The streetlight effect: the hidden bias undermining your AI strategy
32:16 Share

Welcome to today's episode of Lexicon. I'm Christopher McFadden, Contributing Writer for Interesting Engineering. In this episode, we sit down with Justin Graham and Jake Dean from Barge Design Solutions to explore the real-world challenges of implementing AI in engineering.

From avoiding the streetlight effect to saving clients millions through more intelligent data integration, they share hard-earned insights on why clean, connected data, not just flashy tools, is the key to meaningful innovation. So if you're thinking about AI, this episode will show you where to start. Jake and Justin, thanks for joining us. How are you both today?

Wow. How are you? Very well. Thank you. Doing well. Happy to be here. Good stuff. For our audience's benefit, can you tell us a little bit about yourselves, please? I'm Beth Goulder, so Jake can go first and then Justin. Sure. Jake Dean. I'm a technology solutions developer at Barge, and I started my career as a civil engineer. Wow.

But I quickly got interested in geospatial technology and applying technology in engineering context. And over the years, I've slowly migrated to designing and developing engineering software. Fantastic. Thank you. Justin? Yeah, Justin Graham. I am the director of Barge's Information... I'm sorry.

Justin Graham. I am the Barge's Innovation Solutions Center Director and started here at Barge about 11 and a half years ago and over the course of about 22 years of applying technology in the architecture and engineering space. Fantastic. Thank you very much. First question then, Jens.

Can you start by explaining what the streetlight effect is and why it poses such a significant risk when organizations adopt artificial intelligence? Yeah, sure. So it's a bias. It's an informational bias. And it's where we tend to use data that is easily accessible to us, but not necessarily the best to answer engineering questions. We borrowed the term from an old parable about a person looking for their keys in

at night under a streetlight, not because that's where they probably dropped them, but because that's where it's easiest to look. And in our roles of applying technology in AEC, we noticed the shift to cloud services that were making certain processes very, very easy and easy to access data that that service collected and organized. But often our engineering processes rely on multiple systems and

It's not easy to integrate that data between those systems. And so that leads to using the data that's easily accessible rather than all the data that may be necessary to solve the problem or answer the question. Okay. Okay. And is that a common problem for most companies adopting AI nowadays? Kind of a short-sighted or tunnel vision kind of approach to data?

Yeah, so we've seen it, you know, in our domain in AEC, but I do think the problem is general across many different industries, because AI to be really effective, and of course, needs to leverage all the data that is relevant to what it's working on. Okay, great. I guess that brings us on nicely to the next question then. So you talk about turning on all the lights before investing in artificial intelligence. What does that look like in practice for a company?

Yeah. You know, when we say turning on all the lights, really what we're, what we mean is that creating that visibility into either your businesses systems and the data and,

that we think can create the most comprehensive picture. It's going to aid you in that decision-making. And even more specifically, it means integrating all the siloed data that your organization uses in those siloed systems. Once all the lights are on, so to speak, those tools and the solutions that use them become really powerful. It's not just generative AI-based tools.

that take advantage of that, but you know, reporting or dashboards or knowledge bases and, and more, all those become pretty strong in that process. And, but in practice that, that really means just involved. That means connecting data and from design specifications and guidelines or historical project deliverables or financial platforms or HR systems, customer relation management systems and customer service tools and inventory databases and you name it.

So the end solution has access to all that context-rich data, clean data that we think is really helpful in the moment that you need that information. Yeah, the specific example that I like to point to is Barge was moving our headquarters about five or six blocks here in Nashville. And we owned this building and it was kind of a warehouse and

It was stacks and stacks and rows of boxes of physical data, physical information, paper. Bars have been around for 70 years and there was a lot of information in that building. And we went over there to look at it because we're going to get rid of the building and we had to do something with that information. And we're staring at these, looking up at these rows of boxes and boxes.

And nobody wanted to throw them away or they didn't want to do anything with them because they were afraid. But the question was, is like, well, there's boxes of information, but we don't even know what's in there. If you know what you're looking for, you know, it's in this building. It does you no good if you can't find it. And that's there was an that's an example, a real life example of there were no streetlights on in that entire, entire room.

That was going to be my question, with all the companies that have moved from paper-based to computers, how on earth do they get that data digitized so they can use it with AI? What did you do? Did you scan them in or something? Yeah, first off, it was so overwhelming that we actually had to take some kind of approach to it. And then you just begin to look at like...

age, you know, they at least had some kind of labeling on them of when that box was created and stuffed with a bunch of information, whether it be useful or not. So we had to kind of like find a date range. And then after that, you, then you sift through this stuff and you find the meaningful information and yeah, you scan it and you bring it in to kind of digitize it in some way that then will allow you to apply some really, you know, insightful solutions on top of that. Yeah.

Yeah, so you mentioned you set a date range. That's going to be my next follow on one is I guess you would set parameters between like any data beyond 20 years is out of date for what we're trying to do. It's just a kind of I don't know, how would you say, like strategy strategizing what you need for data you're after what you're trying to achieve, right? With this turning the lights on approach. Is that correct?

Yeah. Yeah. It's a, there was a mix of, there was two purposes of finding a date one for like, all right, at what point do this, you know, design practices and principles and guidelines, what, at what point do they age? So there might be some good stuff in there. And then after a certain date, it's not necessarily applicable anymore as technology has evolved. But then the other interesting kind of approach to a date when it comes to project specific information for,

and engineering firms is actually like how long are we required to keep information around? So that was another reason to keep a date in mind when it came to sifting through rows and rows and stacks and stacks of paper. Yeah, that's a good point. Your hand like before, Stripe. If you've only got to keep the records for 10 years, then that's a matter of data doesn't exist anymore. That's a fair point. Okay. So in your experience, what's the most common mistake companies make when they first try to use AI or advanced analytics?

We see it usually a lack of a clearly defined problem. There's a quote that we like, a problem well stated is a problem half solved. So that's one. And then also just understanding the true capabilities of the technology. So it's really kind of two sub issues there.

For the first issue, it's important to differentiate between the perceived or real problem. An analogy that we sometimes use is taking your car to a mechanic, which I just did the other day. And you might tell him, hey, you know, I need a front end alignment. So that's like a solution to your perceived problem that the car is veering to the left or right. And a good mechanic will first confirm that we

real problem. And then, and only then can they define and implement the best solution, whether it's a front end alignment or not. If they simply implement a front end alignment and that doesn't solve the problem, the car still veers to the left or the right, nothing has been accomplished. And we've seen this with, you know, efforts to implement AI without a clear problem. Just nothing really gets solved. You know, it's,

Turns out the ability to transcribe your emails into old English or something isn't that applicable to many business contexts. So clearly define the problem. And then the second problem is just generally referring to technology in a general way, glossing over some details that may be important to what it can actually solve.

And that one is a little bit easier to solve. Just bringing on people that are able to understand the nuances between different proposed solutions early on in the process can really help solve that problem. Or even if you need AI at all in the first place, presumably. Everyone's trying to jump on it. It's a buzzword and you can use it to sell your products. But do you really need this?

So with companies bringing it in for the first time, how much work is involved to get the AI to do what you want it to do? I know it's a very broad question. I guess what I'm saying is you don't want to just adopt an AI and then trust it immediately because it could cost you more in the long run if you start making judgments on what it's doing and it's basically an error. Am I correct there? Yeah. So...

One example that we...

you know, we worked on late last year was a clearly defined problem that someone came to us saying, hey, could you help us increase or improve the way we write health and safety plans, which is something we often do for a lot of our projects. And before, you know, it wasn't, hey, let's just, you know, use AI right away. We first, you know, took time to define the problem. How do we

improve the process and increase the quality of the health and safety plans. And then, and only then did we design a solution. And we first made sure that we had all the data that we needed, uh, that is required to write a quality health and safety plan. And then, and then we saw, Oh, actually AI is really, really good for this. And so implementing wasn't that hard, um, because we have all, we had all the data ready to go. We were able to integrate it. So that was the hard part. Uh, did take a lot of effort, but then actually applying AI, um,

In this case, generative AI to help write a first draft of the health and safety plan was very relatively easy. Presumably it's critical to get the staff who are day-to-day doing these things involved, especially health and safety, for example, director or whoever, must be involved as well to oversee the process, presumably, right? So don't get down the wrong track.

Hugely important. Yeah. And solving, applying technology to solve problems. Hugely important to have that domain experience and really detailed knowledge of the workflow. Absolutely. Yeah. Great. And even, you know, going even further after that, helping us in the design of it, but also owning that application and owning, driving the usage of it internally as

to the organization just is extremely important having that owner kind of take it after it's been developed and drive the usage of it. Yeah, excellent. Great. Okay then, moving on to an actual example then. Can you tell us about the, I think you had a one million pound savings you provided for a water utility company you worked for? What was the core data problem you uncovered and how did fixing it lead to those results?

This is, it's a fun case study, you know, because it is, it started with, it actually started with a goal at the, at the customer level to improve customer satisfaction rates. That was kind of the big, the big goal. You know, what do we, how we got, how can we do that? And the approach we took was actually to look internally at that organization and try to find people.

where is the inefficiency and try to find the silos and maybe some process gaps or that we think would lead to like maybe a better customer experience. Um, technology and data, we, they were applied and they helped us accomplish that goal, but you know, it, they did not improve the customer experience or generate that huge savings that you mentioned by themselves. Um,

It was really the modernization and the digitization of outdated workflows that we think lead to these kind of results. And one major component of that

project was actually a data integration effort. It was turning on the streetlights, you know, it was integrating in systems and data. They had great systems that were ready to contribute to that fuller picture. And it was, you know, those things were like materials inventory and purchasing software and asset condition assessment systems and human resources systems and

customer service systems, all these things, financial packages, fleet management, SCADA, great systems that were just ready to kind of, they were doing great things in and of themselves, but they were better served contributing to a fuller picture. And so that's what we did. And we integrated those systems and we saw that workflows were drastically reduced. And there was a very positive impact to the operations, we think, that improved like maybe service delivery and

at the same time saved a tremendous amount of money in a pretty short amount of time.

And we felt like kind of that savings has from this type of technology and integrations, it's always been there. And I don't know, AE businesses and we've historically, maybe it's just me, we've historically had a hard time quantifying that time savings and that cost savings that we're seeing as a product of kind of true data integration and applying technology.

especially now that you're throwing AI-based solutions on top of that kind of technology. It's like 10x, 50x, whatever you want to say, return on that investment. It's pretty staggering when you take the time to quantify that ROI and to hear you say a million and to even see it is really even surprising to us as we get to apply this stuff and see that savings. Right.

And presumably many large organizations, they'll have very efficient, like you mentioned, very efficient systems and silos of data, different databases here and there. But if there's no oversight connecting all the dots together, which the AI can do, you can only need several experienced people talking to each other, right? Whereas the AI will just do it in the background automatically 24 hours, seven days a week, I guess. Yeah.

Yeah, that's what some people claim. And in some cases it can happen. But, you know, if anything, if anything we're doing is creating a better picture of really what's happening within that organization and, you know, helping. If AI can step in and automate some of that repetitive or that mundane task of centralizing and reporting on that information, that's awesome. We're really interested, though, in like...

the, the effectiveness of people as they, you know, the term human in the loop and those kinds of things. This is really the place that we want to be as like augmenting kind of, um, uh,

professions and bringing the information to them and in a much quicker way there's a lot of time spent in in finding information and searching and we're trying to shortcut that absolutely and avoiding the junk in junk out yeah like missionary with some programs fair enough um

Okay, how do you convince clients that slowing down to fix their data first will speed things up in the long run? Well, yeah, so Jake and I, you know, we have been fortunate, I think, to have a lot of time or some time to at least design and develop and implement some tools that we think solve some pretty real-world challenges and challenges.

And when you take the time to, as I just mentioned, to kind of calculate that time savings and calculate, quantify the benefits of it, and you show that to either your internal leadership, if you're trying to get buy-in or the investment of taking some time, slowing down to fix data, I think that goes a long way. And once, you know, either clients or leadership, when they realize that, how much time that really, as I just said, like it teems to,

spend on searching for answers, you know, across all these fragmented systems or, or how much risk that, um, that they take about even relying on some bad data. Uh, they start to see quickly that data integration, data management, or even just, even just simple, good data hygiene. It really is like an accelerator to investing in that. Not, not a delay. I, um,

You know, don't get me wrong. I do think that that is a it's a really hard thing to do. It's a and I think it's probably overlooked. That discussion of like investing in that is probably overlooked because of how hard it is. And it's not very fun.

to talk about, you know, versus like the exciting world of generative AI or, so there is that dynamic that we typically work through internally and with clients. But we found that when we share those kinds of results, like as, as we just mentioned this, the time savings, it does get people's attention and,

Another specific example, that kind of that point that brings people in is just the searching of information alone. We created this really simple search and filtering tool of archive data. It's just archive project folder. We have a network drive somewhere and it's old stuff. Barge is actively saving estimated $85,000 worth

US dollars per quarter with that tool alone, and which equates to 350,000 US dollars a year. And that just comes from a reduction in time spent for looking for information. I think that is amazing. And it's really very significant. It's an impact that I don't, I don't think can be denied.

couldn't agree more yeah um i've been built some databases in the past and yeah what i call data cleaning um yeah it can it can be an absolute nightmare um it's usually sometimes it takes longer to do that and actually build the system itself it really is a big job um so you can get that right from the start you're on to a winner really um

Anyway, with generative AI being so accessible now, are you seeing more cases of companies building flashy AI solutions that don't solve real problems? I'm kind of touched on this. Yeah, no, I think it's important to go back to because it goes back to, you know, integration just fundamentally being hard and something that really needs to happen within an organization.

So again, I mean, a lot of these cloud services are really well suited to handle specific tasks. But then you have to connect those different systems together through their application programming interfaces, they're called APIs. But that is hard to do. And it takes a lot of critical thinking to do it in the right way.

And so that's where we've seen examples of solutions that can really meet expectations. And the one that we mentioned before, the health and safety plan assistant, I think is a good example of that. When you are able to set these systems to talk to each other, then you can really create these almost magical experiences. The health and safety plan, these are usually 20 to 30 page documents.

when written well and they can take up to eight to ten hours to get to a first draft um that was a manual process under this new project with the health and safety plan assistant we can get to a first draft in 10 to 15 minutes and then you have more time to critically read it and and go through it um to put your critical thinking in there to even increase it further um and

And so, again, that integration is hard, but so worth it. And one thing that we've noticed is a lot of IT departments are ill-suited to actually integrate those systems. They're very well-suited to stand up these systems and protect them and secure them.

to do their function of what they do well, but actually there's not a lot of people that have the skills required to integrate them together. And that's where we see a huge need. And I think companies that are able to build their IT and other, maybe it's a different department with people that are able to integrate that, I think they'll have a huge advantage and able to create AI solutions that are really effective.

Absolutely. Another possible benefit, especially with a health and safety policy, I'm not sure about America, but in Britain, the regulations change quite often and keeping your policies up to date, referencing the right clause of what can be an absolute nightmare. Presumably AI is well placed to help with that sort of review your policy and make amendments as, as, as when required. Absolutely. Part of that one category that, that,

that solution fits into is like kind of compliance and making sure that we're complying with all the applicable health and safety regularizations within that state or region of a state. And so that's one part of that project that just goes out and say, Hey, am I, am I referencing all the appropriate safety codes for this project? Which is, you know, would take a human a lot, a lot of time to do and not very exciting, but it's, it's a lot easier for an AI and they're happy to do it. Yeah.

Absolutely. Yeah. Yeah. Understatement. Yeah. Not fun. Not a fun job. Um, okay. How, uh, how do you approach finding the right data in August organizations that have siloed or fragmented systems? You know, in Jake and I's role here, barge, it's really important to kind of understand the challenge that's to be solved, you know, internally, or that a client needs help with, um,

We get to then do the thing that, as Jake just kind of alluded to, what humans are good at, the critical thinking part.

Um, it's a, it's a fun exercise I think when, uh, because everyone in the room works to kind of a collective understanding of what needs to be done and maybe even knew it all along. But, uh, I think we, we, we can get there with a little bit of, uh, of critical thinking and it just takes kind of really defining that current state. I think in all the broken processes that are involved in that, whether it be workflow processes or, or, or siloed systems, um,

And then we then like to kind of go to the most, like the ideal outcome and trace back, I think, to, you know, which systems can be integrated and where automation can be implemented to get that kind of that desired state.

where you really have improved to really a truly state of data-driven decision-making. And we ended up finding the need to clean up or elevate messy data sources like spreadsheets and such, as well as there's a lot of the, well, this is how we've always done it kind of process. And modernization of that is really important as

And then from there, it's just all about like kind of unifying those sources and modernizing those processes and,

really allowing each system to contribute to kind of that fuller picture, that bigger picture, more comprehensive look at what you're trying to accomplish. Presumably this would be, it's pretty an obvious thing to say, this would be a lot easier for a startup, a new company that's just getting going rather than like with Barge, 70-year-old company. I have an analogy, it's like trying to change, they're not a big old company, it's like trying to change the wheels on your car when you're driving at 70 miles an hour rather than when it's at rest.

I presume. I don't know if you agree. Good point. Yeah. Good point. And, you know, in even in like a company like barges have been around for 70 years, it's you're like, well, okay, this is really hard, but like when it's kind of distilled down to a specific need, it,

becomes a lot easier. And finding that right data, just always taking the approach of like, you don't have to solve for everything right now. Let's at least take action and work towards some kind of step in the right direction. And if it's a subset of a big old mass of data, that's okay. That's good too.

Yeah, that makes sense. Plus, you're in an older company is experienced, it knows really what it needs. They will have a lot of data to Yeah, I guess there's pros and cons for startup and old company benefits for both. Good point. Yeah, yeah, yeah, there's benefits for sure. Great. Okay, looking ahead, how do you see cloud services, AI and data management evolving together over the next five to 10 years?

So, you know, we talked a lot about the importance of AI solutions having access to data that's required to do whatever tasks it's tasked with doing. And I see, I'm hopeful or optimistic that that basic need will help integrate disparate systems. Because again, a lot of these, you know, systems, they have

the engineers that built them have created these nice APIs, but there's nobody or not enough people that understand how to interact with that to fully get the data out and integrate it together. And so I hope that the awesome power of AI to integrate data when it gets the right data will help drive that where we can

easily and securely go and talk with the system to create these data streams that pull from disparate systems. But I think it's going to take a lot of, you know, it's going to take effort and focus and critical thinking to say, yes, this is the type of data that we need to get. And this is how we integrate it. Do you see any potential risks with data leaks and things with that kind of method? Yeah.

Yeah, it's a balance and it kind of alludes to what I mentioned before about a huge focus of IT is security, right? Data security. And that's

somewhat uh like opposing data integration because you have to go in and get the data out right but it's it's like uh an art to go in there and um work with with different teams to say okay this is what we need and you have to be clear but like this is uh you know the precise data and this is how it's going to flow you know to the other system so it's definitely a balancing act um

and again, yes, I think there's always that risky, like I'm, don't get me wrong. I'm a huge proponent of security and keeping things safe. But yeah, you have to find that balance, which is, it's going to be hard. Absolutely. It's more of a problem if you need to access the data remotely. If the system is fully contained within your network in your building, it's less of a problem. But yeah, if you've,

you need to have remote access to the data, then yeah, that's when it becomes an issue. Um, hey, last question then guys. Um, if you could just give one piece of advice to a company excited about using AI, pardon me, AI, but unsure about its data readiness, what would it be? Well, uh, real quick, the benefit of having two guests is you get, uh, two pieces of advice. So I'll go first real quick and just quickly say, uh, from my perspective, you know,

Jake and I, we really appreciate when people act towards these big initiatives. I just take action. This is,

this kind of initiative is a hard, it's hard to start when your conversations are a little bit too general. You know, AI is, we think that it's kind of an easy and quick thing to say because those two little words, um, but there's, there's two little words that kind of mean so many different things to so many different people. And even the same thing can be said about the word integration. Um, it doesn't really mean, you know, it has a meaning, but it's, it's really general. Um, and if those types of general terms stay in your vocabulary, uh,

getting towards action becomes really hard. And so my advice, I would look for, you know, a specific point of friction within that business and, and, and just apply that critical thinking. And, and then you'll have some clarity on that solution for that challenge. And you're really not too far from solving for that specific challenge. And so, you know, just once you take action, that big initiative, I think becomes an easier thing to accomplish. Yeah.

Very good. Yeah. And my advice would be the question of asking yourself, how do you know what you know about your business or what problem you're trying to solve? Because we all carry beliefs about

what we think we know about our business or what we think we know about our engineering problem, but we can't always point to the exact data that backs up our belief. And the simple act of asking that question, I think is really helpful in pinpointing what data you're missing or what you need to go collect. And then you can go create, then it's easy to create the plan, um, to integrate the data or collect the data that you need. Um,

And the great thing is that every business should be doing that regardless whether they want to use AI or not. Just simply trying to avoid that streetlight effect is very relevant across any business. How far should companies go to get their staff involved, especially those with the coalface, with what they need when they're trying to implement AI? Should they be part of the team that helped develop it or at least develop the strategy?

for the process going forward yeah i absolutely i'm a big fan of going to the edge of the organization because i think that's where that the edge of the organization is kind of where it will be applied and without input from that end of it it's um it's not as it's not as advised as i think and and would lead to great design of that that solution as it could be

Absolutely. Brilliant. That's all my questions. Is there anything else you'd like to add? You think it's important we haven't mentioned, gentlemen? I think simply just turn on all the lights and you'll have a lot better and more effective solutions. Sage advice. With that then, thank you for your time, chaps. That was very, very interesting. Thank you. Thank you for having us. Yep, really enjoyed it. Our pleasure.