Putting AI to Work: Practical Use Cases for Construction Safety
November 19, 2026
30 min listen
Podcasts/Putting AI to Work: Practical Use Cases for Construction Safety

How can construction safety teams move beyond the AI hype and start using it to improve safety and risk management?

In this episode of Beyond Prequalification, Highwire's Corey Silverman is joined by Erik Lenke, Vice President of Intelligent Platform and AI Capabilities at Veriforce, and Peter Chong, Senior Product Manager at Highwire, to explore practical applications of AI in construction safety and what responsible adoption looks like in practice.

From streamlining safety inspections with voice-to-text technology to analyzing contractor risk data and identifying emerging risk trends, they discuss how AI can help teams work more efficiently and make better-informed decisions. The conversation also explores the limitations of AI, the importance of human oversight, and why safety professionals must remain accountable for the decisions these tools help inform.

Whether you're an owner, general contractor, safety leader, or trade partner, this episode offers practical guidance on identifying valuable AI use cases, incorporating new technology into existing workflows, and balancing innovation with responsible risk management.

Meet the Guests
923125fa-102a-4474-9b3c-966005031c8a.png

Corey Silverman, CHST

Highwire, Safety Specialist

Corey Silverman serves as a Safety Specialist at Highwire, bringing a hands-on, solutions-driven approach to risk identification and safety management. He collaborates directly with clients to strengthen their safety programs, enhance job site safety performance, and promote a proactive approach to managing risk throughout the project lifecycle.

LinkedIn
1cbf0460-f927-4465-8d8a-f528c0a2d241.jpg

Peter Chong

Highwire, Senior Product Manager

Peter Chong serves as a Senior Product Manager at Highwire, focusing on AI and data-driven product development to improve construction safety and risk management. He brings a user-centered approach to product innovation, working to develop intuitive, practical solutions that address real-world challenges, streamline workflows, and help safety professionals make more informed decisions.

LinkedIn
0e047be9-4eae-48d6-b0da-0e14f2e46c72.png

Erik Lenke

Veriforce, Vice President of Product, Intelligent Platform & AI

Erik Lenke serves as Vice President of Intelligent Platform and AI Capabilities at Veriforce, where he leads the development of AI-powered solutions across the company's product portfolio. Previously Director of Product at Highwire, Erik focuses on translating emerging technologies into practical capabilities that streamline workflows, deliver actionable insights, and help organizations make more informed risk management decisions.

LinkedIn
Transcript

[00:00:13] Corey Silverman: Hello everyone, and welcome to another episode of the Beyond Prequalification podcast. My name is Corey Silverman. I'll be hosting today's episode, and today we're going to be discussing concrete use cases for AI in the construction industry. One thing I've noticed lately is that AI is no longer this brand-new, somewhat mysterious technology that people are still trying to understand at a high level. We've seen AI adoption across the industry for several years now. The conversation's starting to shift, right? Instead of asking, "What is AI capable of in general?" we're moving towards a much more direct, practical discussion about how it can be applied to certain and specific use cases. That's exactly what we're gonna be discussing today. Before we get into it, I wanna give my colleagues here that have joined me a chance to introduce themselves and their roles. I'll start with a little more [00:01:00] detail about my role. I'm a safety specialist with Highwire, and my background is as a construction safety manager. My degree is in safety, and I spent some time in the field in a variety of safety roles, including with a large GC. And so, my role with Highwire is really supporting both our internal teams and clients as someone that comes from the industry. So, Erik, I'll pass it off to you to introduce yourself.

[00:01:20] Erik Lenke: Thanks, Corey. My name is Erik Lenke. I am the Vice President of Intelligent Platform and AI Capabilities within the product organization of Veriforce, the parent company to Highwire. Prior to my time in this role, I was the director of product at Highwire, and now I have transitioned into a role where I specialize in bringing AI products and capabilities to life for customers, not only for Highwire, but across the Veriforce suite of products and companies.

So, I'm really excited to talk to you all about what we're seeing in the industry as well as what we're seeing in technology and how we're marrying those two things together.

[00:02:00] Peter Chong: Hi. My name is Pete. I'm a senior product manager here at Highwire. I've been around for about three years at Highwire, but previous to that role, I did UX design at IBM, and I worked at a few other startups before coming here. These days, I spend a lot of time on our AI and data work, and that's just because of how the industry's kind of shifting. The thing I'm always pushing on is: what is this actually solving? Is this solving a problem, or are we just adding AI because everyone else has it? There's a lot of sparkle buttons in AI. Just if you go to, like, any platform, and so the question is: Hey, how can we make something simple enough? How can we make something that safety managers or project managers actually trust enough to act on? And can they make a call with that AI product? And that is basically my goal right now.

[00:02:54] Corey Silverman: Awesome. Thank you both. So, starting off, like we said, AI isn't this brand-new thing in either construction or safety anymore, and we're starting to see it incorporated more and more in our day-to-day lives. What feels different to you guys right now? What are we hearing from safety and construction leaders about AI and the use of it in the industry?

[00:03:12] Erik Lenke: Right now, what we're seeing is a lot of expectation around access to information. That's not a new thing, right? But how that information is surfaced to the end user has changed. With the introduction of AI, people have less patience for digging through data. They don't understand some of the sort of complexities around why information is hard to find, and they look at AI as a simple way to basically bring insights quicker and more easily, and the expectation is kind of shifting for that to be the baseline.

It's less of an additive value or a unique differentiating capability, and now it's becoming an expectation across the board.

[00:03:56] Peter Chong: Yeah, I agree with that sentiment too.

I think a few things that I've seen are: one, everything is more readily available. Like, if you look in any of your platforms nowadays, they have, like, an AI product, right? And so, you don't need to create, like, a custom build. If you go to any one of these really big construction conferences, you'll see the company's name plus AI everywhere, and there is some AI exhaustion for sure. And so, the availability is a lot higher. There was an NSC survey that came out this year, and in that survey, it said 80 percent of safety leaders are using AI in some way. And so, because that's the case, right, I think there's just a lot of adoption, and I think two, three years ago, that wasn't the case.

[00:04:44] Corey Silverman: Yeah, at Highwire, really since AI took off, and it's kind of gone in cycles for what we're hearing from the industry. You know, when it first started to really take off, there was definitely some fear around it. Some: "I don't know if I trust this being in my software platform."

And that eventually moved to wanting AI and asking, "What are we doing with AI?" Without, like, a really clear direction of what it is they want. They just knew that their leaders were pushing for AI and that they wanted to start using it in the platform so that they could show their leaders, "We're using AI here." And I think now we're entering into a period where people are starting to realize what it is they want from AI, what it is that AI can do. That's what we're starting to hear, so I think you both gave some good examples of that.

[00:05:24] Erik Lenke: That's a really good point 'cause what it talks to is the industry's becoming more educated in the technology. So, a few years ago, everyone heard the term AI, and they felt like it was this new thing. A lot of the technology that AI is built on has been around for much longer than most people know.

It's the new way that AI is consumable through a lot of what we see, the large language models and the different services that provide them, that have kinda hit a consumer-grade product that is now making it readily accessible. Users can now experiment with those capabilities and start to define their own use cases.

That's where we're seeing a lot more customers come to us and ask for AI to be used in a certain way within our product, which is a new dynamic because just a couple of years ago, we were having to go out to customers and both teach and educate them on what AI is and how it can be leveraged. So, seeing that dynamic shift has been really helpful in helping us understand where the market is and where the industry is and where we're headed as a partner with those industries to make sure that we're capitalizing on where it's gonna make the most impact.

[00:06:32] Corey Silverman: Absolutely. So, something we hear from our clients when we discuss AI with them, and something you just hear across the world when it comes to discussing AI, is the topic of responsible AI use and adoption. What do you guys think, and what does it mean to you guys when you hear responsible AI adoption?

And how does that shape the way we work at Highwire and Veriforce as you guys think about incorporating AI into the products you guys manage?

[00:06:56] Erik Lenke: One thing to note about AI is that it is a probabilistic technology.

That means that it does not work on a deterministic sort of pattern where it is doing the same thing to calculate the answer or the response that it's giving every single time. So, when we talk about responsible AI, there's an ease of effort that comes along with using these tools, but it doesn't remove the underlying work that needs to be done and how the human-in-the-loop intervention carries the responsibility of the output and the outcome that they take from AI.

If I was a safety inspector and I was on a job site, and AI told me that something was a low-risk sort of safety observation, something I shouldn't worry about, and I just took that response without applying any critical thought, without applying my own expertise and judgment, there's actually an accelerated risk of that technology because it does not have the definitive ability to calculate with a hundred percent accuracy that that is in fact a low-risk finding.

It is making a probabilistic estimation based on the data that it has available and the way that it is trained against that data. So, as we think about responsible use of AI in safety and in technology, we have to remember that it doesn't remove the work of the human to both confirm, validate, and provide feedback of the responses that the technology provides.

[00:08:27] Peter Chong: Yeah. I think one thing that I'm always reminded of is that the reality is AI, if you really fundamentally break it down, is a very smart intern, right? If you think of like a very smart intern, you know, you can tell it, "Hey, can you go do something?"

The reality is, without the full context of your work or the context of what it needs to do, you're gonna come back with an answer that might be good and it might be bad. Because of that, you think about like the real world before AI, if you had interns working on certain tasks, you have kind of gates and kind of barriers to make sure that intern isn't submitting poor work, right?

And so there's kind of like a framework here, right? When you think about just that intern process and then even with AI, right? It's understanding when you submit something to AI, it's understanding: Where can it break? How can you know that something is broken? What—where is, like, kind of that reflection point? Do you know it's working, is also another question that you should probably ask. Like, if something is going well, how do you know it's working? What's like the success of AI? What is the output? And then the final thing that I really wanna highlight is who owns it when it doesn't? So all work gets done, you submit it to your intern, you get it back. Who is now responsible for that output? And who has the responsibility to make sure that AI actually did something that it's supposed to do?

[00:09:58] Corey Silverman: Those are both great examples. Erik, I really liked the example of, say, some sort of AI hazard recognition tool returning a suggestion of this being a low-risk finding and the risk being there of someone just taking it at their word and not really doing the verification themself, and then all of a sudden you have the potential for an AI hallucination or just incorrect categorization to impact the bottom line of an inspection being done on a project. Something we hear a lot from clients when talking about some of our AI products or just theoretical AI products that could exist or ones that they've seen from other providers is, at what point can I allow the AI to make the call for us, and at what point do I have to come in and intervene and essentially override or take the AI analysis with a grain of salt and make a determination for myself?

[00:10:46] Erik Lenke: Yeah, that's a really great question because I think it, it kind of speaks to the spectrum of risk when it comes to AI technologies and sort of our growing adoption, but there's a very high risk of dependency that we could kind of come up against, which is basically talking about you can't be dependent upon AI to make the decisions for you unanimously.

There is a spectrum that I think every business carries of what are the low-risk, low-stakes decisions that, that you look to fully automate, and some of the reasoning capabilities within these AI services really allow you a new level of sophistication when it comes to applying that reasoning layer.

So, when we talk about that in safety, there are humans at the end of the work that we do. So, if there is a potential risk to a human being, that's obviously something that you don't wanna fully delegate to any technology.

You want to make sure that you have the professional care and the expertise of your workforce that has been trained on this because just like Pete's example of using an intern, if you had a junior safety inspector out on a job site, they would likely have some form of validation of the findings that they're making.

The same thing should be applied to AI technologies when you start to do these sorts of decision-making handoffs where you are continuously sort of monitoring the outcomes and the outputs coming from an AI product.

When we think about effectively where AI can be most useful, it is really the big-data sort of aggregation and reasoning. It's the time to value compression that it provides. So, if we're looking across a multidimensional space of safety where there's a myriad of contributing factors that could contribute towards a contractor individual's risk to a job site or a project, it takes a lot of time and energy to consume, understand, digest, and interpret that data.

A well-programmed AI layer within a technology like Highwire allows you to compress that cycle time down so that we are constantly feeding you insights ahead of you having to go back and rationalize all of that data that can kind of live disparately on its own.

But at the end of that, there's still the responsibility on the individual to make sure that the outcome and the output that they're getting is the right one, and that comes through validation, that comes through feedback that you provide back to the technology itself, and all of those things contribute to high-value utilization of AI.

[00:13:30] Peter Chong: You brought up like a really great point. Obviously, we need those blockers and to validate. I think it's especially important. You brought up kind of like that data layer, right? In that same NSC survey, they found that only eleven percent of safety systems are fully digital. And so, when you think about like how much does the AI actually have context of, it's probably pretty low because it still exists in a mix of paper and software, right?

All of that paper is basically not in the AI context, right? So if there's less material, you definitely need those barriers and to make sure that people are still making those decisions at the end of the day. And I think maybe, Corey, this is a good question for you as like a safety professional. If we did trust AI and completely do something end to end, it gets it wrong, like how detrimental is it?

[00:14:27] Corey Silverman: I think it's only gonna take one mistake for trust to be completely lost in the system. And it's because these small errors in something, whether it's a schedule for construction, whether it's a high-hazard risk assessment, whether it's a pre-task plan, one small error can be enough to completely derail operations, set the project back time cycle, or result in a serious injury or fatality. So, all it's gonna take is that one miscalculation, one error for there to be complete loss of trust in the system. But it's easy to fall in the trap of that as soon as AI gives you something, you can use it right away and not have that verification process.

But I wanna get into the real meat and potatoes of this conversation, which is around some actual examples of really practical use cases of AI in our industry, the challenges they address, and what the actual outcome is. I'm sure most of us listening are familiar with the concept of a safety inspection or observation program, where inspectors walk the job sites and record both positive and negative observations that they make surrounding the work being performed and the conditions on site. Most companies at this point have some sort of software or application they use to log these inspections and observations, and then they have some process for aggregating the data, understanding trends over time, tracking what's open, what still needs to be resolved. But doing this requires a manual user input for each individual observation, and this can end up being a really time-consuming and inefficient process. I've been a part of numerous safety inspection programs that give us a lot of good data, but there's a ton of friction around the time and effort that is required to log all these observations.

I think one of the most practical use cases of AI in construction and using AI to streamline the process is using voice-to-text AI. Instead of manually writing out every individual observation, users are able to just use voice-to-text, talk for ten or so seconds on what they observed, and AI can turn it into a structured inspection finding. So, it really cuts down on the time and the friction required to do these inspections and observations. This is something we've incorporated in Highwire and received really great feedback on. And so, I really like this. It's just—I think it's a great example of a direct use of AI to improve a process.

We have this problem of inefficiency. Here's how AI can fix it, and here's the benefit. It's not surrounded in hype or theory or fluff. It's very much, "We have this problem, and here's how we fix it."

[00:16:46] Erik Lenke: Based on what you just talked about, you're talking about basically an assistive use of AI, where AI is not just basically doing some type of reasoning for the person.

Instead, it's more of what I would call like a convenience layer. So, it is allowing your team to handle more volume of work at scale by taking on some of the labor intensity of going through and conducting things like safety observations.

So as your team gets more experienced using these types of capabilities, they'll start to establish their own judgment around where it's most effective and where they still need to insert themselves. But with capabilities like this, there's a lot of flexibility around like when to use it, when not to use it.

And what we like to do is make sure that all that feedback is captured because within some of the Highwire capabilities that Corey mentioned, that'll help us mature the applications, but it also helps us mature the underlying AI technology as well. Because as that feedback is gathered, we can start to learn how you all are collecting that information, what is most important to you, and AI can start to support better insights and analysis off of the back end of that.

So it's a really good example of how a simple use of AI can be used within a regular workflow that you're conducting every day, but also drive even more value off of the back end of it based on the data that we collect and how we're bringing that through the sort of loop back to the insights and analytics creation.

[00:18:22] Peter Chong: We have a very unique use case. Highwire in general, we, we collect a lot of data, right? For our contractors, when you think about five years of maybe EMR, TRIR, their OSHA citations, insurance and safety programs, all of those materials get siphoned into our platform. And when we think about all that data, it's also important to understand in context how hard it is for a normal user to sift through all that and then come up with an understanding of a contractor's risk.

A safety professional for sure can go through all of that, but even them, you know, if they have a few hundred subs going through each one of those contractors' profiles and looking at those documents is very difficult, right? We provide a score, but that doesn't necessarily give you enough context. It can't give you an answer like if a fall-related citation is a big deal for like a steel erector, for example, right? Is that citation important for this contractor on this next project that might be related to an interior fit-out? Probably not that much, right? And so, it's really about how can we understand the scope of the project, and then provide a response to a user of contractor risks in that context, where, depending on what type of project it is and what the scope of work is, does that risk actually matter for this particular work?

[00:19:56] Erik Lenke: That's really good 'cause I think where that leads us is to sort of the concept of understanding sort of how AI influences the work being done on a site. So, what Corey kinda talked about with inspection capture. Then we're talking about how the additive context that AI can apply around data can start to help materialize an understanding by the user.

But ultimately, where we're going is this concept of like, what are my risk signals? You may hear us talk about from time to time, where is my risk increasing, why, and what can I do about it? So typically, that's a human judgment call based on their experience within their industry, within safety, using a given sort of provider or platform for risk management, similar to Highwire.

But ultimately, what AI can help facilitate is doing some of the sort of initial signal gathering and surfacing some of that to the end user within the products that you're using every day.

That all sort of plays into the concept that many of you may have heard across the industry called predictive analytics, which is really a machine learning concept that has been around for ages, which is looking across a wide and deep data set of historical information. What are the patterns that we can establish, and how can we train AI models against those patterns to develop an estimate of the likelihood of occurrence of a certain event to occur in the future?

Specifically, within safety, I know it's of very high interest: What is the likelihood of a contractor having an incident. That requires an immense amount of data and an immense amount of analysis. As an individual user, it's tremendously hard for me to grasp all of that data and be able to make that judgment call on my own.

So AI becomes an assistive layer to the end user to bring that data to the surface, explain the reasoning of why it's coming to that conclusion, and ultimately suggest that there's this higher-than-normal likelihood that this event could occur, and here are some steps to interdict that process or that outcome rather.

And so, one thing that you'll start to see is those types of products come to life, not only within Highwire, but across industry and across providers. And I think that that's gonna be the real trend that we see adopted.

So, with that, I think that there's a lot of human judgment that still needs to be applied, and we still have a long way to go when it comes to making sure that these models are as reliable as we want them to be, and we can start taking less and less responsibility off of the human aspect, even though I don't see it ever fully going away.

Pete, I'm not sure if there's anything that you wanted to touch on that because I know that you've been at the head of the Highwire development of our image classification and risk identification technology, but feel free to add any additional comments.

[00:22:54] Peter Chong: No, I, I think you hit all the points on kind of like our platform. I think the most important thing is kind of the workflow and, and the thought of this. You know, we initially thought when we did this, we automated the inspection, and we realized that wasn't a good idea, right?

[00:23:12] Corey Silverman: Yeah. So, when we first started experimenting with hazard recognition tools using AI, we started building a prototype for a product that was designed to be hazard identification. You take a picture of something on a construction site; it returns either, "This is a hazard, and this is what needs to be done about it." And as we started testing it more and more, we just felt that we really weren't comfortable putting that much power into the AI's hands. It goes back to the discussion earlier that Erik and Pete both brought up of the example of just taking AI at its word without really verifying. There was too much potential for abuse there, right?

[00:24:00] There's potential for a hallucination or a misclassification. Someone doesn't check it and just logs it as a hazard, and then next thing you know, there's reports being sent around the company of this high hazard on this job site that ends up actually was not really a hazard. We kind of shifted our approach and changed it, and now what we're experimenting with is more of something that returns suggestions of things to inspect. So previously, maybe you would have taken a scan or a picture of a trench, and it would say something along the lines of, "I don't see fall protection here, so this is a major hazard that needs to be addressed right now." The shift we've taken is instead of identifying an actual hazard, it's returning suggestions of things to take a closer look at, like, verify that fall protection requirements are met if there is a fall exposure here. Verify that atmospheric conditions have been monitored before entry. Verify that it's properly sloped and shored. Where it's not actually saying, "This is a hazard; this isn't a hazard," it's saying, "I suggest inspecting these things." So, I think it's a really good example of how AI can be kind of structured in a way that prevents humans from abusing it. In order to create a hazard with this new approach, the human has to say, "This is a hazard." There's no potential for taking what the AI said and just pushing it along as a hazard without the human verification aspect of it.

[00:25:00] Erik Lenke: This is all kind of coming back to like AI is booming within our industry right now, and I think that there's a lot that we can learn and gain from it, and Highwire's gonna be obviously taking advantage of this boom, trying to bring better and more value-adding AI capabilities into our products.

But also, it all comes back to sort of the responsible use of AI, making sure that we're not sort of bringing capabilities online that put our customers at risk and delegate too much work to the technology, keeping that sort of ownership with the human aspect of the team that is managing that work and that workflow and ensuring that there's the proper sort of reliability, training, and governance in place to make sure that you get the most out of the technology itself.

One thing that I wanted to add is it's awesome from a technology perspective to see this type of maturity and modernization happen. I think that everybody, even outside of safety and industry, is seeing with each sort of step model change the improvement of sort of the front-line sort of capabilities that are coming online.

But as we continue to sort of mature and grow, I think that there's gonna be a very significant shift within the safety industry as well as sort of the supporting services like Highwire. And it's important that, like, as we talk to our customers and they talk back with us, we bring this to the forefront of the conversation and understand where and how they wanna see AI leveraged going forward.

[00:26:36] Corey Silverman: I think you provided some really great insight here that will leave our listeners with a lot to think about. I do want to wrap up with one closing question. For a safety team just starting with AI, where should they begin, and how will they know that it's working?

[00:26:50] Erik Lenke: So, I think the key element for getting started is try and be specific with a challenge that you're trying to approach. So, if that is trying to improve the quality of the documentation being sort of added to your various sort of inspection program findings, start there. See if the AI capabilities that Highwire or your provider is giving you actually improve the quality of the outcome that you're getting from those workflows.

The other thing is keep what we call human in the loop, or you may have heard that used across the sort of the, the review and the accountability plane of whatever work that you start to leverage AI through. Make sure that you have sort of expert-led oversight of the quality that AI is giving you, because that'll give you a real sort of understanding of what is doing well, how accurate things are.

Because just like any new technology, you need to have educated individuals measuring the performance to be able to really benchmark your success.

And then lastly, I would talk about what is the actual time savings that you're getting, and has your quality gone down? So, you basically had that specific workflow or that specific area of work that you want to address. You've added that sort of human-in-the-loop accountability plan to make sure that the quality has not gone down through that work.

Now look at what are you actually getting back? Is that time? Is that new sort of insights that you didn't have before? And start to quantify what that is actually bringing you back as a business.

[00:28:37] Peter Chong: I think you added everything that I would wanna say. I think a few things: Start somewhere with low stakes, right? If AI gets it wrong, it's fine. It, it doesn't matter too much, but you can see what happens, right? I think the other thing is that you need to train your people. I think even with AI just so readily available for us, within our day-to-day lives, I think it's a bad assumption to think that everybody understands what AI is capable with. And so even upgrading your workforce or, like, training your workforce, that's so important. And I think there's a lot of free resources. I think the vendors that you all use probably have some sort of training, and they also probably have some sort of AI product that you can use today, right? And so, starting there is also a good starting point.

The last thing that you should incorporate is understanding where your data lives, right? Where does your data live? If it lives on paper, can you digitalize that? How can you get that context layer that we've talked about centralized in one place?

I think the people that can get all of their data in one area are gonna be the people that will win in the future.

[00:29:52] Corey Silverman: Thank you guys again so much. We're gonna go ahead and wrap this up. Thank you all for listening to the Beyond Prequalification podcast. We'll see you next time.