Last spring, a landscaping crew in Menlo Park, California, drove a post-hole digger eighteen inches into what they believed was clean dirt, and it turned out to be a private gas lateral. Nobody was hurt, but the repair bill hit $4,200 and the project stalled for nine days while PG&E dispatched a crew, excavated the damaged section, and certified the repair. The homeowner had called 811, and the 811 locator had marked the main in the street, but the lateral running from the main to the house was never marked, because 811 does not cover private utilities.
That was not a freak accident but the system working exactly as designed, performing precisely the job it was built to do and nothing more.
What 811 Actually Does (and Does Not Do)
Call 811 before you dig, as you have heard a thousand times. It is free, it works, and within two business days, locators come out and spray-paint the approximate locations of public utility mains in the right-of-way near your project: yellow for gas, red for electric, blue for water, orange for telecom. You get a rough idea of where the big pipes and cables run under the street or along the easement, and that is where 811's responsibility ends.
Private laterals, the lines that actually connect those mains to your house, fall outside the scope entirely. Your private sewer line running from the house to the city connection goes unmarked, along with your irrigation system, your septic field, and the abandoned oil tank from 1962 that nobody remembered existed. Add the previous owner's buried dog run cable, the old sprinkler line that was cut and capped but never removed, and the drainage tile that runs diagonally across the yard because the original grading contractor had his own ideas about where water should go, and you begin to understand how much infrastructure sits below a typical residential lot with zero documentation.
According to PG&E's own 2024 data, 1,247 incidents in Northern and Central California involved homeowners or contractors damaging underground gas or electric lines while digging. In 89 percent of homeowner incidents, nobody had called 811 at all. But even when people call, the data guiding the locators can be catastrophically wrong.
Broken Maps, Not Broken Technology
A 2026 open letter from the National Utility Locating Contractors Association should make every builder uncomfortable. It describes locating professionals placing marks "exactly where records indicated, only for excavation to reveal infrastructure in a completely different location. Not a few feet off. On the wrong side of the road. Crossing at the wrong block. Or not shown in the records at all."
What looks like a technology failure is actually a records problem wearing a technology mask. Utility maps in many American cities reflect decades of layered installations by different agencies, different contractors, and different record-keeping systems, some of which were paper-based, some of which were digitized from paper with unknown accuracy, and some of which simply do not exist because the line was installed before anyone thought to document it.
CGA's 2024 DIRT data quantifies the damage: of those 196,977 reported incidents, 25 percent traced back to locating and mapping failures, while another 24 percent went to excavation practices. Purdue University researchers working with the 2022 DIRT data found that 87.84 percent of all incidents resulted from inaccurate location information, which means that even when contractors follow every rule in the book, the underlying data feeding those rules is often wrong.
Total annual cost? CGA estimated it at $30 billion in 2019, a figure that has not been updated and that nobody has challenged, which tells you something about how comfortable the industry has become with treating underground surprises as an acceptable cost of doing business.
Ground-Penetrating Radar Works. It Just Costs Too Much.
For anyone who builds things for a living, the interesting part starts with the technology that could replace spray paint and prayer.
Ground-penetrating radar is not new, but understanding what it does matters for following where AI enters the picture. A GPR unit sends electromagnetic pulses into the ground and records the reflections, and because different materials produce different reflection signatures, a trained analyst reading the output can identify metal pipes, plastic pipes, concrete structures, abandoned tanks, voids, and even the approximate depth and diameter of each object, all from a screen that looks roughly like a very messy ultrasound.
GPRS, the largest private utility locating company in the United States, reports a 99.8 percent accuracy rate across more than 350,000 projects since 2017. They achieve this by deploying both GPR and electromagnetic equipment on every locate and staffing every job with a SIM-certified project manager. Their data gets stored in a platform called SiteMap that creates permanent digital records.
For residential work, though, the cost creates a barrier that keeps the technology confined to commercial projects. A private utility locate runs $300 to $1,000 per site, which barely registers on a $3 million commercial project but represents 2 to 7 percent of a $15,000 backyard fence installation, and most homeowners will not pay it. Most contractors will not suggest it, either, because the liability math does not force the conversation until someone hits something, and by then the $300 they skipped has become a $3,500 repair plus a nine-day delay.
| Project Type | Typical Cost | Private Locate Cost | Locate as % of Project | Avg. Strike Repair Cost |
|---|---|---|---|---|
| Commercial building | $3,000,000 | $800 | 0.03% | $50,000+ |
| New home foundation | $40,000 | $500 | 1.25% | $3,500 |
| Pool installation | $60,000 | $500 | 0.83% | $3,500 |
| Fence/deck/patio | $15,000 | $400 | 2.67% | $3,500 |
| Landscaping/tree planting | $5,000 | $300 | 6.00% | $3,500 |
PG&E puts the average repair at $3,500 per utility strike, which means the insurance math actually works: a $500 scan that prevents even one strike across every seven or eight projects pays for itself and then some. But contractors do not think in portfolios across dozens of projects, they think in individual jobs where the odds of hitting something feel low enough to gamble on, and so the scan never gets ordered and the gamble occasionally fails in spectacular fashion.
AI Can Read a Radar Scan Faster Than a Human. Nobody Sells It for Houses.
Reading a GPR scan is genuinely hard, and the difficulty is exactly where AI enters the picture. Raw output, called a radargram, presents as a grayscale image of hyperbolic reflection patterns that takes years of field experience to interpret reliably: a metal water main produces a clean, bright parabola, while a PVC sewer lateral produces a faint smudge, and tree roots generate noise that can obscure both. When you consider that the difference between paying $800 for an experienced analyst per site visit and paying pennies per scan through an AI model comes down entirely to whether sufficient training data exists, the economics of the problem start to look very different.
Research is further along than most people in residential construction realize.
At Purdue University, researchers published a Bayesian framework combining variational inference and Markov Chain Monte Carlo methods to estimate the location, orientation, and radius of underground pipes from GPR data. Their best configuration achieved 97.5 percent coverage on simulated data and 77.5 percent on real field data. The gap between simulated and field performance is significant, but the trajectory is clear: more training data closes it.
Training data is arriving faster than anyone in the industry seems to appreciate. A 2025 dataset published in Data in Brief released 2,239 labeled radargram images from urban GPR surveys under a Creative Commons license, specifically designed for training deep learning models to detect pipes, cables, and underground voids, while a separate ASCE study used YOLOv5, a real-time object detection algorithm originally developed for autonomous vehicles, to classify underground objects from field-collected GPR images into five categories: metal pipe, plastic pipe, boulders, air voids, and water voids.
At the University of Vermont, researchers built what they call "cognitive GPR," a system that combines ground-penetrating radar with a smartphone 3D scanning app and augmented reality software. Walk across a yard with this system, and it builds a real-time 3D map of what is underground, rendered in video-game-quality imagery instead of grayscale hyperbolas. "The system knows where you are, knows what's underneath, and can show you detailed images of what's there," said Dr. Dryver Huston, the project lead. Edge computing processes the data on site instead of sending it to a remote server.
None of these systems are available for residential use, not as a product you can buy, not as a service you can hire, and not even as a pilot program you can volunteer for.
What a Residential AI Locate Would Actually Need
Forget the lab for a minute. Think about what a residential contractor needs versus what the research delivers.
A contractor working a typical residential lot of 5,000 to 10,000 square feet needs a scan completed in under an hour with hardware that fits in a truck and does not require a PhD to operate, producing a map that says: here is your sewer lateral at 36 inches deep, here is your gas lateral at 18 inches, here is an abandoned irrigation line at 12 inches, and here is something metallic at 42 inches that we cannot identify, so do not dig there until someone investigates further.
Hardware remains the biggest obstacle, though that may be the wrong way to frame it. Professional GPR units from manufacturers like GSSI, Sensors & Software, and Screening Eagle run $15,000 to $50,000, a price that makes perfect sense for a private locating company billing $500 per visit across ten jobs a day but makes no sense whatsoever for a fence contractor who digs three times a week and has never heard of a radargram.
AI rewrites that equation in a way that matters for residential. A trained GPR analyst earns $70,000 to $110,000 a year. An AI model reading radargrams at 77.5 percent field accuracy, which is where Purdue's research currently sits, cannot replace that analyst, and no responsible engineer would suggest otherwise. But an AI model at 95 percent accuracy, a level that the trajectory of deep learning on imaging problems strongly suggests is reachable within three to five years given sufficient training data, could turn GPR interpretation from a skilled profession into a software subscription that runs on the same tablet a contractor already uses for project management.
Drop interpretation cost and the hardware follows. Cheaper, simpler GPR units already exist for concrete scanning, including the Proceq GP8000 at roughly $7,000, and a residential-focused unit optimized for shallow-depth utility detection in the 6- to 60-inch range where most residential work falls, paired with an AI interpretation layer and a smartphone app that shows results in augmented reality, could realistically hit a price point of $200 to $300 per residential scan within five to seven years.
Why Nobody May Build This
Consider the possibility that nobody builds it because the unit economics are genuinely broken at residential scale.
Commercial infrastructure projects pay $5,000 to $50,000 for subsurface utility engineering surveys, while a $300 residential scan would be a commodity margin product requiring the same field visit, the same equipment deployment, and the same liability exposure, all to serve a customer who is perfectly willing to simply gamble instead. A company that scales this for residential is building a low-margin, high-volume service business competing against human optimism bias, which is arguably the most powerful force in homeownership.
Purdue's 77.5 percent field accuracy is encouraging as a research milestone, but it sits far below the 99.8 percent that GPRS achieves with human analysts deploying combined GPR and electromagnetic technology. At 77.5 percent, roughly one in four underground objects gets missed, and on a residential lot with a gas lateral, a water lateral, a sewer lateral, and an old irrigation line, missing one of those four is not an acceptable outcome by any engineering standard.
Here is the rebuttal worth sitting with: imaging AI has consistently followed a curve from 80 percent accuracy to 99 percent within three to five years of sufficient training data becoming available, across medical imaging, autonomous vehicles, satellite imagery, and industrial quality inspection. What limits the GPR application is not the algorithm but the labeled training data, and the 2025 open dataset of 2,239 radargrams represents a beginning, not a destination. Reaching the reliability threshold that would make residential deployment responsible requires something on the order of 50,000 to 100,000 labeled residential scans. Nobody is collecting them yet, which is both the problem and, for whoever starts, the competitive moat.
What You Should Do Right Now
If you are a residential contractor who digs:
Call 811 every time, including for fence posts, because PG&E's data shows that 89 percent of homeowner utility strikes happen when nobody calls, and the free service at minimum catches the public mains, which alone is worth the two-day wait.
For any project involving excavation below 24 inches on a property built before 1980, order a private locate. A $300 to $500 investment is insurance against hitting an undocumented lateral, an abandoned tank, or a utility that was installed before GPS and digitized records existed, and you should specifically ask the locating company whether they use GPR in addition to electromagnetic locating, because EM-only locates miss plastic and clay pipes entirely.
If you are building a new home, require the excavation contractor to obtain a private utility locate before breaking ground and add it as a line item in the contract. It amounts to less than 0.1 percent of total construction cost and eliminates a category of delay that can push your timeline by two weeks while a utility company repairs what your crew hit.
Watch the AI-GPR space closely. Purdue's Bayesian framework is academic today but commercializable within a few years, and the University of Vermont's augmented reality system is closer to field readiness than anything else in the pipeline. When someone packages AI-interpreted GPR into a service that costs $200 per residential lot and delivers results on a smartphone in real time, the economics of preventive scanning will flip overnight, because the question was never whether the technology works but which company figures out the residential business model first.
Limitations
CGA DIRT data is self-reported by member organizations, likely undercounting total incidents, and the 196,977 figure for 2024 represents unique reported damages while actual damages including unreported incidents are almost certainly higher. No published study has specifically tested AI-GPR accuracy on residential laterals at shallow depths versus trunk mains in public rights-of-way, where most training data originates. Cost projections of $200 to $300 per scan within five to seven years are estimates based on hardware price trends and AI interpretation cost reductions in analogous imaging fields, not on any specific product development roadmap, and GPRS's 99.8 percent accuracy rate is self-reported and applies to projects using combined GPR and EM technology with certified analysts, not to GPR alone or to AI-assisted interpretation.