rtificial intelligence and machine-learning algorithms have struggled to make sense of chaotic construction jobsites, but recent years have seen industry firms build the vast data lakes and analytics systems necessary for these machines to provide useful advice on how to plan, schedule and execute projects. In some cases, these AI advisors have become a standard part of some firms’ project delivery methods. But it’s still a challenge to convince construction professionals to listen to these AI advisors, and there are emerging questions of how risk will be allocated once algorithm-driven decisions start to steer projects.
One of the more direct uses of AI in construction has been the project scheduling analysis performed by ALICE Technologies’ machine-learning algorithm, ALICE. The company has made inroads into the industry in recent years (ENR 5/28/18 p.22), but founder René Morkos says that construction may be approaching a tipping point when it comes to AI adoption.
“What I always hear from people [in the industry] is that ‘I really like scheduling, but the number crunching is the boring part,’” says Morkos. “Why would anyone in their right mind want to spend time crunching all the constraints on a project? It’s mind-numbingly boring.”
Instead, the ALICE algorithm extrapolates thousands of possible ways of executing a project by running simulations of a project’s 4D schedule and BIM, readjusting as variable inputs are tweaked in the project “recipe.” Users make adjustments to the inputs, and ALICE tells them how it will affect the construction schedule. But Morkos says the idea isn’t to cede decision-making to ALICE. Rather, it’s about automating the process of generating possible alternate schedules.
As more companies make the investment into collecting and properly organizing their project data, Morkos says that technologies like ALICE and other AI-based advisors could lend some firms a real competitive edge. “The fundamental value proposition of the general contractor is changing. This new ecosystem will be all about integrated data systems, and it will be 20 to 30 companies that take home this prize,” he says. “We are incredibly lucky to be living in this golden age of construction technology.”
Planning out staging for the structural concrete on M2, a $150-million, 20-story residential tower within the 5M development in San Francisco, Michael MacBean, project director for key accounts at Pacific Structures, saw the ALICE algorithm as more of an informed second opinion on his own scheduling instincts. “We used it on preconstruction for that project to validate our approach to the project and check our productivity,” he says, noting he got the most out of it by tapping into his own past experience as a project superintendent. “The algorithm is awesome. Its ability to calculate every which way to skin the cat, if you will, gets that much better if you also have human expertise on construction to make it do its best,” says MacBean.
On the M2 tower, MacBean found ALICE to be a helpful advisor, allowing him to fine-tune his planning without overworking his team. “In a matter of minutes, you can make changes in the way you do your projects. Do you want the crane here or there, do you want eight-hour or 10-hour days, should you recruit 50 workers or 20 workers?”
MacBean was able to validate his approach for the M2 tower and even fine-tune some of the staging for the structural work. “We were able to look closer at how we were cycling formwork on the project, and I was better able to understand my crane demand,” he says. ALICE’s recommendations convinced him to go with a crane-jumped core for formwork instead of a self-climbing core, since the algorithm showed there would be enough crane time available to make it work. “I could have figured that out myself, but it would have taken a very long time. ALICE does some pretty simple math, but it does it very quickly.”
While Pacific Structures and its parent company Build Group are happy so far with ALICE, MacBean says there is a broader industry issue around trusting algorithms. “Selling the idea of putting all this trust in AI isn’t just a Pacific Structures issue, it’s a hurdle for the whole industry,” he says. “It’s a hard thing to talk about. There are a lot of builders across the country with 30-plus years of expertise on how to build.” But MacBean adds that while it can be a slow process to bring AI-based technology like ALICE into construction, it is winning converts among his team.
With the M2 tower erection now going as planned, the next test MacBean has for ALICE will be a high-rise tower in Seattle, currently in the planning phase. For that job, he plans to use ALICE across the entire project, with the whole team engaging with the algorithm’s scheduling recommendations. “Some of these projects are so big, so complex, it can take years for just a couple of humans to consider every way to design and cost it out. But with ALICE, in a few minutes you can have a whole lot of detail on which is the right way to go on a project. That is really powerful to me.”
Firms Build Out Their Own Homegrown AI
Sometimes if you have the data ready to go, it’s better to just build the AI advisor yourself. DPR Construction has been on a multiyear journey to get the most out of its vast stores of project data and is currently building out its own AI algorithms to eventually weigh in on the company’s decision-making processes. “Some of the machine-learning projects we are working on right now, we’re not calling them AI. We’re calling them ‘AI assist’ or ‘human assist,’ ” says Hrishi Maha, DPR data analytics leader. Rather than chase the goal of an all-knowing AI to wisely guide the firm’s projects, Maha’s team is focusing on building out more focused algorithms to serve in advisory roles, augmenting the decision-making processes of human users. The algorithms will offer insights based on the past performance of DPR projects, even if it means sometimes challenging users’ assumptions about how best to build.
Maha says this automated advice will soon be used in bid preparation and project planning. “The goal is to help our business development, operations and scheduling folks make more informed decisions based on historical data so everything is more scientific, rather than someone’s bad feeling about something.”
Algorithms eventually trained on DPR’s historical data could offer insights into which potential projects align with past performance, and even dial in more competitive bid pricing. Maha says these are just some of the benefits of DPR’s effort to improve its approach to data collection and analysis. “If some data comes in from the field, it’s based on the knowledge of the person filling it in,” he explains. “You came up with that number, but how do you support that number? This shift is about removing the subjectiveness in decisions and moving to a more scientific approach.”
But Maha says that these improvements are all low-hanging fruit compared to what his team is working toward. While a comprehensive construction AI is still far off, even limited AI-based advisors could change up how DPR builds. “Looking a few years ahead, an AI assist tool would make a huge difference. It would improve how we work and remove that subjectiveness in decision making. It would let our people move much more quickly,” he says, adding that he expects these AI assistants to go beyond automating bidding and estimating processes, eventually providing actionable, detailed advice on project delivery, based on real-time data collection from the field.
In the interest of getting more usable site data, DPR has also been trying out the wearable workforce monitoring tags from Kwant.ai. “Sometimes it’s not a lack of AI problem, but a productivity in construction problem,” says Charlie Dunn, part of DPR’s planning, scheduling and production planning team. “My experience with labor-tracking data coming from Kwant is there is a period of gaining trust on the site: Do our ops people believe what they are seeing?” He says trust in the data is key, adding: “Then, once it gets into predictive AI territory, we can say, ‘Hey, you said this was the plan, and this thing says you’re not meeting that plan.’ ”
Kwant.ai collects location data from workers wearing its wireless tags, and the technology firm focuses mainly on providing up-to-date head counts and logging safety incidents. But Kwant.ai is working to apply machine-learning and AI to its data sets, says Niran Shrestha, the firm’s CEO and co-founder. “We never try to sell this by saying it will solve all your problems, but if you input all the data it will provide insights for you to take action,” says Shrestha. With an eye on productivity, Kwant.ai can provide workforce recommendations based on past performance. “Can it tell you how many people you need on a crew tomorrow? Well, for example, say you planned on 20 people on an electrical crew. Kwant says you probably need 30, so maybe you decide on 30 just to be safe. Right now it’s only a reference point, but the data we are collecting will improve that over time,” says Shrestha. He adds that Kwant.ai has found the most traction on industrial projects which already have processes in place for standardizing repeatable types of work.
As these machine-learning algorithms begin to make actionable suggestions, it does raise the issue of liability and risk. Shrestha says Kwant.ai is looking closely at this issue, citing how some project owners are already using its analytics data to make tough calls. “We have CEOs and VPs looking at our dashboards, making decisions,” he says. “At some point, we will take over some of that risk.” But for now, Kwant.ai focuses on vouching for the accuracy of the data, leaving the decision-making responsibilities to its users.
“We think of it like self-driving cars: There still needs to be human hands on the wheel,” says Shrestha. “We could have an algorithm that predicts safety incidents and schedule risks on projects, but the users are still directly involved. We are not saying for them to get out of the way and let us do everything for them.”
Legal Liability of AI Advisors
“A fundamental problem with AI is when its learning gets to a level of sophistication where if a problem occurs we are beyond being able to ascribe it to human fault,” says Joseph A. Cleves Jr., partner at Taft Stettinius & Hollister LLP and a specialist in construction litigation. “99.9% of construction risk issues deal with a fault basis, after all.”
While larger firms that develop their machine-learning and AI algorithms internally can manage the associated risk as part of the larger scope of services they provide to their clients, Cleves says it could get tricky if a smaller firm uses a sophisticated algorithm from a third-party vendor to make important decisions. “In that case, those players are not typically going to have the leverage to negotiate a fair system of risk allocation,” he observes.
It’s these smaller firms that Cleves says may need to be savvy about how much trust they place in any third-party vendor’s AI system. “There is the potential for a small user to have an event that causes an existential crisis in terms of liability. They had the wherewithal to purchase the software but not understand the liability.” Cleves says he could see the standard user license agreements common among Silicon Valley technology startups running up against the harsh realities of construction law at some point in the future, if critical decisions based on AI or machine-learning lead to undesirable projects outcomes.
But construction law is not necessarily too brittle to handle the challenges of AI, and contract language can always be changed to accommodate new technologies, notes Tracy Ickes, associate with the law firm Nixon Peabody LLP. “We might start seeing owners and general contractors get more creative in how they assign liability on some of these projects,” she says.
Ickes and her colleagues have been studying the impact of AI and other new technologies in construction, and she says that some of the contractual arrangements between parties may change as these technologies assume more risk. “A general contractor in preplanning could say, ‘I have this technology that will save a billion dollars and shave down the schedule, but it comes with risk,’” says Ickes. “The owner thinks, ‘Whatever, it’s on my GC to get it done.’ But if GCs can’t shift that risk to the tech companies providing the software, they will need to decide if they want to shoulder it all themselves, or arrange for some shared cost.”
Ickes’ colleague, Nixon Peabody partner Aldo Ibarra, thinks it may come down to owners wanting better risk assessments for these emerging AI advisors. “I think once you see AI or semi-AI coming into the construction site, at some point a contractor or owner is going to have to say, ‘If I use your services, we need to rethink the licensing for this technology in terms of something more like a subcontractor contract, since you are basically now part of the construction team.’ ”
While the switch from nonbinding AI recommendations to letting an AI or other algorithm make real decisions has yet to happen, Ibarra says companies should be prepared for what that situation could mean for liability going forward. “Once software is not just a tool, and becomes an actual AI making decisions for you, that is different. At that point, these technology vendors will have to rethink how they license and the types of agreements they have with their users.”
Ibarra adds that while he and Ickes follow the issue closely, they have yet to see any litigation over risk and liability related to construction AI advisors. Ibarra notes it may also get sorted out before a test case happens. “The market is going to have a say in this,” he says. “There’s not just one technology firm out there, so if they want to differentiate themselves from the competition, they might put some skin in the game, say they’ll sign on as a subcontractor to assure you that their AI will correctly advance your project.”
No Time Like the Present
But legal concerns about construction AI are still theoretical, and for now DPR’s Maha is adamant that other firms should take the plunge and get their data in shape for machine learning and AI. “Do not wait for your data to get perfect, because it is never going to be,” he says. Even with “60% to 70% of our data that we can trust is accurate, start with the process, and show others the potential of what accurate data could do.” Maha cites the cloud analytics tools in AWS and Azure as a solid place to begin exploring these insights and build a case for further investment in data science.
Once a company makes that investment in getting their data organized and in the cloud, Maha says it’s important to bring in people with relevant experience in data modeling and analysis to really make it shine.
But he also emphasizes that firms shouldn’t fall into the trap of working on data science or machine learning in isolation—it will take the involvement of the rest of the company to shape the algorithmic insights of any AI assistant. “Be sure to involve the business and project execution side as early as possible, since they’re going to be the ones turning the dials,” he recommends. “Involving them early will mean better outcomes for your machine-learning projects.”