The Future is Now: How Contractors Can Leverage AI for Practical Outcomes

Hello, and welcome to this webinar, The Future is Now: How Contractors Can Leverage AI for Practical Outcomes. This event is brought to you by Engineering News Record and is sponsored by HCSS. Hi, I'm Scott Seltz, publisher of Engineering News Record and your moderator today, and thank you for joining us. Our presenters today are Allen Hurst, Senior Director of Product in HCSS, and Joe Quattrochi, Director of Quality and Process Improvement for the Walsh Group's Heavy Civil Division. Now before we get started, please take a moment to scroll down and explore the webinar console. You can download handouts, click on the speaker images to view their bios, and submit your questions or comments anytime during today's presentation using the Q and A box. We'll address as many of the questions as possible at the end. Today's webinar is being recorded and will be archived on enr dot com. Now, I'm excited to turn it over to Allen and Joe to kick us off. Welcome, gentlemen. All right. Hello. Welcome, everyone. Great to be here. Joe Quattrochi, like I said, Quality Director and Process Improvement Director for The Walsh Group. Happy to be here. Allen, go ahead. Yeah. I'm Allen Hurst, senior director of product at HCSS. I've been in construction technology for about ten years and software for most of my career. And so we've been doing a lot of AI research and development at HCSS. Stuff's really top of mind for us. I'm happy to get started with it. You can always ask questions, and we'll address those later in the presentation. So getting started, we have four topics today. First, we're gonna talk about the three pillars that determine whether AI initiatives succeed or stall. Second, we're gonna talk about how we move from experimenting to operationalizing and using AI in daily workflows. And third, we're gonna talk about what AI agents are, how they work, and how you can deploy them responsibly. And lastly, we're gonna talk about how those systems can improve over time. But we're gonna thread this conversation through a maturity model. So I'll reference back to this periodically, and the progression of the conversation will kinda start at one and move up up the stack there. So we have five levels here. Level one is aware. So your teams are reading about AI. Maybe you've attended some webinars like this one, but there's not a lot of active usage. Level two is experimenting. Someone on your team is maybe testing ChatGPT or another LLM on a on a or testing a particular vendor feature in a tool they already use. It's ad hoc driven by individuals that have passion for it, but not necessarily an organizational strategy. So that comes in in level three where we're operationalizing. So that in that stage, AI is embedded in specific workflows. Leadership is behind it and encouraging that to grow, and data quality is being assessed and addressed. The next four next level is scaling. So there we have the company would have multiple workflows in production and hang agents handling multistep processes, data flowing across systems. Level five is where we're optimizing. So we've got momentum around our AI initiatives. We are getting success in outcomes we expect, and we've implemented some continuous improvement and using it hopefully as a differentiator, for your company. Most of the companies that we've looked at or talked to have really been at level two at this stage. Some were level one and very few kind of above that level, which is consistent with, some industry studies. Although, there was a recent study by RSCS that showed that about twice as many companies are in level twos as they were last year, so we're definitely seeing an acceleration of adoption. And, this model this model is not a judgment. It's just a way for you to assess where you are and maybe what the next steps could be. So I'm gonna let Joe kick us off because he's got real experience with us in a real construction company. Thanks, Allen. So with kind of that progression of the different levels that you're at, Allen kind of mentioned that a lot are at that three, four, five level. And a lot of that is kind of tied back to the three pillars that really is what's needed to take AI or really any initiative to the next level. These three elements, they come from Microsoft, read from their AI implementation guide, but something that we really see with any type of big innovation that you're trying to roll out. It starts with leadership all over the top. Do they have the buy in? Is the financing available from that side? Do they actually understand it and are willing to help push it along? Two is that technical readiness. And we'll talk a little bit more about this later, the governance, data security, that's really important element and really a big issue of why what holds companies back from making that next step that maybe they're ready to start taking the next improvements of what AI can offer, but some of the pitfalls that are out there need to be handled first from the technical side to make sure that it's not going to falter once you get there. But really the big one is that change is hard, and especially when talking about AI, this can be revolutionary for the construction industry, we're already seeing it happen in other industries. So the importance of having those champions out there, trainings and feedback loop is really key and really something that we're striving for ourselves to really ensure that it is a successful initiative all the way around. So that kind of leads to a poll question. I don't if you wanna bring that up. Yeah. So so we're going to let each of you let us know which of these three pillars is the biggest hurdle in your organization today. Just help us just know who we're talking to and give us a sense of how much detail to go into some of the areas. Give you just thirty seconds or so. It looks like these are kinda stabilizing. Oh, a few more coming in. Okay. I think we get kind of the overall trend. So I'll I'll stop it there. So looks like most of us are kind of at a technical readiness, which is interesting. We'll talk more more about that in a minute, like Joe mentioned. But you might, this might hopefully, it may cause you to reassess that a little bit, whether or not that's really the the next thing that's holding you back. We go through in q and a a little more later. Right. So, yeah, go ahead, Joe. Tell us the story. So from a leadership eye, so, know, really it was interesting from the poll that shown that, you know, under ten percent seeing this being more of a things holding back, and that's great. The importance of defining the outcomes of what you're trying to get out of it is key that I think was a big lesson for ourselves. We initially brought AI to our leadership back in twenty twenty three. And at that point, was hit with headwinds. It was, you know, there's too much uncertainty at that point in time of what it really was. The tech really didn't show that it was fully ready and we didn't really fully understand at that point what it could do. We just knew that it could make a lot of change. But it created our initial dialogue and that was the important aspect of it so that as we continue to investigate and look into where things were going with the technology in that and start piling some small little things without putting much investment into it, it allowed us then at the beginning of twenty twenty five to sit back down and have a full on conversation with leadership. And at that point, they kind of already had the background of it, a lot more information in the news, and we were able to present more of a story of what those defined outcomes would be and how we could get there and what it was going to take. And that really helped them get their buy in for us to start taking those next steps. Okay. The next one is technical readiness. So this was a popular response, but, so this really one of the backbones of this is that AI is really only as good as its inputs. So we really need to start with data quality. So if you're using HCSS and, do your activities in HeavyBid conform cleanly to cost codes and HeavyJob? Like, are your activities consistent? Do you have a consistent code book? Heavyjob, are your cost codes? Are they structured to them? It makes sense, so that we can compare, like, activities across jobs. And that standardization is where things break down for a lot of companies. If the field's tracking work, then the estimator did it, then you lose a feedback loop between what you did and what you built. And AI can't necessarily learn from that history if the data doesn't line up and it can't make the connections. So we we've seen, there was a bridge there was a bridge analysis that said that thirty percent of our construction firms or more than half of their data is unusable, and forty five percent had no formal data strategy. So so things like that or if your organization falls into that category, then you probably got some work to do on the data side before you're gonna be able to really take full advantage of the opportunities that AI is gonna give you. The next area is, system consolidation. Every data silo where data lives in totally separate system from another kind of data, it represents a blind spot for AI. The industry's getting better about providing opportunities for, like, off the shelf products are getting better at integrating with one another with regard to pulling letting AI associate data from different sources together. But still, when you're estimating data lives in one tool and your field data lives in another tool and your safety lives in a third, you lose the ability to make connections across your workflows. So fewer disconnected platforms are are inconvenient and less valuable. And the last is infrastructure. So the we we have moved over the last decade or more into cloud first and mobile forward, workflows, but the our our construction industry is is has some special accommodations that we need to make. We still have many job sites that use our software that have that don't have reliable interconnection Internet connection. There's offline fallback. So the strategy that you pursue has to be informed by the environment that it's being performed in. So that can make some construction projects difficult, from an AI perspective since most of it is cloud based now. Other than Allen, I'll add on the last slide, just as we've been making our transition and fully rolling out, one of the pitfalls we realized is the language model had too much data that it had accessible. In our case, it had twenty years of SharePoint sites that it was pulling information from. Much of it was obsolete, much of it's just the permissions weren't quite set up right, so we had to spend a lot of time kind of cleaning that back end up just to limit what the language model was ultimately able to utilize, and that helped vastly improve the data quality we're seeing on the back end. So that all kind of ties into that type of readiness side effects. So this one kind of makes or breaks everything, the human change management. Folks in construction industry are not or maybe skewed to be not early adopters of technology compared to other industries. So there's a skepticism, and AI breeds skepticism across industries. And construction technologies earned the skepticism. The industry has thin margins and high stakes and, honestly, a long track record of tools that are promised more than they've been able to deliver. And so what what tends to work is to start with the pain, not the tool. Ask your foreman or your project managers or your office people, like, what what is the what is the activities that what are the activities that frustrate them? And then, brainstorm ideas for how AI might might address that specific thing. Once it's working on, real projects, not a demo environment, we can start to find people who are already experienced, give them a role, and reduce friction from there. So every extra click that a person's going through in a day is a reason to assess, you know, is there something we can do to automate here? Yeah. This is the area that I spend most of my time. Consistently hear the stories of, oh, I used AI once and it gave me the wrong answer, so I haven't looked back at it. And that was two years ago. So it's taken a lot of our time going to job sites, Allen, like you mentioned, talking to superintendents, project managers, asking them those pain points, and then right there in front of them showing the solution work in front of them. And so here's your question, here's your resources that you want to tie to your spec or contract, what have you, and show them how quickly it could be done and the benefit of the AI. And whenever we show to them, it's mind blowing to them and they're immediately, all right, how do I get that? How do I use it? What's the next step? So, you know, doing that kind of handholding is key to get that initial buy in, but it also can be a challenge, especially for your larger contractors, for ourselves with four thousand full time staff employees across the United States. So we need champions in each of the regions that are going project to project to kind of help show that and show those benefits. But then ultimately from there, one of the beauties of AI is that it can train people itself. As long as you show people how to access it and where to get it, it's as simple as telling them the prompt of teach me how to do this, tell me how I find more, what are things I can ask you to do? And some of those simple prompts will help them understand that the AI isn't just about doing things for you or writing an email for you, it can teach you how to do anything that you're trying to do that much better. Including how to use AI. So don't underestimate the value of asking an AI tool how to how to do something with it or how to how to use it. Absolutely. Really well. Alright. So, this helps to kind of give a simple rubric that you can use. So if for each pillar, you'd ask yourself, are we at level one, two, or three, or above? So if leadership hasn't funded AI or it's not, the experiments aren't tied to business outcomes, then maybe that's a bottleneck. If your data's inconsistent and your systems are disconnected, that's a bottleneck. If there's no training plan and your field teams aren't involved, then maybe change management's the bottleneck. So, the tech, we've seen is sometimes less of a constraint than we than initially thought when we start to really dig in to what's holding these initiatives back. So finding the starting the the the lagging pillar, the most lagging pillar, and starting there is a a good idea to help prioritize these things. There's a poll up here now. So, we're gonna go ahead and let us tell you where you think your organization sits on the maturity model, based upon what we've discussed so far. Who's ever at that level five needs to jump on and start talking. Yeah. I'm gonna hand the presentation over. Right. Okay. Looks like we're stable now. This is very consistent, I think, with with what we've seen in other audiences. And, honestly, it's not limited to construction. I think a lot of industries will see the same distribution. AI is new. It's changing a lot. It's, it's difficult to to scale and adopt as quickly as it's changing. Alright. Okay. Operationalizing AI. So this is the transition from level two to level three in the kind of in that maturity model we've been looking at. The shift is it's straightforward. Experimenting means, someone on your team is using ChatGPT on their phone. And operationalizing means, the AI is embedded into a daily workflow, your team expects it to be there, and they would notice it if it were gone. So you and, hopefully, you can point to measurable improvement, that you've seen or or not, with AI. And so the the here, we also see, we see more companies operationalizing, but still not necessarily it really follows the distribution or less poll. Joe, you have, comments on operationalizing? Yeah, I mean, really, you can go to the next slide. Mean, really, it's making that transition is very hard. It's finding the right tool, the language model we want to be able to utilize, how to implement it, making sure you have that technical readiness change management, all three of those pillars ready to go. But really as far as where to start, that's a big question because AI can do so much, but what is it that you really need it to do is the big question you need to answer. And here's some good things to look at, where do you have good data? Where do you have repeat tasks that keeps happening day after day? And something that is potentially boring that people don't want to be doing so they're going to want to change into utilizing a tool. And a big thing is where is it also a low cost of error, especially early on when you're starting to implement something that is very key. So writing emails, yeah, sure. Drafting a plan that gets redone on every job, okay. But doing structural calcs, probably not the best idea to start off. So finding that right use case to start or, you know, handful of ones that you want to focus on, that's really key versus trying to eat the whole whale at once. That's gonna be a lot better way to try to make that transition. It's not that you need it in every workflow, in every process for every user all at once, it's finding those few key things that's going to make a measurable impact right from the start. That's where you focus and that's how you'll build that momentum. And then with all that, it is defining what success looks like. So that's whether it's success or the goal that you have for each of the individual use cases of how that's going to improve that workflow or reduce errors or speed up that process. A lot of things that we're looking at, it's about how are we giving time back to our people so they can put more time into being out in the field where the work's actually happening versus doing busy paperwork. One of the specific areas that we focus on is for all of our activities, we do construction work plans. And it can be a pretty tedious process of drafting up for each activity a work plan that has all the right details within it so that the crew can actually follow and build the work safely, correctly, and productively. So we've created a scoring metric that actually scores our work plans. And now with the AI inclusion, we're able to use that, utilize that metric to kind of give ourselves a goal of, well, if our scores are currently seventy five percent using AI, we expect a ten percent increase in our work plan scores because we're going to have this AI improvement, but at the same time, it ultimately is going to then provide improvement to the safety, to our quality of our work productivity, and give that time back. So some of those are very easily measurable. Some of them are going to be more just through feedback loops of how we're able to pull the information and see how much time that's provided back to our field teams. So I want to just kind of highlight the metrics are important to understand where we're moving and if we're moving in the right direction and and have something objective. But it's also just high level important to set that outcome. Like, we we will know if this is a successful effort win. Right? And sometimes it's a bigger picture, maybe a more inspiring message than we're gonna save you an hour a week or whatever it is. So some specific use cases, mentioned the construction work plans, another big one for us is submittal reviews. Each job has hundreds of submittals that they have to go through. And while we're not ready to fully trust AI to fully review the submittals before passing them on, they do a great job at a first pass. Where do we really need to dive in? What are obvious things that we're missing? And that's already provide huge benefit of finding errors faster so we can fix the submittals before we even get them submitted over and help us stay out of potential issues. So that's a very simplistic one of, hey, here's our submittal, here's your specification we're reviewing against, compare the two, provide us a checklist on the different elements and what's good, what's not, what we need to review further. So it doesn't eliminate the need of doing the review. It just helps speed up where to look, where to find things versus just going through hundreds of pages and trying to figure all out on your own. Another simple win, the last one on the material tickets, that was a huge simple thing of just creating a simple AI agent to decode handwritten truck tickets that a lot of our projects have, and they'll have a hundred truck tickets a day, decode them and put it all into a Excel tracking log where it used to be a manual boring process of a project engineer taking each ticket and manually typing in all the information. Very simple, just scanning it, loading right in the agent and deciphering it and even tell you, here's things they have high confidence in, here's things where the handwriting wasn't great, but based on past tickets, this is what it seems like it's insane. So another quick win for something that was able to help take repetitive, low risk thing that people didn't like doing, having AI do it was a big one to get their buy in and help them start utilizing it for other things. This is one of the first areas. There are a lot of documents in in construction workflows. So and AI performs really well for tasks like that. We see a lot of customers starting here, and we're starting here. So we've looked into how do how do we pull how do we use AI to make the material workflow that Joe mentioned? How do we take that out of the office and put it in the field even so that when that's captured, that those materials are, you know, are entered into as a cost at the point of entry for every other cost you have. And, also things like you know, there's a lot of information in equipment document. There's a lot of information, obviously, in a bid lighting documents and things like that. So, an h the HCSS Copilot already lets you upload documents today and ask questions about it and, try to cross reference that with your with the data that's in your system. So we're we're finding new and better workflows for this all the time. We're trying to make sure we can systematize that and put it in the tool so that, not everybody has to build the the tool that that Walsh built, for example, for material tickets. Next. So estimating is probably the number one area of interest in some of our polling on, how AI can impact, construction companies, especially in their outcomes. And, we've seen that maybe a quarter of companies are starting to play around with this in some way, shape, or form. So today, the AI assistant, bid document processing, like I mentioned, is a is a key starting point. Bid packages are hundreds of pages, so AI can extract key data out of there and prefill estimate inputs, just reduces the amount of time estimators are spending on data transfer and data entry. Preconstruction professionals spend a lot of time research and anal researching and analyzing data. So here's an opportunity really to optimize their time, give them a lot more time back to do something that's higher value. Data sorry. Go ahead, Joe. Did you have a comment? Oh, okay. So we're also seeing, companies use AI for pulling different data sources together. If there's d r DOT data or historical production data or past estimates from similar work, AI can really help to make connections between all of those and help to advise, to bring that in as input to your next estimate. At HCSS, we're running a lot of experiments around how AI can assist with these different workflows and produce outputs that have we have high confidence in. That's that's where the difficulty is. Some companies that have really good data, we can get really high confidence. Custom companies that have the data that's not as good, it's a little lower confidence. So, we're hoping to as AI gets better, as experiments get better, as we have more DOTs that are exposing more data, like, there's this will only get better over time. So it's something definitely to keep an eye on. And I wanna say, like, our goal is not to replace estimators. Our goal is to free estimators from having to do stuff they don't wanna do. Right? It gives them more time to bid more work. Okay. Safety also has a lot of opportunity to it's an area where construction companies can apply AI. It's not as concrete, kinda, where we stand today, but, really, we wanna shift from as much as we can, even before AI, from lagging indicators to leading indicators. A lot of what we have in HCSS safety, the modules that are not incidents, are are are there to try and implement a safety program that, makes, like, leading indicators visible to those that can do something about them. So we wanna so, traditionally, safety management is kinda reactive. When something goes wrong, it's reported, it's analyzed, and we try to prevent it from happening again. Ai can help us to flip that more to identify things, trends that we're seeing or risks and mitigate them before they turn into incidents. Yeah. So of the this is one of the areas that we're really trying to figure out how to best implement something. Our kind of the aspirational goal is, right now foreman walks on the job site with their iPad in the morning, you're ready to fill up a pre task plan and type in the main activity they're working on. But we'd love at that point for the system to know, all right, well, here's the job they're on, here's the area on that job they're working, have it look at the data already in the system, all right, well, here's some recent issues from inspections of that or things that were brought up from that job, from that specific location, But then here's also, hey, they're working on forming up a twenty foot tall wall. Well, here's projects across rest of the company where they've had issues, whether it's safety issue or quality issue, equipment issue that ties to that work. Here's some warnings that are recent or what have you that it would automatically tell them, here's the things to be aware of for that work and not just be a pre populated standard list, but here's real issues, real lessons that tie specifically to the work you're about to do that they're able to see before they're doing the work without having to search anything, without having to type anything. It just automatically comes in for them. That's the goal that it helps take all these leading indicators and actually puts it into the people that are about to get ready to do the work. So they're ready to go and the train in front of their face, they can actually use it and leverage it versus trying to spend time trying to figure it out because they don't have that time. They got to go build the work, train their people, make sure everyone's doing their job. So that's the kind of aspirational goal we're looking to get to. Very much looking forward to how that shapes up. So, like, obviously, we want to keep people safe. It's a core part of something we wanna do, like, every single day. That's the most important thing we can do. But, there's also a legal dimension emerging. So there's legal experts and have been advocate or speculating that maybe there's an argument that firms who don't adopt AI practices and these specific tools actually face liability after accidents, because it's an available mitigation technique that wasn't pursued. So, it's it's a productivity improvement and a safety improvement and maybe a risk mitigation requirement as well. Distance versus agent. So even among people that know that have done a lot with AI, there's still disagreements and misunderstandings about what agents are. So, talk about agents just a little bit. And if you disagree with what I say, you can ask a question. So AI agents, this is really a transition from level three to level four on that maturity model. And a distinction that matters most is that AI tools in construction today are distance. Like, you you ask a question, you get an answer. Maybe it summarizes a document. Maybe it generates a draft. An agent is a little different. An agent pursues a goal across multiple steps that has some level of agency or, you know, autonomy. An assistant, can ans if you ask an well, if you have an an agent, and you ask it, does the submittal meet the spec, an agent might process the submittal, check it against spec, generate a compliance report, route it for approval, log an audit trail. Whereas an assistant would, answer, does it meet the spec and give you a really straight you know, really simplistic answer. But it it's, it's a it it has its own workflow, but it also acts within defined boundaries. But the caution we raised in, we had a session before that AI agents are a panacea. They're they're not. A lot of what vendors are marketing as AI agents are really rules based automation with a new label. It's it's it's, common to see even features that already existed in products be be relabeled as AI features or, features being relabeled as agents. But a real agent, will can adapt when an encounter is something unexpected. It improves with feedback. It's not it's not just a glorified if then workflow. Right? So it's a fine line, but but there's definitely a distinction between what a AI agent is and what a automation with a very defined path is. But you can't talk about agents without talking about context. So the problem, if you upload a document to ChatTPT and ask it a question and got a confident but wrong answer, or you could unload it you could upload it and get a good answer once and not replicate it, The reason is frequently context. The AI doesn't know about your project. It doesn't know your specs. It doesn't know your standards. They worked with whatever you gave it plus whatever it already knew from the Internet. And so the how how we fix that and give the agent the right information at the right time is where context engineering comes in. And context engineering means controlling what the AI sees so that it produces reliable and relevant examples. So there's, talk about this in three pieces. First is we need to isolate the right data. We need point the AI at a specific project specs, specific drawings or contracts, and not the open Internet. And so without this, it'll find whatever it confirms, answers the question it has even if it's not the best answer. So giving it a specific corpus of data will allow it to use will come to the right conclusion more often. Second is to help define the roles. There's a big difference between telling it to check the submittal and to review the submittal against a specific section of a specific job, you you know, on a specific spreadsheet, right, or or a contract. And giving that specific specificity determines the quality of the output. And third is, structured knowledge, knowledge that, the AI can understand in detail. So things like historical data, company standards, approved material lists. The more structure and context these documents have in them, the more structure and context, the AI can interpret from them. So the the best AI implementations don't put every single document into one AI system and have it sort out the pieces. Kinda the a well defined context gives the AI a map of where to find things, and it pulls in what it needs for the right for the task it's executing. And it's an important distinction. You don't wanna dump your entire project drive into an AI, and maybe you start there, but, ultimately, you want it to know where to look and what to retrieve and, and some rules for how to engage with that information. And there's there's AI topics, like skills and things like that that we could go into maybe in a future session that can go tell you more about the how. How do you implement a map like that? So this is getting a lot of attention right now. Anthropic has published engineering guidance on it. The principle for, for today is you need to tell the AI exactly what job it's doing and to the extent possible how it should do it and what documents to use and what standards it needs to apply. And so, you know, what you come up with in five you know, thirty seconds into your ChatGPT prompt might not be the thing that you wanna rely on. You may wanna put a little bit more thought structure into it from there. Okay. So once we've built some agents, designed some agents, I wanna talk a little bit about agent deployment. So once these agents get deployed, the governance is not optional. We need to define, these three things before deployment, not after. The first is data boundaries. So what what can the agent access? What's off limits? You don't wanna submit a review agent pulling from a different project specs or from the public Internet. Next is action boundaries. What can it do on its own and what requires a human? Document processing may be mostly fully autonomous, but something with a financial or safety liability definitely needs a human checkpoint. Another is audit trail. So every decision that the AI makes needs to be log traceable and reviewable at the key steps in that workflow. Compliance compliance is important, but it's not just for compliance. The the log is also how you improve the system over time. It tells you where it's reliable and where it's making conclude drawing conclusions that are that are erroneous. So that audit trail, it helps you to understand how to improve it. And data security. Data security is a top concern. We hear from enterprise contractors and smaller contractors too. They should really ask every vendor four questions. Like, where is the data stored and who can access it? Is our data used to train the model? What happens if there's a data breach? And how are all these decisions around this logged? So, if if anything is in a free tier or public model, it's probably trading data. So that's another thing to keep in mind as you engage with these tools. If if you don't have an enterprise agreement, if you don't have something that's written down by somebody, you have to assume that you're exposing that data unless you can confirm otherwise. And it's also how the models are free, and they don't out of charity. Right? So this is a key thing. Like, if you don't get a lot from this, I hope you can take this away. This applies kind of at level four and five, but a lot of the steps before that are important because of this. Like, until you have good data, the rest of this doesn't work and, you know, and throughout. Right? Throughout, the it breaks down. The better data produces better AI outputs and better outputs build trust, and trust drives usage. More usage generates more data, and then the cycle repeats. But more data is not necessarily better data. Dumping everything into an AI system can make it less accurate, not more accurate. Right? We talked about that a little bit earlier. The goal is structured relevant data, not all data. So these things work well clockwise, but they also work counterclockwise. So bad data produces bad outputs. Bad outputs destroy trust. There's no trust. There's no usage. There's no usage. There's no improvement. So you can see organizations kill good AI initiatives quickly because this cycle starts to break down. And it all goes back to those three elements. Leadership, technical, and the change management. If you don't have any, you know, one of those fully working, the system breaks down that much faster. So vitally important. Okay. So we have another poll here. So which piece of this cycle holds your company's biggest opportunity to advance AI adoption? We'll give, like, thirty seconds for folks to answer this one. Okay. Yeah. Not surprising. Over the last couple of years, it was all about getting data through different systems, and now it's the need of making sure that data is actually good and useful. Yeah. Recent there. Okay. So fast the the biggest response was better data, and, yeah, that's that's consistent with what with what we see too. That's consistent with what we've seen in our own r and d efforts on this. Right? I mean, it's, there are things we've been able to come up with that work really well for folks that have good data. Right? So that's but not for others. Okay. Feedback loops. So we just talked about feed loops, but, there's some practices, that can help or guidelines for practices that can that can help with this. First, we wanna capture corrections at the point of use. So when a reviewer edits an AI generated checklist, that edit good and could be a training signal, that feedback into your workflow. Second, is to track patterns. Like, what percentage of our AI outputs are accepted as is? What gets edited constantly? What's get rejected? And that could tell you where to improve your context in particular. Right? That's the number that's the first thing to look at is what kind of context do we give this engine, and is that the reason why the why the results aren't what we want? And third is to schedule regular user feedback. So not a survey, but an actual engagement with a customer, or sorry, with a user of, what's working, what's not working, what's noise, what's missing, what's helpful, if these questions help to get, to get better over time. Joe, do you have a comment on feedback in your experience? Okay. Okay. So what's what's coming? And we're running short, so I'm gonna try and go a little fast. Okay. Over the next twelve to twenty four months, some of the things we're keeping our eye on are, on device AI. We're eagerly awaiting to see what Apple and Google come up with as far as better APIs for doing AI interactions natively on the device. As I mentioned earlier, a lot of cloud a lot of AI workflows rely on cloud infrastructure for a poorly connected or disconnected job site that limits the amount of AI we can we can utilize. So we're hoping that, the mobile mobile platforms meet us in the middle there. Vendor consolidation. So, fewer platforms, are doing do more with less AI built in natively. So, less integration overhead then. So if you can get to the point where, you have fewer platforms that need to integrate, then that AI adoption can move move faster with less friction. And richer data platforms. So tools like HCSS Insights make it possible to pull from estimating operations safety and fleet into one view without, building your own custom data warehouse. So the the more you can rely on the richer data platforms, the more, you're gonna be able to leverage AI. Industry specific AI models. So it's, it's important what what we do that we we're training our models on construction data and as accurate as possible, so that we're not we don't we we have a better advantage than pulling from ChatGPT. Right? So it's not the general Internet that's informing its, you know, thirty plus years of estimating data. And, practical advice is to kinda stay close to your vendors' road maps. You're gonna see, you know, what's coming. Are they doing research in this area or not? And use that to forecast how it aligns with your own AI strategy. Is this gonna be a a friction to your AI strategy, or is it gonna enable it? Okay. Oh, we'll close we'll close, with this. But the advantage of starting now is not just about tools. It's an organizational must muscle that compounds. So starting early able even if the tools you use, the workflows you implement are not the first ones you start with because you'll you'll get better, and you'll you'll have information to inform your decisions as you make other decisions. And data literacy across your workforce is important. Trust in AI outputs is built through experience and interactions with it and reinforcement that these things are correct and valuable. Workflows, that improve with every cycle can can occur as a result in institutional knowledge that's captured in systems instead of walking out the door when people retire is also important as we see our the workforce issues in the industry. And AI can can help with that too. So closing up, let's look back at some maturity model. And so what can we what can we do now? That's kind of the, you don't need to be at level five tomorrow. Right? You can start with where you are and find the next best thing to do. We can look for specific actions. Level one or two, we might wanna audit our top three most manual data heavy workflows and have a leadership conversation about, what an improved outcome would be. Then identify champion. You can do a pilot. Level three, you might define an outcome statement for each AI use case that you have now and review the data quality, ask if the data's standardized, or your system's connected. So you can evaluate one workflow and identify what needs to happen for an agent to automate this thing. Then if you're at level four or five, start implementing more structured feedback loops and build a governance framework and talk to your vendors about their roadmaps and how it aligns with your strategy. Joe, you have any other comments, closing comments? I'm sure like a lot of you, I get calls from AI vendors pretty much on a daily basis. There's a lot out there kind of to what Allen mentioned, going see a lot of consolidation in the market. A lot of the big players are building AI just natively into their apps like HCSS and a lot of the other vendors. And so really the people that you already utilize are probably the best ones to focus on their roadmaps, understand where they're at and look into more of the general large language models, whether cloud or copilot or what have you and kind of look to build in yourself because I see a lot of these vendors slowly going away as the system start building up themselves. But kind of like to this point, I know a lot larger in that level one too, just start small, find a couple small scale tasks. Don't look for the giant win right off the bat. Find those small wins and start building that momentum that's going help you in the long run. Okay. Yeah, I think we're ready to hand it over to Scott for questions. Yeah, thank you, Joe and Ellen. This was a great presentation, and we received some really great questions, and we've got about nine minutes to go through these. So a question that I thought about when I was listening to your presentation, we often hear that AI is going to replace jobs. From what you've seen at HCSS and the Walsh Group, how does AI actually change the daily life of a foreman or a project manager for the better? I'll start with this one. So AI is not swinging a hammer, running any equipment anytime soon. So from that side of it, in fact, I think it's going to help bring more workforce into the trades hopefully because we still are dealing with that skill shortage nationwide. But from the project management side of point, this is often asked, am I going to be able to do the work with less people? Yeah, maybe someday, but in reality, right now, all it's doing is changing the workflow a little bit. I was just listening to a podcast this morning that even software coders, that it's not necessarily replacing them, it's just changing them. They're having more conversations. They're able to provide a better quality output. And that's, I think, what it's going to do for our project managers and project engineers and estimators. It's going to be better output, higher quality, but it's also going to provide time back than having to redo things and put more effort and emphasis into the fieldwork. Nice. Allen? Yeah, I mean, I concur with that. You know, AI is not swinging a hammer. And so, like, I think field workers today are are hopefully, hopefully getting value from AI and getting more information to them at the right times or or packaged and analyzed preanalyzed for them. Right? So they're saving saving time doing this so we can get them out of their iPad or out of their laptop and back back on the job site. I think that's that's where we hope the performance comes with. Less data entry and more data capture. So I I think that's the direction you go as far as field goals. Those that are in the office, I think you see a lot of the same changes as you see in other information workers. Right? Their their document their drafting of documents or processing of spreadsheets or aggregating of data. Like, all of those things get easier. So Great. This viewer is asking, what level of diligence and QAQC for the outputs created by AI is the benchmark? It's a great question. What level of? Yeah. Go ahead, Joe. Yeah. I was just gonna say, this is one that at least from my end, it obviously depends on the risk of what you're ultimately asking it to do. Again, if you're asking it to do structural calculations, yeah, you better be doing a whole lot of QA well before you ever roll that thing out. If you're ultimately just having to do something, someone else can be checking later, obviously much less. But right now for everything we're doing, we flat out have said, do not trust the output of AI. Everything needs to be vetted. Can have it write you an email, write up a plan, but you have to sign it, so you better know a hundred percent for sure what you're signing is right. Yeah. Depending on the risk of what you're doing, it impacts a lot, but it's, I think you you use AI for small pieces of the workflow, right, first. And so then you're still having to enter you're still having a human intervene at each step. And then maybe over time, as you get more and more confidence and see a higher level of accuracy or you have the right controls in place, maybe you automate a few more steps of that before there's intervention. I think, ultimately, there has to be human intervention along the way at whatever points the workflow, you know, supports and depending on the level of risk involved. Great. This viewer is asking, given our well organized database, we are interested in exploring how to leverage AI within HCSS heavy bid for takeoffs and the population of activities and resources. Is HCSS planning to integrate more AI driven tools directly into the software to further automate of takeoff input? I can't speak to a specific initiative around takeoff input in AI and heavy bid, but, we are doing we have several experiments going on in AI improvements in heavy bid. I don't wanna get ahead of any announcements, but, that is, like I mentioned earlier, estimating is in a position right now to be the most, like, affected by AI. And so, obviously, we're putting our efforts in there. But takeoff takeoff is not something that in the next three to six months that that we're going to be using AI for. Okay. Hey. For a contractor who feels like they're still at level one, mostly paper based or using disconnected spreadsheets, what's the one small digital step that they could take today to prepare for AI tomorrow? So I I gotta go back to data, like data quality. I think that that's that's gonna put you like, that and learning or experimenting. Getting getting folks familiar and comfortable with the tools and then making sure that the data's in a in a state that it can that you can capitalize on that. And, like, enthusiasm that hopefully you get from the experimentation. That experimentation, I I would double down on that of, you know, even just some of the simple tools that are out there already. You have a ChatGPT, a Copilot, or what have you. As long as you understand the limitations and restrict from uploading things that you shouldn't be uploading that are going be public, being able to just experiment on what it can do, especially with those that are the younger generations, I'll guarantee you, they're already using it on a daily basis for everything. People don't Google things anymore. You AI everything. So just let them naturally do that. They're going to come up with ideas on how to do it. Just kind of have to give them a little bit of a leash to go off, but again, give them some guardrails so they're not uploading cost reports and restricted information up to these public platforms. But so give those guy guardrails, put them let them play with it, and you'll be surprised of the solutions they can come up with. Great. And we actually we have a show show and tell that's one quick comment. We have a show and tell every couple of weeks on our team just, of what you're playing with, just to spread the awareness of what's possible and what other people are doing. So it's a it's a good idea too. This viewer is asking, all the labor equipment materials loaded into HCSS, how actually will AI load transfer a bid activity to an estimate? How will how will AI load a bid activity to an estimate? I mean, some of that depends on the type of work, you're doing. So, if it's DOT work and you and it's a, you know, it's not a design build and you already have drawings, then I think, all of that we were doing experiments with bringing all of that together and combining it with production, history from heavy job or companies that have heavy job also, and kind of setting some initial costing, into the estimate, and obviously with a human involved in verifying and applying those. So it's we try to take the approach of here's some data we think. Here's what we think we've found for you, and then let the and the the human user say, yes. That's right, or no. This isn't right, or I wanna use this piece, but not this piece. So, yeah. But if you don't have if you're if you don't have as much input, then there's there's less we can automate. So a lot of it depends on the kind of work and the kind of inputs that you start with. Great. We got one minute for one last question. If a company only has the budget to solve one big problem with AI this year, like safety, bidding, or scheduling, which one usually offers the fastest return on investment? So those I haven't answered it. I think we've Yeah. Go ahead. Yeah. Go ahead. Okay. Well, I think it comes down to some of the things we said here today. Like, document oriented workflows are are a good place to start today, or in analyzing, or connecting data from different sources. If you've got a bunch of spreadsheets and you've got this other database, like how you pull all that stuff together, think AI is really good at that. So those are a couple that I think the technology is really well equipped to handle today. I was just gonna echo the same thing though. Cost, schedule are probably the three hardest things for AI to do today. I would start with the document processing side of things. That's where you can get the biggest benefit even though it's harder to understand the return on investment. That's where you're going get the biggest benefit and where can you work right now without the risk. Excellent. Well, that's all the time we have for questions today. So please join me in thanking both Allen Hurst and Joe Quattrochi, as well as our sponsor, HCSS. If you have any additional questions, please email them to webinars@bnpmedia.com, and we'll share those with our presenters so they can answer you directly. When you exit the webinar, you'll be taken to a brief post event survey, We'd appreciate it if you could take a moment to share your feedback as it helps us to continue to improve our programs for you. Please visit enr dot com for the archive of this presentation to share with your colleagues, as well as information on some of our upcoming events. Hey, we hope you found today's discussion to be a good investment of your time. Thank you again for joining us and have a great day.