Rethinking the AI-Water Nexus

When Denver Water brought online the Northwest Water Treatment Plant in 2024, that meant staff had to learn a whole new set of operating instructions. The plant, located north of Golden, Colorado, incorporated new technology and adjusted approaches from other water treatment plants. Already, the water utility had secured a licensing agreement with Microsoft that allows for deploying its artificial intelligence interface, Copilot, throughout the organization, but with data secured internally. Copilot can assist with basic tasks like writing emails or technical documents. But in early December 2025, Jonathan Spitze, director of business technology and project management with Denver Water, sat down with staff at that new treatment plant to introduce them to the testing phase for a pilot program AI tool to help them run the new facility. The operating instructions were spread over several lengthy documents, but a Copilot agent, an AI-powered system that can learn, problem-solve, and perform tasks, and is generally built by reviewing massive amounts of text,
had been trained on those procedural manuals. Now, staff can ask it simple questions and get answers with links to the actual documentation.

“They were super excited,” Spitze says. “We see a lot of value from making these more targeted [Copilot] agents.”

“While we’re still early in testing, we’re encouraged by how the tool is supporting operators by streamlining access to information and enabling staff to make quicker, informed decisions,” Stephanie Riley, engineering senior specialist in water quality and treatment with Denver Water, said via email.

Spitze and others in the water sector see this kind of assistance as a key way in which AI could help.

Recent AI advancements have seen the technology leap into modern life with surprising capacities and, at times, seemingly limitless possibilities for solving problems. Water managers, too, have been looking at the technology — and some AI start-ups are targeting the water world, tailoring their solutions to problems like pipe maintenance and leak-detection. The possibilities and promises include more efficient water management, conservation, improved ability to forecast floods and droughts, and boosted productivity in the workforce. But increasing reliance on AI raises questions too, about if and how this technology opens water utilities up to cybersecurity concerns, and how to balance the water savings AI might offer with water consumption at data centers that power it.

State and federal governments have both asserted a desire to harness and lead in this new technology. The Trump administration outlined a goal of seeing the United States as a global leader in artificial
intelligence. The State of Colorado has likewise expressed a dedication to leveraging artificial intelligence and ensuring an ethical, equitable and responsible deployment of AI. All signs point toward a widening
presence of the technology, which seems to further insert itself with every software upgrade.

“The blend of digital technology and AI is the big opportunity to boost efficiency and protect water resources,” says Ralph Erik Exton, executive director of the Water Environment Federation, a nonprofit
organization that provides technical and educational support for water quality professionals.

The Water Environment Federation jointly launched the Water-AI Nexus Center in September to convene utilities, academia, and finance and technology professionals, effectively pairing water and
AI experts to develop sustainable water practices for AI infrastructure while using AI to solve global water challenges. The project arose from conversations at a White House roundtable on water
security and infrastructure in late 2024, Exton says, in which artificial intelligence repeatedly resurfaced in the conversation but with more questions than answers.

The water industry is, admittedly, risk averse: It’s in the business of securing health and safety for communities and economies, and disinclined to make untested changes. The Water Research Foundation has developed a project to share demonstrated success stories to build confidence through quantifiable and real case studies. Trepidation around security and privacy can make it even more difficult to adopt software innovations.

“When you’re talking about human health, human safety, and effects on people’s lives, a machine shouldn’t be making those decisions for us. That doesn’t mean that we don’t use AI as a tool — help me look at different scenarios, and help me analyze the problems I’ve been facing,” says Spitze. “But at the end of the day, it still needs to be a person that says, ‘yes, this is the right thing to do,’ because we want humans taking care of humans. A machine is not going to understand the responsibility that comes with that.”

While generative AI promises much, a 2025 report from MIT found that most initiatives stall: 95% of GenAI pilot programs do not meaningfully impact their companies’ profit or loss. The researchers for that report suggested flawed integration: In part, generative AI often seeks to work on the sales and marketing side, but the biggest gains might lie in back-office automation.

Work is just beginning to outline best practices and standards for some of the thornier issues ringing AI. Those would include building ethical and responsible AI systems that keep humans involved in decisions and make clear what role AI plays, and ensuring equitable access for smaller and resource-limited utilities, Exton says. The group is also working on drafting and distributing best practices and standards for how to balance the quantities of water data centers consume with the water savings AI might offer.

“It’s the two sides of the equation that need to be equally tackled — it’s how can you use AI to benefit what we do around the water footprint and the environment, but also how do we make sure that all of this exploding use of AI isn’t doing further damage to the environment?” Exton says. “Those two things need to be equally as strong and carry equal weight.”

Already, the Water Environment Federation has suggested using air cooling instead of water cooling at data centers in water-stressed regions, which some data centers in Arizona, India, Mexico and South
Africa have already done. However, studies suggest that since aircooled systems require significantly more energy than water-cooled systems, they could end up with a larger overall water footprint. That’s because traditional power generation from gas and coal has a very high water footprint. Where sustainable sources or replenishment isn’t possible, data centers could also be designed to use water for cooling without evaporation. Using recycled municipal water is also an option, as is on-site rainwater harvesting. The goal, Exton says, is to find ways to move from “linear systems that take, use and dispose of water to circular approaches that focus on reducing, reusing, and regenerating resources.” The organization is working on a standard for how to balance water savings and use by AI in part because, he says, now, with different companies using different standards, it’s difficult to assess or compare.

“I certainly wake up in the mornings thinking, we’ve got to move even faster on our mission to help the utilities,” Exton says.

Cutting Losses

When Breckenridge Water Division staff were alerted to a possible leak through their newly installed Kamstrup Flow IQ meters with an acoustic leak detection system in November, water from the ruptured water line was just starting to seep through the surface. The meters use ultrasonic sensors that both measure flow and receive sound waves, effectively listening to the water pipes for noises that indicate a crack. The technician who responded found a skim of ice on the dirt road outside town where the leak detection system had signaled a brewing problem.

“We were able to catch it and get it taken care of the next day instead of being at the mercy of when it really gets worse and surfaces,” says Laura Lynch, water division manager for Breckenridge.

The tourist town faces the challenge of being staffed as though it’s a small town, but having to support 40,000 people during peak visitation. The limits mean contractors were reviewing only about a third of the water system every year, and the utility was living with persistent and undetected leaks and making emergency repairs as they popped up. The new system uses machine learning algorithms to analyze noise patterns from service lines and water mains that often indicate leaks. The new technology could save as much as 580 acre-feet of water per year, Lynch says.

“Without AI, operators would need to manually review thousands of graphs to spot these leaks,” Blake Michal, manager of solution management for Kamstrup, explains. “Machine learning automates this process, making it faster, more accurate, and scalable for utilities with tens of thousands of meters.”

The water sector already includes automation through the likes of smart meters, which automatically track and relay information rather than require in-person meter readings, as well as widespread sensor use and connected sensors embedded in various physical objects to provide real-time data through the “internet of things.” AI could thrive in such a data-rich environment. It is particularly well-suited to analyzing vast amounts of data and spotting patterns and anomalies. It can also combine data from multiple sources, like smart meters and weather forecasts, to improve demand forecasts.

For Breckenridge, staff are still learning to filter out signals from pressure-reducing valves and booster pumps. The system could hand over more autonomy to customers, many of whom keep second homes in town and would call for essentially plumbing or property management services from the water division when their bill indicated a leak. Eventually, a customer portal will allow them to set up personal leak alerts.

Breckenridge spent the winter gathering data, tracking suspected leaks, and doing as much testing as possible, Lynch says. “Then come springtime, [we] have a game plan with what we need to actually dig up and repair.”

Aiding Employees

Denver Water began using machine learning about 20 years ago to identify water mains due for replacement by assessing many factors — age, location, soil composition, corrosion, repair history, and exposure to freezing temperatures — and predicting repairs before main breaks occurred. The utility can only replace about 1% of the thousands of miles of pipe that run all over Denver each year, so it’s crucial to know exactly where to target that work. AI accounts for these many factors and identifies high problem areas.

Denver Water is also in early stages of creating an AI agent to interface with developers and builders who submit their plans to the utility for approval ahead of construction. The lengthy list of requirements often means applications need repeat revisions, and depending on how busy the building season has been, it may take weeks for someone with Denver Water to review an application. This AI system would provide instant feedback on plans and how to match them up to the city’s rules. Even if the AI agent isn’t perfect but gets the drawings most of the way there before a person reviews them, says Jonathan Spitze with Denver Water, that could save significant time in the process for both staff and builders.

Optimizing Operations

Metro Water Recovery, which provides wastewater services for several Front Range municipalities, was already, effectively, drowning in data from tens of thousands of data points coming from field equipment and controls and automated systems. Some level of automation was managing and synthesizing the data for operators, says Joshua Goldman, senior wastewater process engineer, but it was a struggle to utilize and react to that much information.

“We think AI can help there,” he says. “Overall, it’s just going to help us with optimization and better decision-making — let’s find correlations in complex systems that we were unable to pull out before.”

Particularly machine learning models, which can process those many, many parameters and data points and find signals humans might miss, could better inform human decision-makers. Although, for now, those models are better at forecasting arising issues than explaining why the problem occurred or how to fix it in the future.

Metro Water Recovery is about halfway through development with a third-party vendor of an AI agent that would fine-tune “solids dewatering,” the process in which the bacteria that consume carbon, nitrogen, and phosphorus are then separated from wastewater as solids, to save costs, energy and water. Other machine learning projects could optimize disinfection and phosphorous removal.

For areas that can be a challenge to reach because the space is physically small or filled with enough dangerous chemicals or gases that sensors won’t survive, AI’s forecasting abilities could fill in gaps with a rough estimate for what’s going on based on the rest of the data and some non-traditional sources including audio, vibration, emails and other documents. That approach is called “soft sensing.”

Somewhat similarly, AI-powered models can run what’s called a “digital twin,” a digital model that replicates a real-world product or system, and can be used to glean insights on monitoring and maintenance,
predict problems, and optimize operations. A digital twin of a water treatment plant, for example, would use machine learning models and artificial intelligence to find areas for improvement and head off operations disruptions. Models can also recommend optimal chemical doses, saving costs and improving regulatory compliance, and they can predict equipment failures, leading to more effective maintenance strategies and reduced downtime.

Already, the City of Aurora has purchased a digital twin for managing water and wastewater infrastructure and planning for how to make limited infrastructure meet the demands of a growing city. In Boulder,
city managers turned to an analytics software platform, City Simulator, to model the likelihood of a similar storm to the one that created devastating floods in September 2013 and plan for upgrading infrastructure.

“We understand that there’s a lot of opportunities with these technologies and we’re trying to figure out where they plug in and where they make the most sense,” Goldman says. “We don’t anticipate them taking over our jobs or operating our plants for us, but we see them being able to interpret data and help us make better decisions in a more timely manner.”

Saving Farmers’ Water And Wisdom

Think of a spray irrigation system on a farmer’s fields like the zones in a backyard sprinkler system: Each zone can be set to water for more or less time, based on whether grass there is getting enough water, but that takes time and attention. Farmers irrigating with overhead sprinklers can similarly adjust those systems, but deciding how much to change requires analyzing variables including crop type and growth stage, soil profile, and recent and forecast weather. Plus, instead of five or six zones, a field can have hundreds of zones and localized data from hundreds of sensors. Processing that much data is easy for machines, so Perry Cabot, a professor with the Colorado Water Center at Colorado State University, has tasked AI with analyzing those many data points to calculate precise prescriptions for watering fields.

The research project, which is funded by the U.S. Department of Agriculture, is in the second of three years. When deployed, even farmers who have already increased efficiencies in their fields could save about 10% in each watering.

“That last 10% is important,” Cabot says.

He’s using drone and satellite imagery to verify field conditions, checking for crop stress signaled through yellow leaves, and hoping to relay information in a way that will help farmers who wouldn’t have a drone of their own. AI tools might also be able to show that more watering won’t make a significant difference for a crop, including in fields still flood irrigating.

“It’s a tool that allows us to test things that we think we know, and then actually help figure out what we don’t know, such that we could maybe apply it to more simplified irrigation systems,” he says.

As with water utilities, he worries about the loss of generational knowledge about farming as young people drift away from the work of running the family farm. AI — if it works — could help new farmers fine-tune processes without seasons of trial and error, or answer questions when previous generations of farmers aren’t still around to do so. But AI will have to prove itself.

“It can’t just say, ‘Go put this amount of water on,’ and then it’s wrong,” Cabot says. In that case, farmers would give up on using AI. “We have to take advantage of this current time frame where the producers can help us say, yes, that’s a legit decision.”

What he’d love to see is an AI-versus-farmer showdown, perhaps through the Testing Ag Performance Solutions program, which runs projects around the country in which farmers compete to demonstrate best yield or best profits. “I think that’ll be fascinating to figure out how well those two entities can be played off of each other.”

Fast-Forwarding Reviews

When Brian Zavareh was working with the IT department for Denver’s wastewater utility in 2019, he caught snippets of the hours of video staff watch of the inside of sewer pipes, scanning for maintenance and repair needs. In the previous five years, utility workers had reviewed about 1,000 miles of the 1,500 miles of pipes through the city and attempted to accurately and consistently apply one of 249 codes to different anomalies and features within them.

“It is working, but it is time-consuming and error-prone,” Zavareh said.

As a PhD student with the University of Denver’s College of Engineering, Design and Computing, he returned to the utility with an idea: What if he could train AI to review that video for defects and deterioration? And what if AI could be trained to more consistently apply those codes, flagging possible points of failure for repairs?

“It’s not replacing a human, but an AI-enabled environment helps them to do whatever they’re doing faster and more accurately and more consistently,” Zavareh said.

The utility gave him 10 years of videos to train the model and through it, his program demonstrated an ability to review the videos faster than a person. He expects the program to increase efficiency enough to survey an additional 200 to 300 miles of pipe per year, heading off potentially costly ruptures, reducing costs, and lengthening the lifetime for some equipment. AI would help people work through that footage faster and more consistently, he says, supporting staff without supplanting jobs with AI.

He has since co-founded a company, InfraSmart Solutions, through CU Denver’s Smart Futures Lab, a business incubator and accelerator, with one of his professors and secured grant funding for training the AI for defect detection. He’s seen other companies do the same elsewhere, but where InfraSmart seeks to extend the solution is in forecasting deterioration and prioritizing work orders, adding information to the model. including pipe shape, material, and age, all of which can affect the likelihood and timeline of repairs. The result would be a ChatGPT-like program that can essentially converse about maintenance needs. Building that part of the system could take another year and a half, assuming they can get all the data they need to train the model.

Filling In Environmental Gaps

Measuring conditions around the many streams and rivers threaded across the country is expensive, often requiring localized sensors, and so data is rare, says Li Li, a professor with PennState Institute of Energy and
the Environment. Long-term data can be particularly hard to find.

“Usually each site has a small amount of data, has it in different times, is measuring different times at different locations for different decades, things like that,” Li said, “That is a big limitation in water quality fields in terms of how much data has been available and has been a bottleneck for developing a consistent understanding of the river quality.”

She used a deep learning model to fill in the gaps in fragmented and discontinuous data for 1,471 U.S. river sites between 1980 and 2022. She and her team trained a deep learning model to reconstruct continuous daily water temperatures by feeding it data including precipitation, temperature, solar radiation, land use, and soil properties. They compared predicted data with measured data, where available, and re-trained the model until it began to consistently match real data. The results showed a startling increase in riverine heat waves, or instances of abnormally high water temperatures in rivers. The model also reported those heat waves occurred about half as often as air heat waves, and with about a third the intensity, but lasted nearly twice as long. Rivers in the Rocky Mountain area showed some of the sharpest increases, likely linked
to shrinking snowpack and earlier warming in spring. Smaller snowpack results in lower streamflows, with smaller streams that heat up more quickly.

A second machine learning model assessed about 30 different characteristics to determine which factors, such as winter and summer temperature increases and land use characteristics, were most influential in creating a riverine heat wave. The strongest influence landed on higher air temperatures, particularly at night.

Better knowledge about riverine heat waves, she said, can help people managing rivers understand what’s happening in those ecosystems, and plan for how to mitigate effects.

“For me, it’s important to increase awareness,” Li says. While water quantity concerns are often linked to climate change, like floods and droughts, she says, “These studies point out that water quality tends to be affected a lot by climate warming.”

On the whole, water users and researchers are finding success deploying AI. How to balance the water used to power that AI with the savings that technology might provide remains murky, raising its own practical and ethical questions to consider. Spitze, with Denver Water, points to the need to hew to an even wider philosophy aired in the book “Genesis: Artificial Intelligence, Hope, and the Human Spirit.”

“The value that AI will offer to our society is enormous, as long as AI continues to see that its purpose needs to be helping the human race,” he says. “It’s a little on the scary side, but the reality is, I think that’s why humanity can’t relinquish all control.”

Independent journalist Elizabeth Miller writes about environmental issues in the American West for publications including The Washington Post, Scientific American, Outside, Backpacker, and The Drake.