Somewhere under your street, a water pipe is probably leaking right now. Nobody has noticed because the water never reaches the surface — it soaks quietly into the soil, day after day, sometimes for years. AI water leak detection is the technology water companies are now using to find those invisible leaks without digging up a single road, and in 2026 it has moved from pilot projects into routine operations at some of the world’s largest utilities.
This is one of the least glamorous uses of artificial intelligence and one of the most consequential. Cities across the US, Europe, India and Australia are all facing the same arithmetic: aging pipes, tighter water supplies, and no realistic budget to replace everything. Finding leaks faster is the cheapest water a utility can buy.
What is AI water leak detection?
AI water leak detection is the use of machine learning to spot leaks in buried water pipes by analysing data — satellite radar images, acoustic recordings, pressure readings and flow meters — instead of relying on crews walking the network with listening sticks. The AI does not see the leak. It recognises the pattern a leak leaves behind in the data.
That distinction matters. A traditional leak survey is a physical search: technicians walk a section of pipe at night, when traffic noise drops, listening at valves and hydrants. It works, but a large utility may only get around its whole network once every few years. An algorithm reviewing satellite passes and sensor streams can cover the same ground weekly.
How much treated water are cities actually losing?
More than most people would guess. The share of water that a utility treats and pumps but never bills for is called non-revenue water, and it is a headline number in the industry. Analyses citing World Bank estimates routinely put global non-revenue water in the 25–30% range — roughly a quarter of all treated water, lost to leaks, theft and metering errors.
In the United States, the American Society of Civil Engineers has estimated that around six billion gallons of treated drinking water are lost every day, against a backdrop of roughly 240,000 water main breaks a year. For contrast, Germany’s tightly managed networks run at water loss rates closer to 7% — proof that the losses are an engineering and maintenance problem, not a law of physics.
Every litre a utility loses to a leak has already been pumped, filtered, chlorinated and paid for — which is why leak reduction is usually cheaper than building new supply.
How does AI find a leak nobody can see?
AI finds hidden leaks by learning what wet soil, leak noise and abnormal night-time flow look like in data, then ranking sections of pipe by how likely they are to be leaking. Crews are sent to the highest-scoring locations first. Three data sources dominate in 2026, and most serious programmes combine them rather than betting on one.
1. Satellites that see moisture through the ground
Synthetic aperture radar satellites can detect the signature of treated water saturating soil beneath the surface. The raw imagery is far too noisy for a human to interpret — treated water has to be distinguished from rain, groundwater and irrigation — so machine learning models do the separation and output candidate leak zones, typically a radius of a few hundred metres for crews to investigate.
2. Acoustic sensors that learn what a leak sounds like
Permanent acoustic loggers clamped to valves and hydrants record the low-frequency hiss of pressurised water escaping a pipe. Classifier models trained on thousands of labelled recordings learn to tell a genuine leak from a passing lorry, a running tap or a pump. Correlating the same sound across two loggers can then pinpoint the leak to within a metre or two.
3. Flow analytics inside district metered areas
Networks are divided into district metered areas — sealed zones with a meter on the inlet. At 3am, legitimate demand is close to zero, so any persistent minimum night flow is almost certainly a leak. Time-series models learn each zone’s normal behaviour and flag deviations within hours instead of the weeks a manual review would take.
Does AI leak detection actually work in the field?
The most concrete public evidence in 2026 comes from the UK. In April 2026, Thames Water signed a 13-month contract with the North East start-up Origin Tech to deploy its AI satellite system, Origin Orbit, across its distribution network, targeting a reduction of more than 100 million litres of leakage per day.
The trial that preceded it is the interesting part. Over 18 weeks, Origin Orbit identified more than 800 leaks, equivalent to saving 8.7 million litres of water a day. According to Thames Water’s own announcement, 92% of the leaks found were non-visible — they would not have been discovered by anyone driving past. Thames Water has assigned a 20-person specialist team to the programme, with crews expected to find around 25 leaks per week each.
Anglian Water has run a parallel approach using satellite remote sensing from Asterra, aimed particularly at rural pipe runs where acoustic methods struggle over long distances between access points. The pattern is consistent: AI does not replace the repair crew, it decides where the repair crew goes.
Beyond leaks: AI inside the treatment plant
The same utilities are applying machine learning upstream, at the treatment works. The best-established application is chemical dosing. Coagulant and disinfectant doses have traditionally been set by operators reacting to raw water quality; models trained on historical plant data can predict the right dose ahead of a change in turbidity, which reduces both chemical spend and the risk of an out-of-spec result.
Other plant-side uses include predicting pump and membrane failures before they cause an outage, and forecasting demand so that pumping is scheduled into cheaper, lower-carbon electricity windows. It is the same predictive-maintenance playbook now common in other utility sectors — the approach we covered in how AI helps utilities predict transformer failures, applied to water assets.
What AI still cannot do
Three limits are worth being honest about.
- It does not fix anything. Detection is cheap; excavation and repair are not. A utility that finds more leaks than it can repair has simply built a longer backlog.
- False positives cost real money. Every wrong location is a crew, a van and possibly a road closure. Precision matters far more than raw detection counts, which is why utilities run pilots before committing.
- Bad data caps performance. Models need accurate pipe maps, working meters and reliable telemetry. Many networks — particularly older or fast-growing ones — do not have them, and no algorithm compensates for not knowing where the pipes are.
Why this matters in the US, Europe, India and Australia
The economics differ by region but point the same way. In England and Wales, regulators set binding leakage reduction targets, which makes AI water leak detection a compliance tool as much as an efficiency one. In the drought-exposed western US and much of Australia, saved water is supply that does not have to be bought, desalinated or piped in.
In fast-urbanising cities across India and Southeast Asia, where non-revenue water rates in some networks run well above the global average, the constraint is usually metering and mapping rather than algorithms — which means the first phase of any AI programme is boring groundwork, not modelling. That sequencing is easy to underestimate and is where most projects stall.
The broader lesson mirrors what we have seen elsewhere in infrastructure, from AI leak detection in oil and gas pipelines to grid monitoring: the win is not a smarter machine, it is a better-ordered work queue. Water utilities have always known their networks leak. AI is finally telling them which hundred metres to dig up first.
Frequently Asked Questions
How accurate is AI water leak detection?
Accuracy varies by method and network condition, so utilities judge it by how often a flagged location turns out to be a real leak. Thames Water’s satellite trial found more than 800 leaks in 18 weeks, 92% of them invisible from the surface, which is why the utility moved to a full contract.
Can satellites really detect a water leak underground?
Yes, indirectly. Radar satellites detect soil moisture patterns, and machine learning models are trained to separate the signature of treated drinking water from rain and groundwater. The output is a candidate zone of a few hundred metres, not an exact pinpoint — crews still confirm the spot on the ground.
What is non-revenue water?
Non-revenue water is treated water that a utility produces but never bills for, lost to leaks, theft or metering errors. Estimates citing World Bank data place the global figure at roughly 25–30% of supply, making it one of the largest sources of waste in urban infrastructure.
Does AI replace leak detection technicians?
No. AI changes where technicians are sent rather than removing the job. Someone still has to confirm the leak acoustically, excavate and repair it, and utilities deploying these systems have generally kept or expanded their field crews.
Is this technology only for large water companies?
Not necessarily, but scale helps. Satellite and analytics services are usually sold per kilometre of network, so smaller utilities can buy them — the harder prerequisite is having accurate pipe records and working meters, which is where many small operators fall short.
