you go to overhead and special special doors, like it's a table and I can sum those tables, it's not gonna give me perfect doors. So I'm I'm surprised that it got me an F one of one, which is not easy.
Andrew Zukoski: Hey Luigi, for folks who are listening who aren't familiar with F1, you know, is this McLaren versus Ferrari or is F1 something else?
Luigi La Corte: Think of it as as accuracy. Consider it as as like a 100% accurate, consider it would be it would be like an F1 of one. ⁓ and then an F1 of 50, it's like, you know, it's a it's a coin toss. ⁓ but you use F1 when you have an unbalanced data set. And so ⁓ you know, think of it this way. If I were analyzing a contract, ⁓ and there's one, there's one liquidated damage clause, but 99 non-liquidated damages clause, ⁓ could ask me what's your accuracy and I'd say well I found all the the non-liquidated damages clause except for that one. So I have a 99% accuracy when really ⁓ you failed. You you missed you missed the the thing that you were supposed to find, which was the LD clause. So when you have those unbalanced data same thing like in an MRI, like did are we finding ⁓ are we finding a like a cancer nodule or not? ⁓ even though 99% of the images are are not cancer nodules. So you you're using F1 when you have an unbalanced data set. And you normalize ⁓ by the ⁓ the expected amount of false positives ⁓ or sorry true positives. So ⁓ it just helps you understand what the accuracy is ⁓ without skewing that would be favorable towards you know somebody who has a bias. It tries to give you like, hey, zero percent F one means like we basically found nothing. ⁓ get like relative to that example that I gave you on ⁓ on on liquidated damages. I mean like the point, the point. Yeah, go ahead, Mike.
Mike Powers: So like the but I'm gonna push back like concrete like Like you're like you're not that there's not a there's not a defined quantity of concrete that's needed on a job. Like you're delivering trucks that are coming to fulfill the project need. And and you're like you're within a range. Like you're within a range. You're not like, ⁓ I have ten thousand ⁓ you know, vol like cubic feet of concrete for this project and it needs to be a ten thousand or nothing. Like there's there's slop, there's things that fall off the truck. Like I just I don't think being a hundred percent on concrete even matters. Like the concrete bitter is probably gonna know based on the size of the project how much concrete they need. Just by this linear square foot. They don't need do a full concrete takeoff. Cause my bet would be that the architect did not even put all the details to include the things that need concrete, like a curb for an air conditioning unit on the roof. So I think like
Andrew Zukoski: But for the big poor, like, you're gonna get...
Mike Powers: But that's like linear let's that that's square feet on a project. You're like, hey, the project's a hundred feet by a hundred feet and it's a hundred stories. You're not going down to like the minuscule square foot where you need a hundred percent take off of concrete ever.
Andrew Zukoski: That's true. But that is like the first quantity that our customers bring up when you talk about bin based takeoff. And so I think that that's, I think that cubic yards of concrete is like a, a key number. ⁓ it is cubic yards. I'm like, yeah, you, you don't buy it that way. You buy, you buy a certain number of trucks and you keep them coming until the thing is full. Like.
Mike Powers: But I I just I just don't think I would push back that. And I guarantee if you pulled ten general contractors, they're not like relying on an accuracy count of quantity from their concrete sub to within like five percent to say, ⁓ that's a that's the exact amount of concrete that
Luigi La Corte: I think I think they are. I mean like w
Andrew Zukoski: I think they are, and they're often self-performing concrete, at least are. Go ahead, Luigi.
Luigi La Corte: Well, I was gonna well, I I was only gonna agree with you. I mean, they they do this exercise now because they wanna see if their subs are in the ballpark or if they're missing something. So I think G Cs have a lower level of resolution that they're targeting with with quantities. But
Mike Powers: But that but that's the point I I would say like the that like I would bet, I would assume that saying, hey, give me if you asked Claude, give me the estimated cubic yards of concrete needed for this project, it will be within ten percent. And that's all that matters. It doesn't need to be within ninety-nine percent for concrete.
Andrew Zukoski: Hmm.
Luigi La Corte: Yeah, I would agree. I I don't think for the G C it needs to be within ninety nine percent. I agree.
Andrew Zukoski: ⁓ my god.
Mike Powers: So then so then it doesn't mat like why even like spin up why even have a specific takeoff tool that does concrete to that nine nine percent when the nine nine percent doesn't
Luigi La Corte: I don't think I don't think you need a specific tool for for quantities. But I do think the like the biggest drawback that you have right now, the reason why you don't see people doing it, at least at large firms, is ⁓ there's the there's almost no auditability. There's no way of knowing whether it's right. Versus like if the human goes to draw all the polygons, you know, they can refer back to those polygons. And ⁓ Claw just doesn't have the tool use to to ⁓ to draw those polygons yet. Like maybe that'll change. So I think the point here is like it actually does a decent job. Like it obviously does a good job on scoping. ⁓ and it I I I mean, what what we what were your results, Mike, for for quantities?
Mike Powers: So for quantities it it's like again it's within it's the it's it's within plus or minus five percent.
Luigi La Corte: Across every trade.
Mike Powers: Across the trades I think you need to do quantity takeoffs on. Doors, windows, mechanical equipment, electrical equipment, things like carpet, things like concrete don't matter. Things like the amount of paint that goes on a wall. Like that that painting sub is cat like they're not saying, ⁓ we're gonna buy exactly ten hundred gallon tubs of paint. They're buying whatever 'cause there's gonna be need to be touch ups and like when the guy scrapes the wall, like out of the same paint, like
Luigi La Corte: Like
Mike Powers: I just like don't like so there's there's this like bifurcation of trades that matter and things that you have to like order and you have to you can't get from Home Depot or your local distributor need to have accurate accounts. So doors framed and hardware need to have accurate accounts because you can't just go get the 20, the 36 inch wide door down at Home Depot the day of the project. Things like an air handling unit need to have accurate accounts because you can't get because there's all D time on it. But these things that like are commoditized doesn't matter.
Andrew Zukoski: Is that because you bid it out to the trades and the trades are going to do a detailed takeoff?
Mike Powers: Trades own it. The trades own it. Like the like the like I'm like I I just like I think you get in the ballpark with Claude, you can level out the quotes and you can compare your subcontract. You get three concrete quotes from your three concrete subcontractors if you're self-performing and you're in the ballpark.
Luigi La Corte: How how can you bid level without knowing what the subcontractor amount should be?
Mike Powers: God directors do all the time.
Andrew Zukoski: Don't you always, you always, you have four, if have three bids in your own number, the assumption, like you're gonna, your number is missing things too. Like, look at this, like apparently, apparently you missed an elevator. right, that happened, that just, I mean, that's what we hear is like when you're leveling, it's like, we got four data points and I don't know that. I probably trust any one trade number more than I trust my own number.
Luigi La Corte: But then they make mistakes.
Mike Powers: Yeah. Forget the curve. Forgot the curve.
Luigi La Corte: Well right.
Mike Powers: I just I just don't think you can tr like I think the whole point is that you can't trust the drawings enough, you can't trust the documents enough to make a hundred percent assessment on a takeoff. Ever.
Luigi La Corte: But I I think that's your only recourse, right? Because once you enter into the contract, I'm I'm building per the drawings. Like that's that's a it's an exhibit in your in your contract. And if it deviates from that, like I can you know, I'm gonna go after a variation notice.
Mike Powers: You have intent though. But it's you still have intent as a contractor. But you stole an intent as a contract. For concrete, for a concrete pour, you're owned, you own intent. You own that you're probably gonna do engineering slabs, like you're gonna own the engineer system realmant. It's the same with structural steel. You own the structural steel. The majority of the time, the structural steel provider is owning the shop drawings. They're owning like like the thickness, the bull like the anchors, the welding, like they're owning all those pieces. So I don't think that I don't think the I don't think the structural engineer, the structural engineer is gonna sign off. Right, after a submittal, after shop drawings are submitted through the out through the process. But that document that comes out that you're doing a takeoff on.
Luigi La Corte: Mm.
Mike Powers: Like the there's there's so many mis there's so many missing pieces.
Luigi La Corte: But not in a ⁓ in in a design build, yes. But if I'm supposed to just be building per the drawings, the owner and the designer own the quality of those documents.
Mike Powers: Not all the time. Like so I was at the the first job I was on was in the Eagle Stadium, Philadelphia. And the architect wanted to hang ⁓ wings that moved in the concourse. That was the the the contractually on the drawings, there was these artificial wings that were in that were in the concourse that would move with the wind. ⁓ that was spit out to the structural steel subcontractor, W and W steel.
Luigi La Corte: Okay. Yeah.
Mike Powers: ⁓ they could not produce those wings. Because it was technologic it was engineeringly impossible to be it we able to withstand the engineering requirements of being outside plus moving in the Like that that that that there's nothing on the drawings that's gonna tell you differently of that working. And the structural steel guy it wa it was on the drawings, but it wasn't incorrect it was incorrect on the drawings. It was it was technologically impossible to make the way that the drawings were laid out. So
Luigi La Corte: It was on the drawing it was on the drawings though. It was on the drawings. But that's different than saying, Okay, the intent was I had three elevators. You ought to have known that G C versus like, Hey, I only see one elevator bank here. That's different.
Mike Powers: I that's a that's a hole. That's a that's just like the whoever laid out the elevator tell the laid out the elevators wrong. Like I'm gonna say I think that there's so much fuzziness in this world unless it's things that you need to that that you have to like, I need a widget. I need a I need a I need something that pushes air, I need a fan in in the ceiling. That's a quantity takeoff that can be done. And Claude does a very, very good job of saying there's a fan in the ceiling here on the drawing. As soon as you get to conduit wiring.
Luigi La Corte: Mm. Yes, I agree with you.
Mike Powers: Like that's never gonna be accurate. You can never you're never gonna accurately get something off of a two D document that's built in the 3D world. Ever. Never.
Andrew Zukoski: How, how, like, so take the GC, let's accept for a second, I'm trying to understand this, like, except that the GC says, okay, hey, quantity is never gonna be more than plus minus 5%, whatever, can't trust the quantities. If I'm a trade, like, I have to bid the project, I have more exposure to self-perform groups inside of GCs than I do to like strict trade contractors. But I know that they are obsessed with a level of precision that is much higher than GCs have, both for quantities that you would order, but also for quantities that you produce, right? Like they're tracking labor rates. They like close the loop between their production labor rates and their estimating labor rates. If you're installing a, if a crew can install a certain number of units of something per day, like you're doing your resource planning based on how long they're going to be on site. Like, and there's a lot of fuzziness there, but like at some point you do have to be on site or not. Are you planning for 10 weeks or are you planning for 12 weeks? Cause like the job will work at 10, but it won't work at 12. Like what's like, like trade, trade margins are higher. than GCs and it varies, like, generally substantially higher, but they're not like that much higher. You have to.
Luigi La Corte: I think they're gonna wanna be as correct as possible. But I mean if the if the question is can clod do quantities, the answer is yes, it can. You just you you'd have to have confidence that it's gonna be like I think for your quantity type, it's probably more doable. It depending on the quantity type, it's more doable or or less doable. ⁓ like maybe not for concrete volumes, but definitely for doors. Great. You can like you can use clod. In which case, like the there's no real there's no real debate. Like it it can do it. But There's different question is whether or not it works within your organization.
Mike Powers: Yeah, and just like and going back to the subcon tracking position, like mean this is going back to like that clawed article of like, hey, there in commercial construction, there's like 20% upside for AI LLMs because that's like the manual process of things that they're doing on a computer. But to back to Andrew's point of like, hey, I'm running ⁓ I can run hundred linear feet of conduit conduit a day per team. That has nothing to do with anything in LLM world. How do you do 200 linear feet? ⁓ I have to have two teams. there's not a robot that's like like laying conduit. It's labor ⁓ or prevailing wage labor. Like that's that that's a mechanism that nothing around AI is gonna touch for the foreseeable future because LLMs aren't gonna solve physical world problems like some of the physical world models that are coming out. So I think like that point of like, hey, the twenty percent of my work as a subcontractor needs to be done, can live with that being or minus ten percent accurate on quantities, the whatever you want to do? I think the answer is yes. If that helps you get a a response faster, it's helpful. Or if it helps you like find the gaps. Like it has like an estimating assistant that's sitting there, like, hey, I'm gonna go through all the documents as this estimator because I own the documents as a mechanical estimator. All right, I'm gonna have a bot or an agent go through and like also do the same kind of process that I'm doing. as a mechanical estimator that's been in the space for 20 years knows this mechanical engineer never puts fire on the drawings through the penetrations. All right, so every time I have a slab penetration, there's there's mechanical, there's conduit that's running through the floor. I need to add ⁓ dollar fifty per penetration. wait, but the drawings show just line of this conduit that's running through for the electrical or for the mechanical piping, the sprinkler lines. I know there needs to be three. Why is there three? Well, I knew like here's the reason why there needs to be three because of this code piece. And I think like Until those pieces, until that mechanical subcontractor is like embedded with that mechanical engineer and they're both incentive the both incentives are aligned to make the document as accurate as possible, doesn't there's no reason to solve for the document issues at Because they're gonna be solved. There's no incentive for mechanical engineer to put fire stopping on every penetration through the slab. ⁓
Andrew Zukoski: I mean, it like it can be straightforwardly inferred.
Mike Powers: Yeah, the yeah, like the per the the mechanical the mechanical subcontractor is inferring that they need to have that there. if you're reading the documentation, the documentation's not a hundred percent accurate about that. Or the spec says that you need fire stopping through the penetration, but the drawings are never gonna show that. ⁓ then the then the engineer has no incentive at all because they're billing by the hour. They're already squeezed by margin to sure that's accurate.
Andrew Zukoski: ⁓ okay, so you're Okay. So you're saying a literalist interpretation drawings, like, well, not that miles to go before that. can literally count the stuff on the drawing. Okay. ⁓ That's I feel like that's different from saying you can take a drawing set and use given just a drawing set, come up with.
Mike Powers: Correct. matter. Doesn't matter.
Andrew Zukoski: accurate quantities for what's necessary by applying knowledge and understanding.
Mike Powers: For the thing for the things that are accurately that need to be quantified, like fans or windows, where there's an actual like someone's actually doing this for like for windows, you're doing lighting, whatever you like that is yes, that's solved. For the things that don't matter, you don't need to have ⁓ windows on the wall or three light fixtures in the ceiling here. So paint, drywall, concrete, that doesn't So it's it's it's not a it's there's no reason to be a hundred percent accurate on those pieces. So you're you're bifurcating the the processes and you're like verticalizing a supply chain where things can be verticalized. the things that can be verticalized are things that are engineered, things that you need to like have a system design around, things that need to be like produced in Mexico and shipped to the United States because you gotta ⁓ chain manage that. You gotta make sure that that light fixture is there. But it's drywall or paint or concrete, like being within 10% doesn't matter. Like you're it's it's not gonna make or break the project.
Luigi La Corte: To who? Doesn't matter to who. Not to the the sub that's doing the partitioning.
Mike Powers: To anybody. To anybody. Like the painting the painting is not making or breaking the project because they had to buy ten percent extra paint.
Luigi La Corte: It's not like they have it's not like they have like twenty five percent margin.
Andrew Zukoski: Might be if they have to install 10 % more paint. Like what's a typical painting, painting, I mean guess you can recoup it through like through a change order to a certain extent like hey we.
Mike Powers: It's brut it's like brutal. Is that that painting subcontractor though is just betting on the fact that they can paint faster and better than their competitor.
Luigi La Corte: Definitely do it definitely does matter.
Mike Powers: And it's the manpower piece, it's not the takeoff piece.
Luigi La Corte: Well, I mean like do end up multiplying the square f like the square footage by a labor productivity rate. I just to think to think that to say it doesn't matter is like what I what I disagree with because they spend a lot of time making sure it's as accurate as possible. And if it didn't matter, why not just take the area kinda grow like grow square feet and apply a number and just be like, we'll see what it is. It doesn't really matter. Whereas reality, like they spend a lot of time going through to make sure it's as accurate as possible. So if it doesn't matter, why do they do that?
Mike Powers: I don't think painting subcontractors do.
Luigi La Corte: Say it again.
Mike Powers: I don't think painting subcontractors do.
Andrew Zukoski: Drywall, drywall on concrete definitely do.
Mike Powers: I just think there's so many there's so many dependencies on a project like that you're like already putting 10% extra drywall on a project because the comes sweeps by and and ⁓ and half of it. And that's not a change order. That's just kind of fixed on a punch list. I I feel like it's like it's like you're you're working toward it's not like self-driving where you need to be 99.99% accurate to save somebody's life. So if you can get to 90% accurate, all my position is if you can get to 90% accuracy by launching it through Fable, is good enough for the market. Because you don't need to be 99.99% and those additional nines that you're trying to drive for are just increasing the cost to provide the ROI that's needed.
Andrew Zukoski: That's interesting, right, Lake?
Mike Powers: Like if you're a painting subcontract paying a hundred thousand dollars a year on a takeoff tool, doesn't make sense.
Luigi La Corte: If it if if you spend ten last point on that, if you spend ten grand, if you spend ten grand for a takeoff tool, okay, it saves you from having to overspend relative to your budget on projects. Let's say your average contract value is a million dollars. That means you're spending an extra hundred thousand dollars ⁓ ⁓ on material, it more than pays for itself on one.
Mike Powers: I would say the controlling of your items on a project your process for QAQC is of a variable than the takeoff that you've done and the portion.
Andrew Zukoski: like if the infield QAQC. for some trade share.
Luigi La Corte: Help me ⁓ help help me help me understand that. So you're saying what actually happens on site is a variable that I can't control. And so it doesn't matter whether I hit but the thing is like it could hit me twice.
Mike Powers: No, it the the QAQC is what you can control on a job site by how you manage your team. We were gonna stick with a painting concrete contractor as an example. are you managing that process on a job site, which is the majority of where they're gonna lose money as to saying like, hey, ⁓ I you know, I bought a hundred of paint as opposed to a hundred and ten gallons of paint. So that quality, QAQC process is more important to that subcontractor than the initial estimate takeoff.
Luigi La Corte: But then I get hit both ways. I'm both short on quantities and I'm low on then I'm also hit on QAQC versus like I could only be low on QAQC unless I'm missing something.
Mike Powers: I just think I don't I I they so again, like, hey, we're doing the estimate, we're doing a takeoff. ninety percent accurate on my takeoff for paint by using whoever whatever. That's checkbox. Being ninety-nine percent accurate on the takeoff at that point is not balancing the profitability of you once you do the installation as a contractor.
Luigi La Corte: So what do you think is if so let's let's say that were true. what do you think is stopping subs from just doing their quantities in in cloud right now?
Mike Powers: I think it's education. I think some are. Some c some contractors are. think it's just like why like there there's a fear of like, hey, what actually happens when I upload my doll documents to Claude? Like who owns data? getting approval on twenty dollar Claude subscription is actually a hard thing for subcontractors right now, unless you have someone in IT that's like educated in that model. ⁓ we've talked to contractors still have on-premise email servers. Like and and the idea of like buying an AI that they don't really understand is is a huge hurdle. So that's my my entire pitch is that you should always you should like you should try it in Claude. Whatever you're trying to do, set up a skill, try it in Claude. you get ninety percent of the the accuracy that you need, I that's good enough for a
Luigi La Corte: You it'll
Andrew Zukoski: It's, it's this like 90 % number, like I'm very sympathetic to the argument that you don't need perfect accuracy, especially in QTO. that makes a ton of sense to me, but the 90 % feels low, man. Like if it's, if the arrow bar is greater than your median, like gross margin on a project, like something about that gives me the heebie.
Mike Powers: W well you're not just launching it. You're not just launching it. You're not saying, hey, send it to Claude and then just send the quote back to the general contractor. still having human in a loop, so you're supplementing it. How is your estimator? Is your estimator over 90% accurate right now?
Luigi La Corte: I would say most estimators at subs are probably at around like ninety-eight percent accuracy. That's like my vibe.
Andrew Zukoski: greatest. Yeah. They only do one scope, they do it day in, day out. They've got tight connections to people performing a work.
Mike Powers: Ninety percent accuracy on a document set that's not accurate though.
Andrew Zukoski: Yeah, because they've got all the rules of inference that you described earlier. They live that all day. I mean, it's interesting to like, you guys remember when Zillow did auto buying of homes, right? They built a model that could automatically price a home in a given thing. it like, they lost a ton of money. And the estimates were probably actually pretty good.
Luigi La Corte: But it doesn't like the accuracy component you can carve out.
Mike Powers: Yeah. There's
Luigi La Corte: They lost so much money.
Andrew Zukoski: The trick was they set themselves up for like rampant adverse selection where everyone who was in the house knew more about the house than Zillow ever could. And so if they thought they could do better than Zillow, they didn't sell on Zillow. And if they thought Zillow was better and they were going to get in the market because there was disgusting black mold infestation that was going to come out immediately when someone walked in the door, like they sold to Zillow and Zillow got fucked. Right. And so you like, you like you get If you drive 90 % accuracy and you're 10 % high, you don't win those jobs. They don't contribute to your margin. You only win the ones where you come in low. There's this selection that means that the error is accentuated in projects that actually win in a way that is not to your benefit. And that makes me wonder if that 90 % is like a man that feels low.
Mike Powers: But you but you're not you're not just like right like so hey all right I'm going back to the painting example. getting a budget number to my general contractor using Claude. Hey, I'm gonna launch this document send it to Claude, ask it to do a painting takeoff. All right, here you go. Here's my here's my 90% budget number. you as an estimator are still going through process to like invalidate those pieces. Claude does a great job of giving you the steps, giving you the data, giving you, like you can set up a skill to do this and you can find like you're exploring the gaps. Like you're looking at, you're having this like a sit again, it's like this assistant that's helping you for 20 bucks a month. And like that, like it's never gonna be you're you know there's no reason to be 100% because you're never gonna like there's no reason to be 100% because the documents that are never 100% accurate.
Andrew Zukoski: Yeah.
Luigi La Corte: That but that argument though doesn't hold in my opinion doesn't hold water because if I this is a lump sum contract and the owner has responsibility for design errors, I'm accepting those documents as is. If there's an error with the documents, I have recourse. could sue you. If you're not gonna give me substantial performance because you I I have to put in this elevator bank that was never accounted for, that's a that's a problem. So yes, the like the drawings are never completely accurate, but it's still like that's my contractual basis. If this were to go to court, they'd say, okay, it's an app it's exhibit A in the contract. There's clearly 200 200,000 square feet. The owner's asking for another 40,000 square feet that's never shown on the drawing. It's like I actually have legal standing to sue you for it. So I think that's what they end up relying on. But if we were gonna say Claude is gonna be an additional component, like a check, okay, well then then what we're not saying is it's gonna replace estimators or it's gonna replace estimating tools because estimators still need to go perform that takeoff themselves.
Mike Powers: 100% you own intent. You're never gonna offload that unless that unless the subc unless the software provider is absorbing all liabilities, you be a subcontractor and say, I'm just gonna like rely on the quantities of the software, which means you're gonna have to human verify everything that you're doing on a takeoff. You're gonna have to trust but verify the application. So that delta of ninety percent accuracy to 99% accuracy. Doesn't actually matter. And there's like the there's the psychological trick. Like when you go to an Apple store, for example, they have their computers at whatever degree, so that you go grab the computer and bring it up 'cause you feel ownership. Is it actually better? for a user to have a 90% accurate tool because then you're feeling ownership of the documentation by going through and validating all the different nodes that were inaccurate being 99% accurate because that 1% is going to screw you over because you're not trusting and verifying the process or the ⁓ the the process.
Luigi La Corte: Then what we're describing is a tool that's not as good as their current their their current practice. Versus like, okay, let's let's take let's take like a like a like a company that I think is like is is is really good. So like Bob Yard, for example. They started with like a single trade landscaping. when do takeoffs, you know, don't just do like counts, linear area, they break down into different types. ⁓
Mike Powers: Is
Luigi La Corte: So there's symbols, there's hatched areas, there's like r irregular polygons, et cetera, right? And so like they'll put that on the screen for you and you can just click. My so people are people are using it because it's very fast, but because it it maps to their current process. And at the end of the day, the estimator is still responsible. If you look at what Claude can do relative to that, I would say it fails in situ because it doesn't meet them where they are, but it's a good check. So but like, okay, so then if the question is like, can can Claude do quantities? The answer is yes. Can it do quantities the way subs currently do them? I would say no. It doesn't, it doesn't pass that test. So I like I'm st I'm still long, I'm still long quantity software because still need a human to verify, but unless Claude can just spin that up and like make a model on the fly that allows you to verify these accuracies. Sorry, ⁓ ver verify the accuracy. Like I I don't think that's gonna change.
Andrew Zukoski: Gents, I want to take a sideways for one minute. You talked about how education was a big barrier to trying this out. Can we get a tour? Do you guys want to do a tour of the setups that EG to drive quantities and scopes from Fable? What it like? What's the skill word? What's the...
Luigi La Corte: Yeah. Mike, why don't you go? I can show mine right after.
Mike Powers: Yeah. So anytime you do anything in Claude that you're doing a repetitive task. So there's two different systems in Claude that are powerful. One is a project. A project has an end date, so it closes out. So it's something that's like short term. The other tool is what's called a skill creator. And anytime you do anything in Claude that you're trying to do a repetitive task, for example, like a quantity takeoff, you want to ask the skill creator, so if he's backslash skill creator, ⁓ to build a skill, like say build a skill to do a quantity takeoff of a mechanical set of drawings. So this is going to take time to do, ⁓ but I'm running this in Fable Max. ⁓ and then you can call that skill to do a task. ⁓ so if I go over to this tab, it's actually processing this document. It's going through and checking all the different drawings. I hope that probably doesn't help. ⁓ so this is an example of what, oops, if I can zoom in. ⁓ what Claude is doing in the background.
Andrew Zukoski: I don't see any tab.
Mike Powers: So this is a quantity takeoff. It's going in and actually boxing out all the mechanical equipment. It's putting founding boxes, actually highlighting the different sections of the drawing. ⁓ So now this skill is repeatable and reusable on any of the documents that I want to upload in the application. And you can tweak and process that skill. So to go back to this position of like, hey, it's not in my workflow. Well, every estimator is doing a little bit different. You as an estimator can build a skill that follows your process exactly. And then it's repeated over and over again in the application to catch all these different things. ⁓ and you can just continuously upload document sets to do that. So ⁓ this is how Fable does it. again, like this is the idea that it gets you 90 to 95% of the way there. It can create different reports through different Excel exports, ⁓ different things that help you drive, like, hey, this is what's valuable to me as a user. ⁓ to get you to that point where now, like, I go look at the challenge areas, I go look at the penetrations through the slab, or whatever those things are that you're historically, you know, there's scope gaps around. ⁓ So again, this is this is out-of-the-box quad. So building that skill, ⁓ building out that process, updating the skill and continuing to evolve it around how you want to use the application ⁓ does a really, really good job for 20 bucks a month. And it can and it ties directly to the app the the way that you do it as an estimator and not kind of pigeonholing you into the process of how a software wants you to do it.
Luigi La Corte: So the does Claude actually draw the bounty boxes around the PMGs? What does it look like?
Mike Powers: Yeah. Yep. so these were bound ⁓ I lost you guys. So example ⁓ it put bounding boxes around ⁓ the pieces of quick here. ⁓ These lines. ⁓ I so I don't know why it's so blurry, but ⁓ yeah, ⁓ the bounty boxes. that bounding is getting better and better. I think Gemini does a better job with bounty boxes, but that just ⁓ has a better visual model, but yeah, they'll do bounty boxes around specific sections. you run into things with like context windows. So it's like, hey, let me just upload the plan set for the mechanical, the drawings you have, and it's gonna get you ninety percent of the there. Out ⁓ of box. ⁓
Luigi La Corte: Okay, so let's get okay. Let's let's ⁓ let's go round table with our with our conclusions then. So three questions. One, can Claude do quantities? Two, should subs Claude to do quantities? and would be an interesting third question? will they will they be using Claude for quantities? So those are the three questions. Can it do quantities? Two, should subs use it, three, will they?
Mike Powers: Yeah.
Luigi La Corte: Mike, what are your answers to all three?
Mike Powers: Yes, yes, and maybe.
Luigi La Corte: Andrew.
Mike Powers: So as a software pro again, at the whole point I'll say as a software provider, I think it's very important to educate your customers on how how the how these systems work. Because that's the value prop. Like you want to be consultant to them and that then teach them how to use these models. and that's that's the value prop you can provide as a software provider because they're not gonna have enough time to know, hey, do I use Fable, do I use Opus or do I use Gemini 3.5. ⁓
Luigi La Corte: Yeah, I agree with I agree with that. I agree with that. Bandre.
Andrew Zukoski: Yeah, I'll go yes yes no.
Luigi La Corte: Yes, yes, no. Okay. And I'll say I'll say yes, no, no. But I mean it's undeniable that they can do quantities. I think anyone who says at this point that they don't is like either they don't know or they don't want to know, but it like they can. What I just don't understand about the harness though is like how how how is it making an inference on area if it's if it doesn't have measurement tools? Is it just taking
Andrew Zukoski: Why wouldn't you give it measuring tools?
Mike Powers: Is use the tools.
Luigi La Corte: Well the the harness the harnesses don't the harnesses don't come equipped with like a scale measurement tool like a polygon trace.
Andrew Zukoski: Yeah, we got code for, there's Python libraries for that and Python libraries for opening PDFs. Have either of you guys tried out Bluebeam's MCP server? Have you used it? It's hilarious.
Mike Powers: And yeah, bluebeam I gonna say blueb. Mm-hmm.
Luigi La Corte: Hilarious how.
Andrew Zukoski: It's a local MCP server, when Claude is doing something, ⁓ UI has to be open and goes unresponsive and says, performing MCP operation. And I think that's, I think it's funny.
Luigi La Corte: But I don't think that's that's not what Cloud's doing when it's when it's generating an area. I don't think it's calling the Bluebeam MCP. It's probably using like a Python library, sure, but like those are not it's not like it's not accurate, right? It's like what is it doing? Like open C V and then it's just and then and then what counting the pixels, like maybe.
Mike Powers: So
Andrew Zukoski: maybe it should. Yeah. I mean, not a bad start. like, guys, like, you gotta, like, how do you not know this? Like, are we putting forward and we don't understand the, like, chain of technology that connected the drawing to the number? Like, there's a little bit of, like, wait, what? Like...
Luigi La Corte: Well, I don't know. I don't know what it's specifically doing. It just comes up with the number. Like with the quantities, it makes sense. Like I could take an image, and like the the the the image models are good enough to say, yes, there is one AHU And like AHUs are e are I'd say they're easy because they all have text tags and you can just OCR it. But if I have a floor plate, again it like it I don't think it's drawing the polygon. Well, how does it know?
Andrew Zukoski: Thank Yeah. Yeah. Yeah, I don't think anybody should use that then.
Luigi La Corte: Andrew's changed. He's a yeah, he's an he's a no no no.
Andrew Zukoski: Yeah, I've changed. I changed my mind, but like, yeah, I don't think we're gonna have this like unknowable black box. It's like, like it could be magic.
Luigi La Corte: Well yeah, that's why it's it's not useful.
Mike Powers: Never but never it should be in a noble black box.
Andrew Zukoski: I'm Zach Glantz.
Luigi La Corte: It is a bla it is a black box. Like it's good as a check. I I will agree. Like it'll say, Hey, like I think you've missed this elevator. That's fantastic. That's really useful. All right, we'll cut it there, but you heard it here first. yes, cloud can do quads. All right, everyone. I'll stop the stream.
Andrew Zukoski: Yeah. All right.
Mike Powers: When spent fifty dollars on Fable tokens and
