How We Achieved 23% Water Loss Reduction with AI-Optimized Distribution

Non-revenue water — the gap between water treated and water billed — averages 40% across Sub-Saharan Africa and 30% in North Africa. In Casablanca, Morocco's largest city with 3.7 million residents, the figure was 34% when we began our pilot with the national utility in June 2025. That means one-third of every liter of treated, pressurized, potable water disappeared before reaching a customer — lost to leaks, theft, metering errors, and aging infrastructure. At a daily production of 600,000 cubic meters, 34% non-revenue water represents 204,000 cubic meters of lost water every day. The energy cost alone — pumping water that never reaches a tap — exceeded $45 million annually. This is not a water problem. It is an infrastructure intelligence problem. And infrastructure intelligence is what Harch Technology was built to solve.

Our approach deployed the full SENSE-THINK-ACT pipeline across Casablanca's distribution network. The SENSE layer ingested real-time pressure, flow, and acoustic data from 12,000 sensors installed across 4,200 kilometers of distribution mains — one sensor every 350 meters on average. The sensor deployment was the largest of its kind in Africa and required custom hardware to handle Casablanca's water chemistry, which includes high mineral content that corrodes standard pressure transducers within months. Our solution: titanium-housed sensors with sapphire pressure windows, manufactured locally at one-third the cost of imported alternatives.

The THINK layer applied two machine learning models in parallel. The first, a gradient-boosted decision tree trained on three years of historical flow data, predicted expected flow at every sensor node given time of day, temperature, and seasonal demand patterns. The second, a graph neural network operating on the network topology, detected anomalous pressure gradients that indicated leaks — including slow leaks below the detection threshold of conventional systems. Together, the models identified 847 leak locations in the first 30 days, of which 792 were confirmed by field inspection — a 93.5% true positive rate that far exceeded the utility's previous detection accuracy of 35% using manual patrols.

The ACT layer translated detections into automated responses. For large leaks — defined as flow anomalies exceeding 50 liters per minute — the system automatically isolated the affected segment by closing motorized valves and rerouted flow through redundant mains, reducing service disruption from an average of 18 hours to under 45 minutes. For smaller leaks, the system generated prioritized repair tickets that included precise location coordinates, estimated leak volume, and recommended repair method. The prioritization algorithm was critical: with 847 leaks and a repair crew capacity of 15 per day, intelligent prioritization was the difference between meaningful reduction and a backlog that grows faster than it shrinks.

After six months, the results exceeded every projection. Non-revenue water dropped from 34% to 26.2% — a 23% relative reduction. Daily water recovery: 46,800 cubic meters. Annualized energy savings from reduced pumping: $12.3 million. Repair cost savings from early detection: $8.7 million — because a leak detected at 10 liters per minute costs $200 to fix, while the same leak detected at 200 liters per minute after main failure costs $15,000. Customer complaints about water pressure dropped 31%. And the system improved continuously: as more data flowed through SENSE, the THINK models refined their predictions, reducing false positives from 6.5% to 2.1% over the pilot period.

The Casablanca pilot demonstrates a principle that extends far beyond water: AI-optimized infrastructure doesn't just improve existing operations — it creates capabilities that didn't exist before. Before this deployment, the utility could not detect a leak until it surfaced or a customer reported it. After deployment, the utility could identify, locate, and prioritize leaks before they caused damage. The difference between reactive and proactive infrastructure management is not incremental. It is transformational. And it is replicable: the same SENSE-THINK-ACT pipeline that reduced water loss in Casablanca is now being deployed for grid optimization, irrigation scheduling, and mineral processing quality control across Harch Corp's verticals. Infrastructure intelligence is not a feature. It is a platform.