What problem does it solve?
Most leaks in a water network are underground and never seen by customers. Thames Water says a huge proportion of its leaks are underground and hard to find. Northumbrian Water says it has traditionally relied on visible signs such as pooling water or pressure changes, and often on customers to report them. Siemens quotes a VA SYD engineer who says that, until recently, the utility could detect only the largest leaks that disrupted service, while the smaller, slower leaks, often from small service pipes, account for the lion's share of non revenue water.
Listening devices find more, but they add a second problem. Yorkshire Water has nearly 40,000 acoustic loggers that listen each morning; when one flags a suspected leak an analyst confirms it and an inspector goes to site, and some alarms are background noise mistaken for a leak. Each false alarm costs a site visit, so the cost of detection rises with the number of sensors unless the software separates leaks from noise.
How does it work?
- Collect signals. Acoustic loggers or sensors on fittings record the sound of water escaping, flow meters and smart meters report flow and pressure by zone, and satellites provide ground moisture, temperature and soil disturbance data.
- Classify the signal. A model learns what normal operation looks like and what a leak sounds or looks like. Yorkshire Water's trial taught an engine to distinguish a leak from background noise in audio recordings. Siemens' SIWA Leak Finder is trained on historical flow and pressure data to separate anomalies from regular operation and assign a cause to each one.
- Locate. Northumbrian Water says FIDO Tech's AI gives technicians accurate areas of interest to investigate, even when no water shows at the surface.
- Confirm. An analyst reviews the alert and an inspector goes to site to pinpoint the leak.
- Repair and feed back. Close each work order with the result, a confirmed leak or a false alarm. The reports on this page do not describe retraining, but those labels are what a team needs to tune or retrain a model and to track its false alarm rate.
- Audience
- Employee facing
- Autonomy
- Assist
- Adoption
- Early adopters
- Channels
- Internal tools, API and system to system
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
No public deployment has disclosed a measurable outcome yet.
Value drivers: Lower cost to serve, Risk and loss reduction, Employee productivity, Customer experience.
Indicative value
A water utility that loses 50 megalitres of treated water a day to leakage
USD 109,500 to USD 1 million
Treated water cost avoided per year
How this is calculated
Formula: dailyLossMl * recoveredShare * daysPerYear * valuePerMl. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Water lost to leakage dailyLossMl, megalitres per day | 50 | 50 | The reference utility. Editorial assumption, replace with your own leakage figure. |
| Additional share of that loss found and repaired recoveredShare, fraction of leakage | 0.02 | 0.08 | Editorial assumption. Replace with the share your own trial or zone data supports. |
| Days in the year daysPerYear, days | 365 | 365 | Constant. |
| Value of one megalitre of treated water valuePerMl, USD per megalitre | 300 | 700 | Editorial assumption for the production and treatment cost of water. Replace with your own marginal cost. |
What it leaves out: Counts the production cost of water not lost. It leaves out the cost of sensors, software and inspection and repair crews, the value of water in a drought, regulatory leakage targets and penalties, and the avoided damage from bursts.
Who already uses it?
4 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Thames Water
United Kingdom · Energy and utilities · 2026
Thames Water, which the release calls the UK's biggest water company, says it now has more than 75,000 acoustic sensors on its network, up from 21,000 in 2025, and that they have nearly doubled leak detection accuracy. It also uses an AI driven service from Origin Orbit that reads satellite data for moisture, temperature and soil changes that point to underground leaks.
In a 19 week trial the company says Origin Orbit helped it identify over 800 leaks, saving an estimated 8.7 million litres of water a day. Thames Water also reports that it found 88% more leaks in July 2026 than in July 2025.
That count is for the company's whole leak finding work, which the release describes as AI, satellite data and acoustic sensors together, in the driest July in England in 190 years, so it shows more leaks found, not what the AI contributed.
No outcome disclosed.
Northumbrian Water
United Kingdom · Energy and utilities · 2025
Northumbrian Water began a six month trial in November 2025 with FIDO Tech, whose acoustic monitors attach to fittings on water pipes and use AI to analyse the sound of water escaping underground. The AI tells engineers which sounds are leaks and which are not, and gives technicians an area of interest to investigate even when no water shows at the surface. The sensors can stay in place or be moved daily to sweep a wider area. The company reports that early results from the trial had already helped it save nearly 6.4 megalitres of water a day.
No outcome disclosed.
Yorkshire Water
United Kingdom · Energy and utilities · 2021
Yorkshire Water has nearly 40,000 acoustic loggers on its clean water network that listen for leaks each morning. When a logger flags a suspected leak an analyst confirms it and an inspector visits the site, and some alarms turn out to be background noise. In 2021 the utility trialled a machine learning engine from Siemens and Artesia Consulting that learns to tell a leak from noise in the audio recordings. The utility reported that false alarms had the potential to fall by up to 60% and that the system could find more true leaks at the same time. The 2021 release does not give a later rollout result.
No outcome disclosed.
VA SYD
Sweden · Energy and utilities · 2021
VA SYD supplies drinking water to more than 546,000 customers around Lund and Malmö in southern Sweden. It tested Siemens' SIWA Leak Finder, an AI based algorithm trained on historical flow and pressure data that identifies and classifies anomalies in the pipe network, in a proof of concept area with 5,000 consumers. Siemens reports that the utility detected simulated leaks as small as 0.5 litres per second, and that non revenue water in its drinking water network could be brought down from 10 percent to 8 percent. VA SYD ran the software in its own data centres, and Siemens says other Swedish municipalities have adopted the same solution.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- A hydraulic model or at least a zone map of the network with pipe materials and ages
- Flow, pressure and meter data by zone, with history to train on
- Acoustic recordings or sensor readings, with outcomes of past investigations
- Work order history that records whether each alert was a leak
Systems to integrate
- SCADA and telemetry for flow and pressure
- Smart meter or advanced metering data
- Geographic information system for the pipe network
- Work order and field service system for inspectors
- Sensor or satellite data provider
Complexity: Medium
Hardware and models come as a service from the sensor or analytics vendor. The utility needs a network model with pipe, zone and meter data, a sensor roll out plan, data links from the SCADA or metering systems, and an inspection team that records the outcome of every alert. VA SYD had to run its software on its own servers, which needed its IT and operational technology teams and Siemens.
- 1
Pick a zone and fix the baseline
Choose a zone with reliable district metering and measure its night flow, leakage and the number of investigations per leak found before any change. VA SYD began with a proof of concept area of about 5,000 consumers.
- 2
Connect the data
Link flow, pressure and meter data and the hydraulic model to the analytics platform, and settle whether it can run in the cloud or has to run on the utility's own servers.
- 3
Test on known leaks
Check the system against past leaks and, where possible, simulated ones. VA SYD simulated leaks and reports detecting some as small as 0.5 litres per second.
- 4
Put analysts on confirmation
Route every alert to an analyst who confirms or rejects it, and record the reason. This builds the labeled data that lets the model improve and shows the false alarm rate.
- 5
Change how crews are dispatched
Agree which alerts inspectors visit first, using the confirmation record, and track repair time from alert to fix.
- 6
Widen coverage
Extend to more zones once the false alarm rate and the cost per leak found are known, and mix sensor types where they find different leaks.
Guardrails
- A person confirms an alert before an inspection is raised, until false alarm rates are known
- Location estimates are shown with a confidence level, never as exact values
- No automatic valve operation or pressure change is triggered by the model without an operator
- Smart meter data tied to an address is handled as personal data under data protection law
KPIs to instrument
- Leaks found per month and per 100 km of mains, against the same period in earlier years
- Share of alerts that turn out to be leaks, and false alarms per leak found
- Time from alert to repair
- Night flow or estimated volume of leakage by zone
- Cost per leak found, including sensors and crew time
Human in the loop
Analysts confirm alerts and decide which to investigate. Inspectors locate the leak on site, and network engineers decide on pressure management and repair priority. Leakage managers review the false alarm rate and the leaks found per inspection every month.
Common failure modes
- More leaks found mistaken for better detection
- Finding more leaks can mean better detection, but also a drier summer or more sensors. Thames Water's 88% rise came in a record dry July and with a larger sensor network, and the release does not split it by method. Report leaks found alongside leakage volume.
- Alert volume outruns crews
- More sensors and more alerts than the repair team can handle leave known leaks open. Size crews and priorities before adding coverage.
- Models that do not travel
- A model tuned on one pipe material or soil type misreads another. Test in each new zone and keep the analyst confirmation step.
- Savings that cannot be traced
- Company reports of litres saved rarely state a baseline. Fix the baseline and measure leakage volume before and after, not only leaks found.
What are the risks and rules?
EU AI Act
Depends on design
Annex III point 2 makes AI systems high risk when they are intended as safety components in the management and operation of the supply of water, and Recital 55 defines such components as systems that directly protect the physical integrity of critical infrastructure.
Recital 55 names systems for monitoring water pressure as examples of safety components, so a model that monitors flow and pressure may fall under Annex III point 2 when its failure could directly risk the network's physical integrity or people's safety, even without automatic valve control. Analysis of acoustic or satellite data used only to plan inspections is the likelier case outside Annex III. Assess each design on its own.
Smart meter data tied to a household is personal data under the GDPR.
Guidance
- Annex III, high risk AI systems referred to in Article 6(2) (European Union, Europe). Point 2 lists safety components in the management and operation of critical infrastructure, including the supply of water.
- Recital 55, safety components of critical infrastructure (European Union, Europe). Defines safety components as systems that directly protect the physical integrity of critical infrastructure and names systems for monitoring water pressure as examples, the test that decides the tier here.
- Directive (EU) 2020/2184 on the quality of water intended for human consumption (European Union, Europe). Article 4(3) requires Member States to assess water leakage levels and the potential for reducing them, which is the reporting context for any leak detection program in the EU.
Controls to put in place
- Inventory entry for each model with an accountable owner and a description of what it may trigger
- Records of every alert, its confirmation or rejection and the outcome of the inspection
- Quarterly review of false alarm rate and missed leaks found by other means
- Access control and security monitoring on the telemetry and sensor networks
Frequently asked questions
- What results do water companies report from AI leak detection?
Thames Water says a 19 week satellite trial with Origin Orbit helped it identify over 800 leaks, saving an estimated 8.7 million litres a day. It also found 88% more leaks in July 2026 than in July 2025, but that count covers its whole leak finding work in a record dry month, not AI alone. Yorkshire Water reported in 2021 that false alarms had the potential to fall by up to 60% in a machine learning trial.
Northumbrian Water says early results from a FIDO Tech trial saved nearly 6.4 megalitres a day. All are company reports. None gives a leakage volume baseline to check the savings against, and the Thames 88% is a count of leaks found for the whole program.
- Does AI replace acoustic loggers and field inspection?
No. The reported systems sit on top of sensors and inspections. Yorkshire Water's engine learns to separate leaks from noise in recordings from loggers it already has, Thames Water uses AI on satellite data alongside more than 75,000 acoustic sensors, Yorkshire Water has an analyst confirm each flagged leak before an inspector visits, and Northumbrian Water's technicians investigate the areas of interest the AI gives them.
- Is AI leak detection high risk under the EU AI Act?
It depends on the design. Recital 55 names systems for monitoring water pressure as examples of safety components of critical infrastructure, so a model that monitors flow and pressure may fall under Annex III point 2 when its failure could directly risk the network's physical integrity or people's safety, even without automatic valve control. Analysis of acoustic or satellite data used only to plan inspections is the likelier case outside Annex III. Assess each design on its own.
How to cite this page
Blits.ai AI Use Case Library, "AI leak detection for water distribution networks", last verified 13 October 2026, https://www.blits.ai/ai-use-cases/water-network-leak-detection. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 13 October 2026: First published