How to Actually Use AI for Quantity Take-Offs transcript

Recently, when GPT-6 Astra launched, it was touted as the AGI era. Well, by our testing, it only scored 47% on simple quantity takeoffs. Any construction business that is doing their takeoffs at 47% accuracy isn't going to be around for long. Does that mean I think using AI for quantity takeoffs is useless? Well, absolutely not. I just don't think it's useful in the way people think it is. So, in this video, I'm going to show you how I think AI is genuinely useful in the quantity takeoff process.

I'm going to walk step-by-step through the entire process, from what templates and documents you should have set up before you try to run it to actually using it to organize and structure your construction drawings, to then using it to customize a template library of assemblies to the specific items you're going to measure on your construction drawings, how to actually use it when measuring, counting, taking areas, and then using it to export your quantities from your software into usable output formats, whether that be an estimate, using it to create a bill of materials for a supplier, or generate a schedule.

I think AI has huge potential to save people time when doing takeoffs, just not with the actual measuring and counting. I think if you expect to be able to just take a set of construction drawings and dump them into Claude and get a perfect bill of quantities out of it, I think you will be disappointed. Okay, so first, I want to explain the architecture of this setup. So, I do my takeoffs in ZZ takeoff. It's a low-cost assembly-based takeoff software.

All an assembly-based takeoff software means is that you For every item you count, it automatically derives the labor, plant, materials, and subcontract costs associated associated with that based on the assembly. So, you create this ratio of quantities that you actually measure on the drawings. For example, I'm measuring linear meters of cable tray, and from that, I'm deriving how much cable tray material I need. That considers wastage allowances and the fact that I buy the cable tray in 3 m lengths, that the fact that I need 0.4 man hours per linear meter to install, which calculates my labor, that I need a bracket every 2 m, so I need 0.5 brackets per meter, that I need a splice plate every 3 m.

So, once I set up that assembly structure, all I then do is I measure on the drawings the linear meters of cable tray, and then it derives my secondary quantities from this. Then I either connect ChatGPT or Claude to ZZ takeoff. So, this is via an MCP connection. So, if I open up my Claude account, I go to customize, I go to connectors, I just have the I just have ZZ takeoff set up as a connector. Therefore, when I go to and create a new chat, I say, "Can you give me a summary of my quantities on the electrical plans in ZZ takeoff?" And then it will be able to basically access and read this data.

With the MCP, you can use your Claude or ChatGPT subscription just as you would when you're writing any prompts, but now it can pull all the data from your takeoff software. Now, while this is running, I will just make a comment about purpose-made AI takeoff software. I know there's quite a few on the market. I am personally not that bullish on this technology. My main reasoning for thinking this is that while like these AI models aren't necessary Well, I just have to approve the connection to Claude.

So, it can read the data, but the comment I was going to make is while there are purpose-made AI takeoff tools at the moment, and they are performing much better than Claude or ChatGPT. From what I have seen, they seem very expensive for what they are, and again, from what I've seen, none of them are claiming to be 100% accurate with takeoffs. And if they're not 100% accurate, you still have to go through and check that they've done the take off correctly, which I guess I question how much time it actually saves you.

Now, the other reason I'm not don't necessarily think you need to go with a AI specific take off tool is I think these numbers will continue to climb as the models get better and better, and I'm not sure how long this will take, maybe a couple of years, maybe sooner. I eventually think take offs will be a solved problem, which just means the base models like ChatGPT, Claude, Grok will effectively be able to do highly accurate take offs.

And therefore, I question how much sense it makes going and investing in an expensive take off software, storing all your assemblies, building out all your infrastructure within the software, and getting locked in to the vendor. And the final reason I guess I'm not I did I sort of question the value in a lot of these like purpose-made take off softwares is if take offs are really your problem, and you're spending all this time doing quantity take offs, then I would look for a high-quality person on like Upwork or Fiverr, where for $10 an hour, they probably could do a complete take off in a couple of hours, and it would be accurate in the format you want it to.

So, I really think if take offs are your problem, and you want fully automated accurate take offs, then just find someone on Upwork to do them for you. But again, that is just my opinion. If you do see an AI take off software that you think looks good, most of them let you do free trials or at least give you pretty detailed demonstrations before you have to sign up. So, I would just recommend trying and experimenting.

But I like this setup because I think it actually leverages AI where it makes sense in the take off process, and will actually save you a lot of time, and it also gets better as the base models like ChatGPT and Claude get better, which they almost certainly will. Okay, so the first thing I want to comment on is the background information you need to actually make this work. And the first thing I think every business should have is a template list of assemblies.

Now, this doesn't necessarily need to be in your take off software. All it is is a list of the primary quantities you measure on the drawings and what are the secondary quantities you derive from it. So, for example, for every light fitting you measure on the drawing, you have your mounting clips, you have your junction box, you have 1.5 m of cable, you have 8 m of circuit cable, blah blah blah, you have your labor associated with that, whatever you however you choose to break it down for your specific trade, just have a simple ratio of what are the primary quantities you measure, what are the secondary quantities that you derive from it.

This doesn't have to be job specific, it should just be a standard business library based on the typical work you do. Because then all you're doing for each of your projects is going to be taking this standard set of assemblies with all your productivity rates in them and customizing them to the specific project. The other thing I recommend always knowing is what is the output format you're eventually going to want the quantities in.

So, what are you actually measuring the quantities for? Are you using them for procurement, in which case you want the materials broken down and product specifications nominated nominated, or are you using them to input them into your Excel estimating spreadsheet? And then what is it So, what is the eventual output format? Then the next thing I do is I basically just take my drawings and I tell AI to restructure them so they're easy to read.

I don't know about you, but when you get a 2D set of construction drawings and you're like, for example, this is a road project and it was originally an individual set of sheets, but what I got AI to do is to merge them together so they're easy to read. This is just a HTML file, AI's done the conversion from PDF to HTML. Just makes it so easy to go through and navigate your drawings. Or for example, I've got a set of architectural drawings here.

I break it down into architectural, electrical and mechanical. I can flick through all the drawings, it extracts all the name, it groups it by drawing type. And then I can click on tags and I can see the references to the tags and schedules and I have this on a split viewer. Just makes it very easy that when I first get a set of construction drawings to fly through and understand the scope. Much more much I find this much easier than scrolling through a huge set of 2D drawings.

So, maybe that's not very useful for you, but I often find the hardest thing about when you first get a set of construction drawings is actually working out what everything is. The next thing I do, and I've already done this on this set of drawings, is I just tell Claude, I say, "Can you please restructure my drawings so they are grouped into a logical folder structure?" Now, I've already done this on this set of drawings, but what I AI then does is I just get it to group it.

I mean, obviously this is very specific to the project, but for example, on this set, it's grouped it onto ground floor and first floor because it's commercial project, and it has broken it down by the different types of drawings. Lighting, lighting small power, electronic systems, data, and fire security. I find this much easier than just having a big long list of drawings and that I'm constantly scrolling through to find stuff.

And Claude can just do this via the NC pay. The next thing I do is then I take my business list of assembly, my standard list of assemblies, and I say, "This is a standard set of assemblies for me doing electrical quantity take-offs. Can you please use this standard set of assemblies and customize it to my project specifics? For example, take the light fitting assembly and customize it to the specific types of light fittings I have on the project." So, I copy that in.

My voice dictation tool didn't work amazingly, so the prompt copied perfectly, but what this will then do is it will take and I've already started a take-off on this drawing, so it's not really strictly necessary, but what it will do is then look at, say for example, my light fitting assembly, it'll look at the legend and say, "There's six types of light fittings." It will create an individual assembly for each of those light fittings.

And this will be based around this specific uh ratio of primary versus secondary quantities that I already have set up in an Excel document. Now, I'm not sure about you, one of them like most time-consuming things I actually find when I'm doing a takeoff is going with a new set of drawings and saying this is the B1 light fitting assembly, this is the B2 light fitting assembly, or if you're doing like trenching, these are all the different types of trench types.

Now, you obviously have to check how it's doing this and how it like whether it seems to have done it just to actually make sure it's done it accurately, but really all it's doing is saying there's eight types of light fittings, I'm going to create this uh specific product best matches this item in my assembly library. I'm going to take that and I'm going to generate that new assembly for the project. Now, this will probably take a while to run.

I probably should have run it in ChatGPT and put it in fast mode. Basically, you can just see as it's working here, it's actually going and extracting information from the project documents to go and customize this template list to my project. Okay, so 6 minutes later I said 6 minutes later, it's still running. Unfortunately, Opus particularly I notice is just a very slow model, but what it is pretty much going to be doing is it's going to go through and add to this library of assemblies.

And once it's done, I should have in my assembly list a complete list of everything I have to count on the drawings, which leads into the next step of the process where we actually measure. Now, maybe I have a disappointing perspective on this, but often I find the actual acts of counting and measuring stuff often isn't the slow part of doing a quantity of of doing quantity data. I often find once you have all your assemblies set up, the act of saying uh we're going to count the C1 light fittings, creating your assembly, and then clicking everything on the drawing just doesn't end up taking that long.

Like to me, the hard part is all that stuff around it, like actually building and customizing the library of assemblies. So, the actual measurements itself is often something like and it's unfortunately the part that AI does at least accurately at the moment, but is actually something that I personally think doesn't take that that long. If you do it yourself, you know you're going to be like more or less 100% accurate.

You can get Cloud to do the actual takeoffs itself. Unfortunately, you can't do that via the easy takeoff MCP yet, so you'd have to do this on a local set of PDF drawings. Although they are actually releasing a feature that will help do it. If you are going to get AI to do your takeoffs, there's really three things that I find improve the accuracy a lot at least for what I'm experimenting with. So, number one is just putting in place a reasonableness check.

So, making sure that if for example, you explicitly tell it and this is something I found that seems to prevent a lot of errors is I always just give it some general rules around doing takeoffs. Say for example, it is measuring the length of cable tray. What it will try to do is create a like extract the vector of this information on the PDF. It will try to say it will try to scale the drawing and get the exact measurement.

What I find sometimes if you run that process a couple of times, they will give you wildly different results because clearly somewhere along the line it's making a silly mistake. I find just getting it to do a reasonableness check where to say that for example, like sanity check the cable tray. Say there's three lengths running vertically along the building. You know from the architectural drawings that this like the distance from gridline B to F on a scaled drawing somewhere else or maybe from an explicit measurement if there's something else in the architectural drawings.

If you know conceptually that that length is 15 m, then this should be roughly 60 m because there's four lengths. And I always make sure in my instructions that I always make sure I want it to be roughly correct rather than precisely wrong. So, that's strategy number one. Strategy number two is to get it to check the drawings for the quality of the underlying data in the drawings. So, some drawings are compressed image files.

Some drawings have a vector layer of data on them where you can highlight text. If you can highlight text on the drawing, it means doing something like counting it can do very accurately because it can actually go and extract the text tag C1 and then write a script to count the number of instances of C1. It can do this almost 100% accurately, but if you have a set of drawings which are a compressed image file, so there's no actual layer of vector data on top of these, then what it will do is it will try to compress the symbols and just as an aside you can see it's starting to import my assemblies here with my breakdown of it that I can actually go and start using, but sorry, that was just about the assembly line, but what you find is if you're dealing with compressed image PDF drawings, it cannot count them accurately because it is trying to screenshot small sections of the drawings and distinguish between symbols, which if you actually look at how dense this section is, it does it tends to not do it very accurately.

So, that's why the third strategy, so the first strategy was getting it to just do a reasonable check. The second strategy was getting it to look and understand the quality of the data in the in the actual drawings and say whether or not it can do the take off accurately. The third strategy I find is to actually do a QA pass on the drawing. So, you get Cloud to go through the drawings and check things like the scale is 1 to a 100, that the actual page is the right size, that the drawings are oriented correctly.

Like, this was super interesting, but I found if you have a set of drawings and the drawings are incorrectly oriented, often AI will still try to do the take off on the incorrectly oriented drawings and it massively drops it massively drops the accuracy. So, we typically recommend running a drawing analyzer workflow, which basically goes through and does a QA pass to the drawings to check for errors and issues like this.

So, we are going to be releasing some more testing soon on the individual strategies and how they can bump up the accuracy, but we just haven't finished the testing on that. But, that is what we've actually found works. Now, the final I guess this is a bonus strategy. So, we have the first three from proving how well it does a takeoff. The bonus strategy is actually when the model is unsure, it blows up sections and gets it gets you to verify its count.

So, it returns a document says, "I count six power points in this screenshot. Can you please confirm if that's the correct?" Because we know in certain situations it would be out of count. Like say if we have a um a symbol that's all on its own, it would be out of count that relatively accurately. But, if you have a dense congested symbols like this, this is where our files. So, where it knows where it's not going to be able to count them accurately, it actually returns that to you.

Okay, but as I said, the easy takeoff MCP is now currently just text-based. So, if you do actually want to run an AI takeoff, you have to use it on a set of PDF drawings and get AI to read the PDF drawings. And if you want to do that, those strategies will help. As I said, we will release some more testing on this specifically. Now, one comments I would make is I do like while I wouldn't recommend do using AI to actually do the full takeoff for a lump sum job, there are situations where I do think getting AI to do a takeoff can be quite useful.

This would be if you're doing a budget price, maybe you're checking a client's bill of quantities, you're checking um you're doing a supplier cross-check so you're seeing if the quantity they've quoted makes sense. Even to an extent scheduling where you don't have to be 100% accurate for your like your quantities to get an indicative project schedule or even to just more or less roughly estimate the duration of activities.

Tasks like that, I actually think using AI to do a quantity takeoff, making sure it does a reasonableness check and a sanity check, I think is actually quite an effective strategy. What I'm talking about here is if we are preparing a lump sum estimate for the project, I would not recommend getting AI to go and measure all the cable tray lengths. So, I think you're going to be very disappointed when you under quote the job.

Now, the final part of this is actually taking the quantities you've measured and actually extracting them and moving them into a format that is useful. So, say I This is a very very simple example to show, but this would be like the first thing I did is I asked AI to give me some of my quantities. You can see it can extract the quantities and now I can begin to work with those quantities. So, I could turn this into a workflow where I say, "Take those quantities and input it into my Excel estimating sheet." And then it's just taking the quantities and moving them into the format you want.

You could say, "Take this bill of quantities and use it to populate my subcontractor pricing schedules." There's really a hundred different ways you could do this. I guess at the end of the day, what I'm saying is the opportunity for AI to help with quantity takeoffs isn't to do the specific counting and measuring, it is to help with everything around it, which is often, in my experience, the time-consuming part of doing takeoffs.

Now, the final recommendation I would make is package all these workflows into skills. We have a construction takeoff skill that we use when we want AI to actually do a takeoff. We have an assembly skill, which is where we want AI to customize our standard list of assemblies. Or I would like it would be you'll save a lot of time re-explaining yourself every time and you'll increase the right reliability of all these workflows if you take them and you package them into a standard set of workflows.

And then you just run these same skills on every single project you do. But as I said, I was doing this demonstration in Zizek's Takeoff. If you do want to check it out, I'll put a link for it in the video description.