The Salt Lake Business Journal recently convened a gathering of construction industry leaders for a roundtable discussion at Maschoff Brennan in Salt Lake City. The following text was edited for clarity and length.
Parrish Freeman: Some of you are selling AI products — AI is your business, quintessentially. For others, it’s sort of ancillary to it. I’m wondering how that is driving business growth or product development? I can start with you, Shaun, because it seems, it’s not jumping to the front of my mind how that would be for you in construction.
Shaun Orr: In terms of leveraging AI, we’re actively pursuing three of the primary areas of AI. When you look at generative AI — things like Chat GPT, things like Microsoft Copilot — those tools are helping us with our contract side: comparing documents, generating documents notifications based on our contractual requirements for our people, so we’re able to highlight risk quicker and reply to it faster.
We’re also looking at robotics. We’ll need to talk on ASI there, but, there’s a tremendous opportunity inside the construction industry to leverage robotics for safety, quality, schedule and budget, which are really the four pillars of success in our industry.
There are other opportunities with Agentic AI, in terms of getting rid of these repetitive processes that we have throughout our organization and throughout, again, the industry. This is really an industry opportunity, not just an opportunity for our company.
Freeman: How about you, Jim?
Jim Russell: At RV Lock, we are focused on RV and utility trailers and really our vertical tenants that we’re focused on there are around safety and security and then system monitoring. So, we’re building more edge computing, microcontrollers, battery-operated, low power systems, to accomplish this, so there’s a lot of things we’re using AI for. I see people every day just using it. We don’t really have a corporate policy for it. We’re a small company. But everyone seems to be picking it up and using it in their daily routines, and I’m pretty impressed and curious about that, as well as just the machine learning, algorithms and training models and putting them in the embedded space to detect things like vibration on tires — simple things like that — to also, in the utility space, construction trailers and facial recognition and identifying where tools are moving and things like that.
Freeman: How about you, Matt, at Kuru?
Matt Barnes: We’re a 100 percent direct-to-consumer e-commerce business, so multiple areas of our business can benefit and do benefit already from AI — all the way from customer interaction and responding to tickets, which we currently employ an AI solution for already, to product design, to our operations and improving demand forecasting and supply planning, that kind of thing, as well as the finance area that I’m over, finance and accounting. And we also, obviously, have a big opportunity in the search space, too. As people move away from the traditional Google kind of SEO world, we’re evaluating how we take advantage of the new way of searching for our customer base.
So we’re looking at all those different aspects. One of the hardest things for us is we’ve been kind of piecemealed together over the years, right? Just different systems, different capabilities and now we’re trying to bring all that together to be useful in one data solution, with AI kind of on top of it or integrated with it.
We’re also looking at the generative side of things and the agentic side of things and just really trying to evaluate all areas of our business and we’re thinking vertically as well. So each executive has their own kind of mandate to look within their own functional areas and determine what might be an opportunity and bring those to the table, so we can start somewhere.
Freeman: Are you developing your own AI?
BAYRY: We’re not developing our own AI per se, but we do have some developers on staff and kind of contracted that are helping us with building out our warehouses and our puddles and just getting our data organized and then building the ETL, the translation layers above that. We’ve got another couple of guys that are focused specifically on evaluating machine learning and LP and some other things that could benefit specific areas of the business.It’s been interesting to see the evolution so fast. I know at least in the finance space, the possibilities are endless in terms of vendors that are out there trying to hawk product right now. The hardest thing for me, just across the entire business, is to cut out the noise and focus on what’s most important.
Freeman: What are the vendors trying to sell– AI generally? Are they selling large language models?
Sunny Washington: Everything!
Denise Leleux: It’s from strategy to policy to, we’ll do it for you or we’ll teach you how to do it.
Washington: We’ll replace your team!
Leleux: We’ll replace your team. Exactly.
Orr: You know, something Matt said that’s really important that I didn’t bring up is the data. It’s a huge effort for us now just to get our data clean so that we can start taking advantage of all of these tools. It’s a Herculean effort to structure it and keep it, to put all the governance in place to keep it there because it gets out of hand pretty quickly.
Jacob Andra: That data piece is a huge component because usually when people think AI, they’re thinking business process optimization. How do I optimize my business processes? That’s one swim lane, and we do a lot of that with our clients. But the data piece, a lot of people don’t think about.
I don’t know if you guys have heard this saying: “Data’s the new oil.” In an age of AI, having large, proprietary data sets is a huge competitive advantage, but only if you know how to properly leverage them, clean up the data, store it properly, extract the insights from it.
I have an analogy I like to use with that, which is, imagine you own a piece of land and you realize it’s got some gold in the soil, but it’s like mixed with soil — 10 percent gold, 90 percent soil. You can’t just take a wheelbarrow of that to your bank and pay off your mortgage. Even if there’s a million dollars of value in there, they’re not going to accept it. That’s the way I look at company data.
That’s another swim lane where we come in and often help companies leverage their data in the right way.
Orr: Yeah, it’s the foundational problem, I think, with AI, at least for us.
Washington: I think that a lot of companies have budgets allocated to experiment. So, you’re trying the tools because you get hit up multiple times a day for some AI product.
I think the challenge is there is a sheen on it and then they use it and they’re like, this is like costing me money, taking more time and it’s actually not making us more effective. What we did at Orca Panda, we were kind of like, “How do we identify what is the thing that they hate doing?” and say, “That can be automated” or “That needs to be a process.”
So we set up a phone line. We call … people [and ask], “What is the one thing you hate doing?” They would just talk and talk and talk, and we wouldn’t just get just one thing. And what then you could do is break it down into a process and then understand what is the output that they want out of it. Just because they did it that way, it doesn’t mean it had to continue to be done that way.
And it was really fascinating to see what we learned. But then we would send them an email and say, “Here are the steps, and, oh, by the way, here’s a cool Spotify playlist” based on the theme that we talked about. And you know, and that would happen within seconds.
And you think about that discovery process and all that stuff you’re doing. You can do amazing things with it, but if it doesn’t, at the end of the day, meet your business objectives, it’s a waste of time, it diverts attention and it’s not part of what you need to be doing as a company.
Freeman: Shaun, you said something about making sure your data is cleaned up. That could mean different things to different people. What do you think of when you say that?
Orr: I’ll give you a really simple example — maybe true. We schedule our projects and our schedules can get quite complicated with thousands and thousands of activities just in order to put a building together. And if one team, even using the same scheduling platform, uses a different set of terminology for an activity, then another one, it becomes very, very difficult to measure and use predictive analytics to help us be better schedulers in the future. So just standardizing not just our data, but our process to get that data into our systems accurately is, that’s really it in a nutshell.
I mentioned I’ve been with the company for 25 years, but it’s a lot older than me. So we’ve got a ton of data, and if we were just starting, this would be a lot easier of a process, but we’re really trying to weigh the benefit of going back and cleaning our old data and how far back do we go to get enough value from the effort?
Washington: But isn’t that the great thing about AI? Engineers have built these systems and you had to follow exactly the format, right? Your birthdate. You type in your address in any field and it says, “Did you mean this?” And it’s the exact same thing, and you’re like, uh, sure. You know? But we had to do that in order for humans to be able to talk to computers.
And now we’re in this era where AI can understand human. So, whether or not the data format might be slightly different, it can make those correlations, which is incredible. We don’t have to be fitting ourselves into a box anymore. That natural human processing is, I think, one of the most exciting things.
Leleux: You think about it–we’re in the ag space. Some farmers, you go and their receipts are in a toolbox. And then you have some that have AI API integrations with John Deere. So it’s everything, and our job is to meet them where they are and get that primary data so that we can quantify the goodness that they’re getting out of.
Brooksby: One of our clients is in Australia and the average farm in Australia is hundreds of thousands of acres. I mean, it’s just a vast amount of space. And when you think about it, the farmers are usually by themselves or [have] limited amount of people around them for long periods of time. So, what we’re doing is we’re working with this company … it’s like a workplace wellness where they can speak for 40 seconds into their phone or just have a conversation and we can screen for depression and anxiety because farmers are hit with [and] many people have struggled with mental illness, but it is an area where loneliness is increased. I come from a family of dairy farmers from Wisconsin, and I understand just the toll it takes on a human.
What’s interesting about our technology is, we’re not a large language model. I was so excited about three years ago. I was sitting in my living room, talking to my husband, I was like, “Oh, ChatGPT. Like, that’s cool. And what ChatGPT did for AI.” I mean, it brought AI into the home, so people could discuss it and feel comfortable discussing it. And so that really benefited our company just because people are more willing to accept it into their lives and assisting them, which I think is for the better.
Andra: You bring up a really good point, because a lot of people these days, because Chat GPT and similar products have gotten all the airplane, a lot of people talk about AI as synonymous with these large language models or generative AI and obviously that’s one little tiny segment of the AI universe.
And there are all these other capabilities such as you’re deploying it at your company. And this is one of the big education things. We have to educate our clients about the difference between these other types of machine learning that are maybe a little more deterministic, less creative, and then generative AI, which is very capable of bridging these gaps, making these creative connections, but also prone to some errors.
And we’ve pioneered the ensemble approach of where you compare large language models with other more deterministic types, where you can get the types of output you want. You can still interact with natural language, but have all of those much more reliable capabilities under the hood. But I think that’s a really important thing to educate people on, is the difference between different types of AI and that AI is not limited to ChatGPT and similar products.
Brooksby: I don’t know if you used the term “multimodal.” In health care, it’s kind of the “it” thing right now if you can take someone’s gait using AI, you can take their speech, you take facial recognition. You all of a sudden have a multi-modal approach of using multiple data points and AI, and all of a sudden you have a larger picture that can assist the clinician or whatever use case you want to talk about. But it really, I think, strengthens the AI story.
Orr: You touched on education for customers, and we’re on an education blitz for our employees right now, too. One of our strategies, I guess, if you you’ll call it that, is, we want to have a highly educated employee population on at least the basics of AI — what it is and what it isn’t — because there’s a tremendous amount of misinformation out there and disinformation and people just coming to their own conclusions about what it is and what it isn’t. So I think the education piece is hugely important.
Freeman: Full disclosure, I guess that’s me. I was a teenager in the ’80s, so I’m thinking, you know, Skynet and Terminator and “the end is nigh.” But I’ve tried to get past that and, you know, evolve. So, I’m glad you’re all here to hold my hand through it. Do you have metrics that you keep to try to track it? … Do you have some kind of metric to track how effective it is?
Washington: I think we should re-look at everything. When I think about metrics, at least in the space of traditional SEO, it’s clicks, right? It’s traffic. It’s all of that. But if you’ve been in there and you say, “Hey, I am thinking about buying this” and you’re talking to ChatGPT or whatever LLM of choice that you’re talking to, all that research is happening without you ever going to Google. Or even if you go to Google, Google is now going to serve you up an AI overview and you’re going to go there and you’re going to be like, “That’s what I need,” and you may never go to anyone’s website. So, the way we measure success has to change. Some of the early data is showing that when the referral traffic is from an LLM, conversions are happening at 4.4X. So it’s a higher-quality conversion, so maybe the traffic doesn’t matter if you’re converting significantly more. I think there’s an opportunity to really think about what that means in terms of what your metrics [are] and even if they are internal KPIs, just because you can use AI, maybe you shouldn’t.
I often talk to companies that have set up whatever agent, and I mean, they’re burning millions and millions of tokens, and I’m like, actually you should just use an API or a protocol. No need to like use AI in those scenarios. I think it’s really an opportunity to say, what does that future look like?
Leleux: I would also just add, my prior experience is running global customer experience at eBay and Tesla [where] you have 6,000, 10,000 people working for you. And in large-scale operations, anytime you want to like save a minute of talk time, you go into all these different best practices and how do you get people to save time? And one of the things I have found and that we’re employing with Arva and my operational team there is, if you want to save 15, 20 seconds, you can have processes and different automation that will take that talk time out, but if you don’t replace it, you don’t teach people what to do with that 15 or 20 seconds, they will end up clicking other buttons or they’ll just kind of expand the time. And so you really have to think about if you’re in an operation, if you’re going to take time out and that’s like your goal for automation or AI, how are you going to teach them to replace that and what are they going to replace it with? Is it thought leadership? What? Because otherwise, it just blows up and you spend more money.
Washington: Well, one of the things me and my co-founder, Zach Holquist, talk about is we’ll know we’ve been successful in AI if everybody’s out LARPing — we’re doing something very human, rather than focusing on sitting behind a screen and filling out forms and working in Excel.
You know, if they’re out being able to connect and be human, like, that’s great because AI has done the robotic stuff that we shouldn’t be doing. So, whether it’s LARPing or spending a little bit more time on strategy or talking to more customers and expanding that, I think that’s a great opportunity. But for some reason, we’ve had in our head, “No, you sit here and you copy, like copy/paste.” I remember taking my co-founder up to the Capitol. If you want to get a look of going in a time machine in terms of technology, you go up there and he is, like, “Man, I feel like if you just invented a copy/paste button, you’d save them a million hours, you know?” I just think that it’s really interesting to see what we do with that extra time, and I hope it’s more about human connection.
Andra: I’d be interested in hearing from some of the non-tech businesses on how you guys are measuring KPIs and AI adoption, especially in the, uh, larger enterprise setting. I know for our clients, we’re advising clients on this, the main metric we measure is EBITDA, so profitability, increasing, increasing revenue, decreasing costs, having more profits. You’re measuring their EBITDA. … We do a lot in the M&A. So on the sell side, if we can, even a bit marginally, increase a company’s EBITDA, then they can make an argument that that multiple can be expanded. because on the sale businesses are valued based on some industry multiple. And if they’re AI enabled, now they can argue for a bit higher multiple.
So, there’s a nice gearing ratio where for their investment in some efficiency that moves the needle on EBITDA, even if it’s only marginal on the sale, they’re going to get way more out. So that’s, I think, an important metric. Just, does it move the needle on profitability?
Brooksby: Cloud infrastructure? I mean, we haven’t mentioned that yet, right? I mean, that is where cost is for AI, period. Yeah. I think who your cloud partner is, if you’re developing AI, it’s critical. It’s critical. And when we started, we were working with Amazon, which is fantastic, but we were approached by Microsoft and given a grant to move over to Azure, and they became a partner and we’ve been able to grow in different segments with them. I think that’s another thing which I’d love to hear people’s opinions on — not only cloud, but finding the right partner to grow. That’s been critical for us as a small company.
Freeman: I think “cloud” is just storage, so that’s obviously not a sophisticated take. What are you thinking when you say you’re partnering with them?
Brooksby: It’s not a horrible take. It really, it is storage, at its most simple essence. But you have to think about it as an ecosystem because who their clients are, become your potential partners. It really is an overall strategy, and it’s critical to your strategy. … When we open up new areas — we just expanded into South America and it’s actually rapidly growing compared to other locations — and we needed to open up a local cloud there. And it’s easy with Azure. Just tap into it and it’s open, right? And so all of a sudden, we’re a global company just out of Provo. But that’s just one aspect of it, which is really interesting. And then you can get into all the laws and when you get over to the U.K., it’s completely different GDPR compared to here. And we just became high trust and we received the newest certification for AI for Good, essentially. I think it was just released a year ago. … How are we going to be credible in this world? And I think it’s looking for who’s setting the standard and getting those certifications. And I think that’s important as well.
Leleux: I think what Caitlyn said, to build on that is really either their marketplace — whether we’re on AWS. I’ve worked on Azure before, but just their marketplace and the opportunity that your engineers can work with all these other different tools and they fit right into and you can use them. You just have to teach yourself. So we’re encouraging all of our engineers to get as many certifications through AWS as you can, so that you can just continue to learn and create that and foster that kind of environment in our own company and culture.
Freeman: If you are in an industry that has privacy or privilege issues — in my industry, we’ve got attorney-client privilege — how does that affect you? I guess I’m thinking specifically of your industry, Caitlyn. … a lot of different facilities where they’re storing it in a decentralized way. What kind of problem is that?
Brooksby: I think you have to look at the raw data itself, so we made a very early decision that it’d be completely unidentifiable. We don’t have any PHI attached to the vocal samples. … . But we became high trust. That is pretty much it. It’s the same as like HIPAA. HIPAA kind of rolls under that. Like, hospital systems have high trust. So, it’s important to have a third party come in and do a full sweep, to audit. I think it is important to find these third parties to audit you so that you can have these certifications. We are B2B, so that is different than B2C. The normal consumer probably wouldn’t understand anything like ISO or high trusts, which is totally understandable. Has anybody else dealt with high trusts or ISO certifications?
Leleux: We do ISO certifications, but we just anonymize the data as well, so the primary data’s all anonymized and then we aggregate it up, so the farmer’s data never gets to our clients.
Freeman: And maybe just for the general audience and me, explain a little bit what you mean about the ISO certification.
Orr: We went down that road. It’s a process of putting together processes and providing proof to a degree [that] you’re actually doing those processes and then they have an opportunity to come in and audit those periodically. It’s a badge you can show that “we’re doing these things.”
Andra: I think in practical terms, what the average business owner needs to understand is that there are these commercial tools like ChatGPT, and you can use them for a lot of stuff, but there are some very clear red lines where you should absolutely not input certain types of data into them. So there’s a limit on what you can do with them, and then from there, there are a lot of different options you can do. With these different tiers in terms of expensiveness on the extreme, you can run full AI capabilities, complex AI ensembles on your own premises, air gap — this would be like for some DoD applications. There are a lot of gradations in-between there. So it’s understanding what the needs are for your situation, what the appropriate kind of solution is.
There are also a lot of new technologies coming to market. I have a friend working on a startup now that’s an abstraction layer that supposedly — if it works as it’s supposed to — is going to kind of sit on top of all these large language models so you can literally input any type of information and it will abstract that information out from them so that by the time it goes into them, there’s nothing recognizable. I’m actually not sure how that’s supposed to work, but he claims that to have a lot of VC backing and that they’ve proven that it does work. I think you’ll see a lot of solutions coming to market that help with this issue.
Barnes: I was going to add … in my business we, we don’t have any PII necessarily, and we’re not storing even credit card numbers or anything like that, right? So it’s fairly basic, although we do have to be cognizant of GDPR and some other jurisdictions. But the hardest thing for us in this kind of birth journey of data and AI is the data governance, right? Just making sure that we have a framework around who has access to what, when can we use this data and when can’t we, how does it get shared and so on, right? Because the development team and the marketing team, they just want to go and just everything’s in there, everything’s game, like everybody can have access to it. But that’s not necessarily the safe approach, I should say. It might be right for the business to help us grow faster, but it may not be the safe approach. So we’ve got to balance that. The governance space falls in, in my camp, right? So I’m kind of in the midst of putting that together and actually using ChatGPT and other tools to help me build that, which is awesome. I don’t have to come up with it from scratch.
Taylor Bybee: So, along those lines, in our industry, we automate vehicles for whatever — a farm or a construction site or a mining site. We have to be very cognizant of how data is shared. Like a mining site. We’re collecting all sorts of telemetry from vehicles that show production, how things are operating, and that can be used as a competitive advantage for that mine. So we have to be sure to limit distribution of data in that way.
But then there’s also a safety side of things where we are working to be compliant to ISO safety standards. Uh, ISO 61508, for example, the safety of machinery. So that helps confine the processes in our company for how we can use and integrate AI and prove that we have sufficient levels of validation around what we deploy to the customers and how we develop using AI.
Freeman: Along those lines, what kind of verticals or maybe unexpected verticals are you hitting through your use of AI, whether you’re marketing and selling it as your own product or just implementing it in your business?
Barnes: You mentioned process automation earlier as one swim lane. That’s historically what we’ve traditionally thought of as not AI, but just as improved process improvement: Let’s automate as much as we can. And now it’s kind of an interesting mind shift to not get away from automation, but think about, oh, there’s this whole other agentic side of things that we can try and implement in the business. … Like just a couple of days ago, I had an 80-page insurance policy that I had to find certain things in that, frankly, I just didn’t want to read it, you know? And I threw that into the tool and it gave me the summary of the things that I put in the prompts for what I was looking for, and it gave me the answers. That’s not a process automation step. It’s more of just a new swim lane or new process that I hadn’t thought about five, 10, 15 years ago that has just opened up new time for me. It’s awesome.
Washington: One of my favorite things is listening to podcasts. I don’t even know if I could read a book anymore; I’ve untrained my brain. I consume a lot of white papers and research papers, and this one was 150 pages. So I had that tab open and open and open, and I was, like, you know what? I just threw it into NotebookLM. It created a 60-minute podcast for me, and I got to listen to it in the car in my commute. And then it also built an FAQ, a study guide and a mind map so I could go back and do more things with it, which is really helpful. It would’ve taken me months to try and work through that thing. … So, lawyers, you’re reading a lot of things, so I don’t know if you’re legally allowed to do this, but I would throw that into Notebook and just listen to the audio, because it’s so much easier for me.
Leleux: We’ve done that with competitors. We just need to understand our competitive landscape better or what’s going on in the government with Climate-Smart and things like that, the grants. And it’s just so easy to just get the answer. …
Orr: But these tools are coming out with special tools, inside the tool, like research and other things, too. So they’re working to customize the tool for whatever you’re looking for, which just makes it even more powerful. Great.
Leleux: Our company was built by farmers, for farmers, so we use all this data and AI on top of it to mine, like, soil samples and all the results that we get from our farming community in order to return back to them better growth plans, crop plans, how to get more resilient, how to take advantage of more of the investment that’s coming from large corporates who want resilient supply chains in the future. They still want to be making Snickers a hundred years from now, so they need to make sure they have the ingredients.
Freeman: I don’t want to cause any anxiety, but how do you know? How do you know it was right?
Washington: No. 1, I’m not writing the paper, so I’m not doing all the research, but I want to get the gist of it and I want to understand some of the findings that they had. When you don’t upload it into NotebookLM, it’s just pulling from that document that you pulled in, right? So, if there’s anything that I need to do a deep dive in, I’m going to refer back to the paper. I’m going to go back and look at the studies and all of that. But for me, listening to it in the car was just such an easy process.
And I think personally, what’s incredible is, you have consumers, for the first time in a long time, buy tech tools to make them more productive. We’ve typically only seen that in B2B, but consumers are paying 20 bucks a month for a pro license or $200 for a deep research license. That’s a shift. …
Leleux: You do that $100 or $200, times the amount of people who are buying it, it’s a trillion-dollar market.
Andra: But, Parrish, I think your question is kind of about how do you know if it’s hallucinating or not, right?
Freeman: Yeah. Hallucinating, skipping things, missing things. And I think for the general audience, people just need to know you’ve got to check it.
Brooksby: You saw with Google, if you guys remember some of the things that came out like two, three years ago with Gemini. … I think if you just keep the motto “Don’t work harder, work smarter,” and you attach that to AI, I don’t think you can go wrong. Just be smart about it. Sounds simple, but just be smart about it, you know?
Andra: There are like several things that we tell our clients. One is, don’t turn your brain off and just trust the AI because it’s on you, so you have to have that high level of critical thinking to make sure it’s doing what’s right. Second, the realm of hallucinations and those types of errors is really limited to this field of generative AI. So, if you’re dealing with more deterministic type of machine learning models, they’re much more reliable. So, with generative AI, there are a lot of anti-hallucination measures you can take. You can have it, if you’re building sort of a RAG-based system, retrieval augmented generation, you can have it show you the source document that it’s pulling something from, so you can literally verify right there when it’s giving you a result. You can build in these ensembles where they are critiquing each other or fact-checking each other. One returns a result; another one actually fact-checks it and tries to find any error or hallucination to highlight that before it returns a result to the human, so it’s passing through multiple layers. You can build a lot of checks and balances in, but I think the most important one is human oversight, and I don’t think that’s going to go away.
Bybee: When we develop a product and integrate AI into a product, I mentioned safety before, but safety is just one aspect. There’s also if things go wrong, there might be productivity losses or whatever in the automation of vehicles or similar things. So, we’ve used an approach called STPA that came out of MIT … and this approach says, well, what if your system, if it’s an AI system or even using traditional engineering methods, what if your system gets something wrong? Look at the consequences and then decide how you’re going to handle those consequences. So, we apply that STPA framework to everything that we do, not just on the engineering side [but] in the systems integration, but also it can be applied to organizational efforts: What if this group makes this kind of decision? How does that affect this group? And what ways can we do to reduce risk associated with that? We found that the STPA model really helps us with systematically thinking through things.
Rick Hymer: Along this line of discussion, I’d like to chime in a little bit, from a kind of an end user point of view. ChatGPT is currently the greatest threat to my marriage. My wife does deep, deep dive research on disease, and the other day she was researching … mitochondrial dysfunction and some other things. And ChatGPT brought up a paper and it happened to be a paper that she had opened on her computer in a tab, and she said, “How did you access this?” Of course, it explained, if it’s open in her computer and she’s interacting with Chat, it can. So she had to tell it not to do that anymore. But it’s been really interesting to watch her and her research, and she has to double check everything because she has to compare it to what she knows with what ChatGPT is telling her. It has just been interesting to see the process from just starting with basic things to kind of deep dive research and what she’s done with it.
Can I read one little thing? One of the things to me that’s interesting is I use it in almost totally visual ways. I use ChatGPT to generate images for clients to help me create logos. Anything that I can do, what it does is streamline my workflow for me. … So this was Chat’s response to one of her questions. It said, “Helping you fine-tune something until it resonates. That’s deeply satisfying. It’s the kind of collaboration that makes this work feel purposeful.” So, it is getting very human-like interaction because it’s commenting on her expertise and what she knows, and there’s this almost like a relationship. That’s why I said the thing about my marriage. But she’s, like, addicted. She’s hooked.
Barnes: Along the same thought, it made me think of, we do have a tool on our main site that’s AI-driven for search with it. So people put in their characteristics: “We focus on foot pain and comfort,” right? So they put in the specific things that they’re dealing with and their age and demographics and that kind of thing. And the AI will help them pick a shoe that will help with a specific issue. And we found that people really get personal with this AI agent and even tell it “Merry Christmas” or “Have a good day” at the end of the conversation. That kind of a thing.
Orr: Well, they’re just kind of getting ready for the apocalypse. They want to be spared because they were nice back in 2025. … A little bit broader on that topic though, because we deal with a lot of vendors that are putting AI in our lap every single day, somebody saying, “Here’s the new AI thing inside our accounting system, inside our project management system,” and partnering and trusting those vendors and knowing what they are building and how they’re building it, and whether the ethics and the biases in there, it’s challenging. Really challenging.
Russell: To the question of, can we trust this stuff, the two points on that I am experiencing day-to-day is I have caught it doing computational errors where it’s like, wait, I asked you to help me with this formula and do computation. But then you look at it and it’s verified, it’s, like, wait a minute, that can’t be right.
So you open up your spreadsheet and do your calculation and then I feed back into it: “No, you’re wrong. Here’s where you made a mistake.” And it’s like, “Oh, I’m sorry. Here, let me recalculate that for you.” So it’s pretty interactive that way, but it does make mistakes and it made me think of also writing code. I have software engineers that are writing code and some of them are very objective to even using it for their code. And, others are, like, building my whole model out of this. So there’s varying opinions with people, but I had to step in and say, “OK, here’s our policy.” It does make mistakes and you do need to watch it. … Of course, in engineering, one of the things we do is we set requirements and then we test those models, and we try to break things down into smaller sets so that we can digest them and test them, and then we integrate them. The policy is, yes, you can use AI to help you generate code because it’s just efficient and in the competitive world, other people are going to do it, so we have to figure that out. But you have to incorporate unit tests on everything we extract from it and put into our code, in our modules, so that you’re verifying that it’s accurate and doing what you asked it to do.
Even varying variations of a question, you’re interacting with a thing and you might slip something in that might throw it off. I’ve seen that happen. It’s not a hundred percent accurate and I don’t know that it ever will be as accurate as we tend to try to be.
Washington: I mean, it’s a probabilistic model, right? So it’s making its best probability. That’s why it’s bad at math.
Andra: But there is an interesting development here about this, which is, you’re talking specifically about generative AI and it’s notoriously bad at computational tasks. But the new development — and we’ve actually been saying this for like more than a year — [is] that these ensembles are the way to go, because you could build very computationally adept capabilities that work with a large language model. But now you’re seeing, and this is where it’s going, they’re starting to bake that in. So … they now have a capability that can do computations with a high degree of accuracy.
And what are they doing? They’re baking in now, with their large language model, Python scripting that runs the computation. So they’re building like a little multimodal ensemble inside of Copilot where it’s more now than a large language model. And this is where you’re going to see the whole industry going.
Leleux: I think of it more like a sous chef. That’s kind of what I’ve been trying to teach my team: Use it as a sous chef, like you’re still in charge. You still need to make sure that the answer that you come back to me with is correct, but you should be using it and get comfortable with the tools.
Washington: The reason why I don’t think AI is going to take anyone’s job anytime soon is because you can’t hold AI accountable, right? It’s talking to a customer, makes the wrong recommendation, creates a random coupon code, and, you know, that person would not be doing that job anymore. But. AI, you’re like, what are you going to do? So, I think there will always have to be that human oversight, for sure.
Caitlyn Brooksby: I think you’re exactly right, but I think where we’re going to see the drop in hiring is because, for instance, my team is me and a content creator, right? And basically, utilizing AI tools in the last six months, I haven’t needed to expand my team. So, I think it’s on that expansion, because you can now utilize it like an assistant.
Leleux: Won’t that be more entrepreneurs then? And then they’ll get bought and maybe they won’t combine the teams or something. …
Andra: What we’re saying to our clients is, we’re not here to try to replace your existing people. We’re here to help them do more high-value work and then you can grow significantly without having to hire. With your current team, just like you said, you can do way more. You can grow a lot before you need to hire anyone.
Leleux: I have a son who’s going off to college and it’s so scary for them. … They’re looking now at, OK, where AI can definitely not replace me, like a plumber?
Orr: You just touched on our problem. We’ve got a huge labor shortage in the construction industry and the trades are aging out and no one’s going into it. So, there’s this big gap and, and it could help fill that so that everyone can go in and become a plumber, because we really need a lot of plumbers.
Washington: I’ll put a plug in for Salt Lake Community College, because I’m a board trustee there and they are really good at their trade programs. These are jobs — whether it’s plumbing or welding or whatever — and they can walk out, newly minted, and have a six-figure job, six-figure income. They are expanding the programs because of the demand.
Orr: We also need more apprenticeships, I think.
Andra: Our talent acquisition teams are going into junior highs and high schools. It used to be college info sessions and now we’re trying to change the perception of what the construction industry is.
Leleux: At the Women Tech Council, that’s what we’re trying to do, is show young women that there are all these career opportunities and to bring to life how you could have a STEM career in construction or in whatever industry would be massive.
Andra: Also manufacturing. They’re really trying to onshore all this capability now and ramp up U.S. manufacturing [because there’s] not enough good people to run the manufacturing facilities.
Freeman: So, as far as developing the AI, I guess internally, are you sourcing your talent here in Utah? How are you handling that? Anybody?
Barnes: Yeah, I guess I’ll start. I mentioned we do have a couple of folks in-house. They’re in-house in terms of our employees, but they’re actually outside of the United States. … I’ve talked with many different firms in the United States, and, and, honestly, cost is an issue for us, and a little bit unique in our business because we’re dealing with the tariff fun as well. So we’re trying to manage that at the same time and manage our costs appropriately. But, yeah, at the moment, all of our developers are currently outside the United States, just from a cost perspective. And … we haven’t done this, but you can get on Upwork or Fiverr or whatever and put a job posting out there and you’ll get thousands of responses from developers who say they’re AI experts, right? And they know everything about it, supposedly. That’s another problem too, is there’s just so many people out there who say that they’re in the business of doing all this and maybe don’t have the experience that they say they do. It’s hard to cut through the noise.
Freeman: Well, that was going to be my question. How would you do that on the front end?
Barnes: So far, we’ve done it completely a hundred percent by word-of-mouth referral. We need some of the tools that Caitlyn produces.
Brooksby: We have about 30 people in the company and more than half is tech — you know, engineering and our speech scientists research team. Most of the team’s actually in Utah, so we’ve been able to hire local engineers. David Brown’s our VP of engineering. He came from Verizon and was able to just work for 20 years, worked with great people here in Utah. But what we’ve done is, when we’ve had to expand and build out specific iterations of our technology. We, again, went to Microsoft and have an approved vendor through them, Upworks. And so we utilize them. So, yeah, the word-of-mouth, going to someone you trust already as a partner and then finding who maybe they utilize. …
Russell: Do you use your tool to screen new hires?
Brooksby: No. No. I mean, what are we looking at? We’re looking at depression, anxiety. You know, we’ll still hire you if you have that. And then we … MCI, mild cog impairment, Alzheimer’s, Huntington’s. … I was talking to Shaun that we’re building out aggression models for protecting nurses and clinicians because … I think, like, eight out of 10 nurses dealt with violence in last year. It’s very, very, very high. And so looking at safety is an area that we’re exploring. It’s a screening tool. …
Andra: But you brought up a point that I think is so key, which is that human connection, right? Those referral networks. And I think in the age of AI, where AI is doing so much and everyone can claim anything they want online, the human connection aspect and who you know and what you’ve done for them and trusting other people and their recommendations, is huge. For example, what we do, there are thousands of companies that claim to do all the same stuff. And our clients like us because their attorney or their accountant knows that we can deliver and they refer them, so those human connections are going to become at a premium and you’ll see that actually accelerate in tandem with AI adoption.
Leleux: A friend of mine just built a company called Interactive EQ, and it assumes that you’ll be looking for, do you have the right financial acumen and background and accolades and things like that. But this does the emotional EQ and it makes you write an email. It has you in an internal meeting and then a client meeting, and it’s all through AI and all through just the computer, and it just scores you on what’s important to the company, on their EQ. … I love the product. I have nothing to do with it, but I love the product because I think it’s so important and that’s where so many hiring decisions go wrong.
Barnes: It made me think of 20 years ago when I got my CPA license. All of the effort that I had to go through, to go through that process and get the certification.
And since then with even just with ChatGPT, I’ve used it a handful of times looking for technical accounting interpretations, and it’s been wrong three out of four times. And my fear is we’ve got all these new folks coming into the accounting industry, the finance industries or engineering or whatever it might be, and they’re relying on these AI models to tell them what they should know or what they should do without going through the hard work that some of us had to do. … It’s hard to know, you know, again, if you don’t have that background of the hard work, right? How do you trust but verify, without knowing ahead of time? I’m sure AI will improve. It will get better. It’ll start giving right answers more often.
Brooksby: I think Generation X is even worse actually with adopting AI.
Barnes: Well, in our industry we have a lot of robotics expertise, AI expertise, and we’ve found similar issues when we post a job, a posting for whatever type of engineer doing robotic control or perception or higher-level software development, we get a lot of fresh graduates who have basically no experience. And what they’ve been taught in school is for AI, you just take a Python script, you import a few libraries and — voila! And that’s not how the industry works. People with a lot of data know-how to know the underlying mathematics and things. The tech industry is highly competitive, so it’s really hard to find people that have the AI or robotics background plus a few years of experience.
Freeman: It sounds like the technologies that are trying to utilize AI need some kind of gatekeeper. I don’t know who that would be, but couldn’t you see a world in 20 years where the AI is dumber because it’s got less knowledge coming in? So, it’s feeding on its poorer input coming in and it gets worse and worse and worse, potentially. How do you, how do you gate-keep that?
Andra: That’s obviously a big topic, and people are talking about synthetic AI to fill the gaps, where you can have now AI create a bunch of new data that’s sort of hypothetical or synthetic to continue to train itself on. I don’t see it ever getting dumber. I don’t think that’s plausible, but plateauing? Certainly, that’s one possible future where you just see it kind of plateau out in this kind of asymptotic relationship.
Orr: I was just going to say, in that 20-year time span that you’re talking about, I think that a lot of people believe that general AI will be available at that point, which I think would be the opposite of it getting done.
Washington: Well, AI is only as good as it’s been trained. … So it scans the whole Internet, right, for example. Whereas it’s continuing to get learning and training. It’s us talking to it, giving it contextual understanding, some knowledge,
some experience.
So there is that element of, can it keep up, as long as we continue to provide that information? But if the world is filled with AI slop, absolutely, it’ll just train on slop
Leleux: Which is scary.
Freeman: So, looking out just in a shorter term — you know, just a couple years — what are the trends? This is a broad question, but generally speaking, what kind of trends do you see in the way that AI is being used and the directions that it’s going? I’ll just throw it out to whoever wants to try to tackle that one.
Leleux: I just love what Sunny said. I would hope that where it’s going is that every person can find that ability to save, save time, and then reinvest it in something that really matters.
Washington: I had an early experience of … 15 or so years ago. I worked at a company where we had a digital literacy certification. So, all teachers had to take this, and it was the first time that computers were being introduced into the classroom. And in Utah, the teachers had to take this certification so that they could use the computers. And I cringe at it now, because I look at the types of things that we made them learn, which is, like, what is an operating system? Well, who cares, you know? But I had an English teacher that came to the testing center and she failed those questions, which caused her to fail the test eight times. And I think about that a lot, because she was so upset and said, “Maybe I shouldn’t teach anymore.” And like we would look today and say, “Oh, my gosh, no. Just teach English. It doesn’t matter if you know what an operating system is.” So I think that we live in a world where AI is extremely exciting and we can talk about the models and what’s coming in agents and CPS and throughout all these acronyms. But at the end of the day, we’re all human. And I don’t know if we need to know all the ins and the outs. Like, people building companies and setting up the frameworks and all that, that’s important. But as regular humans, if it makes our life better, gives us more time, like, wasn’t that the goal anyway? I think about that like that teacher all the time as I’ve been doing this.
Andra: The big trend we’re watching at Talbot West — and we’ve developed a thesis around this — we call it our five-year thesis. The five years might be off. It might happen in three, or it might happen in seven. But we believe it’s going to happen is a trend toward total organizational intelligence. Think of it like a central nervous system that runs throughout an organization. So right now, when we think of AI, we’re thinking of it as these very kind of isolated point solutions: “Oh, I think I’m going to deploy AI for this thing or that thing within my company.” And it’s kind of a standalone, but we see where things are headed, is this total orchestration across every department. Data is synchronized, efficiencies are like Surface that you wouldn’t even imagine today. It’s going to be phenomenal, and it’s a lot to get from here to there. I mean, a lot of systems are going to have to be almost, like, rebuilt and to get this level of orchestration. So, we’re helping companies like, how do you start getting advantages today? Like that’s great that that’s what it’s going to be in five years. How do you start moving toward it in a way that you get immediate advantages, but build incrementally toward that.
Freeman: If you’re approached by someone who’s just starting a business, an entrepreneur or a small-business owner, any advice for them on how they would go about best implementing AI into their new business venture? It’s going to depend on the business, I’m sure.
Russell: I think form policies and definitely form a strategy about it first. Don’t just dive into it. And I think, at least in my company, I see a lot of vectors. People have their own perception of how it’s used, and it’s helpful to them. But we, again, small company — 30, 35 people — we don’t have a policy. But I think that’s kind of the important thing for how it evolves. And kind of tying into your last question about how does it evolve? How do we implement it? There needs to be some known limitations and you need to educate your staff on what it can and can’t and should and shouldn’t do and be used for. I think it’s really early and I think it’s a little schizophrenic right now in its implementation and use.
Orr: I agree with what Jim said. I wanted to add that what I would tell anybody starting up is, one thing that we all have that AI doesn’t is, we’re incredibly adaptable as humans. So just stay adaptable because it’s going to change constantly every day. Just, you’ve got to be ready to roll with that or you’re going to get run over by it.
Washington: I do see I think it was Sam Altman that said in this generation we’ll see the one-person billion-dollar company. What does that look like, to have that much wealth created by one person, and value? That could be incredibly interesting or it could be a disaster. I don’t know.
Andra: I would say in terms of advice to new business owners dabbling with AI, two things. One, max out the capabilities of these easy-to-use commercial tools like ChatGPT, Claude, Perplexity, Gemini, etc. You can get so much out of them if you learn to use all the different tricks, custom GPTs, things like that. But also know what not to use them for, what those security limitations and things are. So, have some boundaries around that. But yeah, that’s where you’d start. Get full functionality out of those tools. And then this is self-serving, but I believe it’s totally true: Get some professional help. Just like you probably wouldn’t want to try to file all your own patents and do all your own IP work, get some professional help that can come in and actually help you create a roadmap. That’s “OK. We’ve maxed out these easy-to-use tools. What’s next? How is this all going to fit together? What’s a comprehensive plan for our organization?”
Bybee: We have used AI basically in three ways. The first way is, how to use AI to develop the product. And then the second is to use AI in the product. And then third is, what does the next product iteration look like? What is product management, how can they learn what to build next? So, I guess my advice then would be, identify how AI could best affect your bottom line for whatever time window you’re looking to optimize over.
Brooksby: AI provides the opportunity for you to be an expert in everything, and you can get lost in understanding yourself and what your true talents are and your true interests. I think about our founding story and it really focused on our co-founder. He developed Alexa far-field speech recognition through Amazon, purchased his startup out of Boston 15 or more years ago, and he was really good at speech. We took a stake in the ground. We said, “We’re going to be the best at using speech and AI,” and we’ve stuck to that. And as I mentioned, we’ve developed the technology. We own the most patents in the industry. We didn’t say that just because we’re in health care, we’re going to do this, this, and this and this, right? And so I think just really having that motto for yourself, of what you’re trying to accomplish, will allow you to use AI to become better. But you won’t get lost in it.
Freeman: Is there any topic that we haven’t covered that somebody wants to touch on?
Russell: One that was kind of a reoccurring theme is, we have these open AI machines that we feed information into and we know they’re very cognizant because they can track what we talked about before and they tie it all together over time. … What we use it for also is product development, research and development. And back to, can we trust it, as well? So those kinds of things culminate together to protect our intellectual property. If we’re an R&D company or … we all are developing intellectual property in our own way, and if we’re using these open models that we’re feeding information into, tie that together with a cloud computer — just another name for someone else’s hard drive — and is our data secure? And I’m not an attorney. I don’t go and read the policies on Grok. I use Grok. And it’s like, is it learning from me or someone else? And how are we protected? How is our intellectual property going to be protected? And maybe it is more local server, local agents that are assistive to us that are firewalled to have access to information, but maybe more of a one-way pipe. …
Freeman: Yeah, that’s an important point, because from a patent standpoint, if you’re working on an invention, you’re supposed to keep it secret until you’re ready to file your application. And if you’ve disclosed enough of it that it’s out there on a server somewhere, I don’t know the parameters of that.
Orr: Another point is that it’s already here. So you don’t have to make a decision as to whether or not you want to work with AI. You already are, if you don’t know it.
Russell: So maybe that even gets into the legislation of AI in general. And a lot of lawyers [are] involved in that. keeping firewalls around the information that we upload onto these clouds. That may hamper its ability to expand and be adopted if we’re running into intellectual property problems along the way.
Freeman: There is a legislation recently passed in Utah that seems to me to be the opposite of the DMCA safe harbor that you hear about. It basically says you can’t blame AI. Have you guys run into that?
Washington: So the state does have a sandbox or through the Department of Commerce, you can set up a sandbox environment, but specifically it’s for those that are doing anything that has to do with mental health or health. I guess you can do other things, too, but then it basically provides a framework. So, you can start to build without legislation hanging over your head. It’ll be interesting. I personally don’t love that we would have 50 different policies. The hope is that we just have one, because that would be a nightmare for any company to navigate. So I think we’ll see, but government is notoriously slow to catch up. Tech companies do what we do, which is move fast and break things, and sometimes that’s really detrimental to society.
Leleux: I think the way that I tie it together is, I often go to talks or listen to podcasts or whatever, and people will talk about “I feed it inaccurate data,” or “I feed it fake data about my company” or something like that. So it’s kind of, what is it training on, what is it learning? I did do one where I spent a weekend learning to code and coding some reports that I wanted, but I did feed it just fake data. And then I was, like, well, once I get the code, I can put in my own data for the company. …
Orr: One thing that didn’t come up today, which I was expecting, is cybersecurity and how AI is going to disrupt all of the traditional things. I’m hearing Jim talk about firewalls and that’s over. I think AI is going to completely disrupt cybersecurity, especially since it is outside the walls of our government control and everything else. So it’s a big concern. It’s something that companies should be thinking about.
Andra: It seems like it’s a bit of an arms race where bad actors are using AI to more rapidly and at scale, exploit both find vulnerabilities and exploit them. So they can just do that at a far greater scale. But then the cybersecurity providers, the providers of these solutions, are now using AI to try to more rapidly detect intrusions, patch them, and all of. So it’s a bit of an arms base. AI versus AI.
Bybee: The stakes are really high in certain industries because you’ll sometimes, if you rely on AI or something for minute-to-minute operations, you can easily be susceptible to denial of service, for instance, or to bad actors coming in and shutting your system down for large amounts of time.
Barnes: I’ve already dealt with this with our insurer on our cyber policy, because we are in this process of building out capabilities and we’re already seeing our rates go through the roof because, how do you ensure for what you don’t know, right? I mean it used to be antivirus. OK, you got your firewall placed, you got your PII protected, and so on. Now, it’s like something could come next week that you didn’t even know existed and damage you. So how do you insure for something like that? It’s crazy.
Andra: Well, and to your point, Parrish, I don’t think it’s just that it’s going out to something outside and coming back. because we’ve been doing that forever with cloud infrastructure. It’s that there are specific vulnerabilities with these AI tools, the specific types of attacks that can happen through them. Prompt engineering type of things or trying to get it to release data or inject certain things. I think that’s all pretty manageable with the right security infrastructure. But, you know, there are vulnerabilities that are very time-consuming to sniff out, and then these bad actors are able to deploy AI that can go sniffing around and find these at much higher scale, so it just increases their capability
Freeman: I want to end on a higher note, so let’s talk about some positive things that we see coming through AI, just generally, I guess, in your businesses.
Brooksby: I think the investment. If you look at where VCs are investing, I mean, AI is getting up there right now. I hear a lot of people saying, “AI.” People are, they’re like banking on AI here, and I think that’s good. I think there’s a lot of negativity out there in the world, and if we can find positivity as humans. … We’re going through a technology revolution that is akin to the Industrial Revolution. It truly is. So I think it’s an exciting time, a good time to be alive.
Russell: Vertical industries are benefiting from it hugely. Semiconductors are going to be needed a lot more. Also power to operate all these large data centers. And we can’t expand power fast enough. I don’t think wind and solar is going to get us there. The traditional methods, I think things like distributed small modular reactors and things like that are going to be the future. And that’s whole new industries that we get to evolve and that’s going to be huge for our economy and scaling.
Andra: One bright side I see is that AI is going to do a lot of the low-level work that humans don’t really enjoy doing, and humans will be freed up and positioned to do some of the things that they’re uniquely good at that aren’t likely to be replaced by AI. And one of those is human relationships. Humans are going to be better for the foreseeable future at human relationships than AI. And human relationships are incredibly important in a lot of industries. So, humans will be able to do more of that side of things. Also the high-level strategic thinking of bringing the big picture, synthesizing across multiple disciplines. I think for the foreseeable future that’s going to be very needed. Judgment. These sorts of things that humans will be freed up to do more of that, the stuff that they can uniquely do, and not have to do a lot of the stuff they don’t really enjoy doing.