AI for Finding Buried Utilities: A Well-Measured Problem and an Unproven Fix

How CodoraTech is funded: CodoraTech is supported by advertising and, in some articles, by affiliate links. Where an article contains affiliate links, we say so at the top of that article.

The Common Ground Alliance’s 2024 DIRT Report, published on 28 August 2025, analysed 196,977 unique damage reports and concluded that conditions had worsened rather than improved. Separate CGA modelling puts the annual cost of underground utility damage in the United States at 83.2 billion dollars. Those numbers are solid, attributable and large enough to justify serious investment in better locating. What could not be found for this article is a published independent evaluation showing that machine learning reduces strike rates. What follows is the size of the problem, the physics that constrains it, and one real company working on it.

Two Common Ground Alliance figures that are not interchangeable

The DIRT Report is the authoritative record of damage to buried infrastructure in North America, and its 2024 edition analysed 196,977 unique damage reports. The CGA Index, the association’s composite measure of damage-prevention conditions, moved from 94.0 in 2023 to 96.7 in 2024, in the wrong direction. CGA President and Chief Executive Sarah K. Magruder Lyle framed the finding by saying that incremental change is not enough.

A separate CGA product models the national picture and arrives at roughly 668,999 damage incidents a year. The two figures measure different things and should never be added or substituted. The DIRT number counts voluntary submissions to a database; the national number is a statistical estimate of what actually happens, including everything nobody reported. One further caveat applies to everything in this section: CGA’s own web properties returned errors to automated retrieval when this article was researched, so these figures are taken from trade press quoting CGA and should be checked against the association’s published report before being relied on commercially.

For every dollar of repair, nearly sixteen of consequence

The cost model is the part of CGA’s work that changes how the problem should be framed. It puts total annual cost at 83.2 billion dollars, which the association notes is nearly three times the widely cited 30 billion dollar benchmark from 2019. The breakdown matters more than the headline, because direct repair is the smallest component.

• Community and environmental costs: 32.6 billion dollars
• Business and economic disruption: 31.6 billion dollars
• Operational response: 7.8 billion dollars
• Human impacts including injuries: 6.2 billion dollars
• Direct utility repairs: 4.9 billion dollars

The quarter of damages that no software can reach

The DIRT root-cause table is where a technology argument either survives or does not. In the 2024 data, failure to notify the 811 one-call service accounted for 24.54 percent of damages, an excavator failing to maintain clearance after verifying marks for 16.07 percent, a facility not being marked because of locator error for 11.94 percent, a facility marked inaccurately because of locator error for 8.58 percent, and improper excavation practice for 6.75 percent. The top ten causes together account for 85 percent of reported damages.

Read those figures as an addressable market and the picture sharpens. Just over 20 percent of damages trace to locator error, whether a line was missed entirely or marked in the wrong place, and that is the specific failure a better sensing or interpretation system could plausibly reduce. Roughly a quarter trace to an excavator simply not making the call, which no imaging technology and no model changes at all. That single split should discipline any claim that artificial intelligence is about to transform damage prevention: the largest single cause is a human process failure upstream of any sensor.

Clay soil, clutter and 51 documented challenges

The second constraint is physical. A peer-reviewed review of ground-penetrating radar for underground utility detection, published in the Indian Geotechnical Journal in 2026, concludes that GPR faces persistent challenges affecting data quality, interpretability and overall survey reliability. Two are named directly: signal attenuation in clay-rich soils, and overlapping utility clutter in congested urban ground, which makes individual targets difficult to separate. The authors synthesise 51 specific challenges into nine categories.

The implication for the machine-learning case is uncomfortable but important. A model can improve the interpretation of a returned signal, and interpretation is genuinely one of the named weaknesses. It cannot recover a signal that clay absorbed before it reached the target. Where the binding constraint is soil physics or a record that was never accurate in the first place, the limiting factor is not a shortage of training data.

Exodigo is real; the independent test is not

One named company is doing serious work here. Exodigo uses non-intrusive subsurface imaging, fusing multiple sensor types with artificial intelligence, to map buried pipes and cables. National Grid Partners announced a multi-million-dollar strategic investment in October 2022 without disclosing the amount, and National Grid deployed the technology at sites in New York, including a test location at Yaphank where it found utility lines absent from existing as-built records. The chief executive is Jeremy Suard, and the company said it had worked with more than twenty organisations across energy, utilities and transport in the United States, Europe and Israel since commercial launch in June 2022.

Exodigo also claims that it identifies 20 to 50 percent more utilities than alternatives, attributed in its own press release to CGA analysis from 2022. That is a supplier’s claim in a supplier’s announcement, and it is presented here as such rather than as a measured finding. No independent, published, quantified pilot showing artificial intelligence reducing utility strike rates was found for this article, and no state 811 centre appears to have published measured results from an AI ticket-screening deployment, although several vendors sell risk-scoring products for one-call tickets.

That is the honest state of the field in September 2026: a very well-quantified problem, a physics limit that no amount of computation removes, a real company with real deployments and an unaudited performance claim, and no published evidence yet that any of it has moved the damage numbers. Anyone selling a different version of that story should be asked which independent evaluation they are relying on.

Sources: Utility Contractor · Heavy Equipment Guide · Indian Geotechnical Journal · Exodigo