Automation
Architecture review before implementation
Implementation scope and rollout planning
Clear next-step recommendation
A blueprint for a custom AI workflow that transforms raw pressure and flow data into prioritized leak repair orders, directly targeting the 20-30% of treated water lost before it generates revenue.
Non-revenue water (NRW) represents a massive operational drain, often exceeding 20% of treated volume. Manual leak detection is slow, reactive, and misses small losses that aggregate into major financial waste. A custom automation workflow directly addresses this by continuously ingesting SCADA, IoT sensor, and district metering area (DMA) data, applying real-time anomaly detection to pinpoint pressure drops and flow discrepancies indicative of leaks, often before they surface.
The architecture's value is in its closed-loop control. Detected anomalies trigger a synchronization with a calibrated EPANET hydraulic model to isolate the leak segment. The system then generates a work order in the CMMS (like IBM Maximo or SAP PM) with confidence scores and estimated loss volumes, enabling crews to prioritize by financial impact. Implementation requires robust data validation, model calibration agents, and approval gates to prevent false positives from overwhelming field teams.
A custom AI workflow transforms raw pressure and flow data into prioritized maintenance actions, directly reducing non-revenue water and protecting utility margins.
This workflow automates the detection of leaks that represent pure financial loss. By identifying and quantifying losses in near real-time, utilities can recover 5-15% of non-revenue water (NRW), directly improving the bottom line. The system integrates with hydraulic models like EPANET to estimate loss volumes, translating physical anomalies into concrete financial impact for capital planning.
Automated detection shifts maintenance from reactive, costly emergency repairs to planned, efficient interventions. The workflow analyzes sensor data to pinpoint leak locations within a 50-meter radius, reducing crew dispatch time, excavation area, and street restoration costs. It also prevents catastrophic failures that lead to service interruptions, regulatory fines, and expensive damage control.
Continuous pressure anomaly detection identifies weak points and transient events that stress the network, enabling proactive reinforcement before pipe bursts occur. This extends asset life, defers capital replacement costs, and mitigates the risk of major service outages and associated reputational damage. The workflow creates a defensible, data-driven record for infrastructure investment justification.
The system eliminates weeks of manual data sifting by engineers. AI agents continuously analyze SCADA/AMI feeds, run anomaly detection algorithms, and auto-generate preliminary work orders with location, estimated severity, and recommended crew. This frees senior staff for complex analysis and planning, turning data scientists into force multipliers rather than data processors.
By modeling water loss across the entire network and forecasting deterioration rates, the workflow provides a precise, dynamic picture of infrastructure needs. This moves capital planning from a calendar-based or reactive model to a condition-based and predictive one, ensuring limited budgets are allocated to the highest-impact pipe segments and pressure zones.
The architecture is designed for phased rollout, starting in a high-loss District Metered Area (DMA). A pilot can be deployed in 6-8 weeks, connecting to existing SCADA and GIS, with clear ROI measured within a single quarter. This low-risk, high-clarity approach proves value before scaling network-wide, building internal buy-in and operational familiarity.
This blueprint details a production-grade, multi-agent AI workflow that automates leak detection and water loss modeling by orchestrating sensor data, hydraulic models, and maintenance systems to reduce non-revenue water and operational costs.
This workflow directly targets the financial and operational drain of non-revenue water (NRW). It automates the continuous analysis of SCADA pressure and flow data, replacing manual, periodic reviews with real-time anomaly detection. By identifying potential leaks early—often before they surface—the system prevents escalating water loss, reduces emergency repair costs, and protects infrastructure integrity. The architecture is built for integration with existing hydraulic modeling platforms like EPANET or InfoWater, ensuring credibility with engineering teams.
Implementation requires a robust orchestration layer, typically using frameworks like LangGraph or Temporal, to manage the stateful workflow between specialized agents. The Anomaly Detection Agent employs statistical and ML models on time-series data, while the Hydraulic Model Agent executes calibrated network simulations to pinpoint leaks. Critical controls include confidence scoring gates to route low-confidence alerts for manual review by engineers, ensuring operational trust. The final output is a prioritized work order in systems like SAP, Maximo, or custom CMMS, complete with location, estimated loss rate, and recommended crew dispatch, closing the loop from detection to action.
A production-grade leak detection and water loss modeling workflow integrates sensor telemetry, hydraulic simulation, and maintenance orchestration to reduce non-revenue water and operational risk.
This agent continuously ingests pressure and flow data from SCADA, IoT sensors, and AMI (Advanced Metering Infrastructure) systems. It applies statistical and ML-based anomaly detection (e.g., isolation forests, pressure transient analysis) to flag potential leaks in near real-time, filtering out normal operational noise. It validates data quality and routes raw alerts to the Hydraulic Model Orchestrator for confirmation.
Upon receiving an alert, this agent automatically retrieves the relevant network segment from the EPANET or InfoWater model. It executes a calibrated simulation, comparing observed vs. modeled pressures/flows to confirm the leak's existence and approximate location (zone-level). It handles model versioning, parameter updates from recent calibrations, and exception routing for data gaps or model failures.
This component estimates the volume and financial value of water loss using the FAVAD (Fixed and Variable Area Discharges) principle or other hydraulic relationships. It integrates tariff data and assigns a priority score based on loss volume, critical asset risk, and historical repair data. High-priority, high-confidence leaks are pushed automatically to the work order queue, while lower-confidence cases are flagged for field verification.
This agent interfaces with the CMMS (e.g., SAP, Maximo, or custom field apps) to generate structured work orders. It includes the suspected location, priority, recommended equipment, and links to the hydraulic model results. It assigns crews based on location, skill, and availability, and can trigger notifications via SMS or mobile app. It manages the feedback loop, updating the model and anomaly detection logic once field results (e.g., actual leak size, location) are reported.
A critical control plane that logs all agent decisions, model inputs/outputs, and field validations for auditability and regulatory compliance. It manages approval gates for high-cost interventions and orchestrates the periodic retraining of anomaly detection models using confirmed leak data. This layer ensures the system's decisions remain explainable, defensible, and improve over time, integrating with data lakes and governance platforms.
Built on an event-driven architecture (e.g., using LangGraph for agent orchestration), the workflow integrates key systems: SCADA/Historian for telemetry, GIS for network geometry, Hydraulic Modeling Software (EPANET API), CMMS/ERP for work execution, and Data Lake for analytics. Deployment choices include cloud-native for scalability or hybrid for latency-sensitive control, with robust API management and fallback procedures for offline sensor scenarios.
This blueprint details a phased implementation for a custom AI workflow that automates leak detection and water loss modeling, integrating SCADA, hydraulic models, and maintenance systems to reduce non-revenue water and operational costs.
Phase 1 establishes the foundational data pipeline, ingesting real-time pressure and flow telemetry from SCADA and IoT sensors into a time-series database. Concurrently, we deploy initial anomaly detection models, trained on historical data to flag deviations from expected hydraulic behavior. This phase delivers immediate value by providing a centralized, clean data lake and early leak alerts via dashboard, while de-risking the project by validating data quality and integration points with systems like OSIsoft PI or Siemens WinCC before complex modeling begins.
Phase 2 integrates the calibrated hydraulic model (e.g., EPANET) to contextualize anomalies, distinguishing leaks from normal operational changes and estimating loss volumes. The workflow orchestrator, built on LangGraph or Temporal, manages this logic, routing high-confidence detections to Phase 3. Here, the system automatically generates prioritized work orders in a CMMS like IBM Maximo or ServiceNow and pushes them to field crew mobile devices. Governance is enforced through confidence scoring, human review gates for major mains, and full audit trails linking sensor data to maintenance actions, ensuring operational trust and regulatory compliance.
Comparison of operational and financial KPIs for water loss management before and after implementing a custom AI workflow for leak detection and modeling.
| Metric | Current Manual Process | Custom AI Workflow |
|---|---|---|
Non-Revenue Water (NRW) Detection Cycle Time | 30-45 days (monthly billing analysis) | < 24 hours (continuous anomaly detection) |
Average Time to Confirm and Locate a Leak | 5-7 days (customer calls, field surveys) | 2-4 hours (automated pressure/flow correlation & EPANET simulation) |
Percentage of Leaks Identified Proactively | 15-20% (reactive to surface reports) | 70-85% (predictive from pressure transients and model deviations) |
Annual Cost of Water Loss (for a 100k-connection utility) | $2.5M - $4M (15-25% NRW) | $1M - $1.6M (target 8-12% NRW) |
Field Crew Efficiency (Leak Repair Work Orders) | 3-4 investigations per crew per day | 8-10 targeted, high-confidence repairs per crew per day |
Data-to-Decision Labor (Analysis & Reporting) | 2 FTE engineers for model calibration and report generation | 0.5 FTE for workflow oversight, exception review, and model retraining |
Audit Trail for Regulatory Compliance & Capital Planning | Fragmented spreadsheets, SCADA logs, and paper reports | Unified digital ledger linking sensor data, model runs, alerts, and work order closures |
Capital Avoidance (Deferred pipe replacement via targeted rehab) | Low; replacement driven by age and failure rate | High; prioritization based on modeled criticality and actual loss severity |
A production-grade workflow that integrates SCADA telemetry, hydraulic models, and maintenance systems to autonomously detect leaks, quantify losses, and trigger repairs, directly reducing non-revenue water and operational cost.
This workflow automates the continuous analysis of pressure and flow data from water distribution networks to identify anomalies indicative of leaks. It replaces manual data review and reactive response, converting SCADA streams into actionable intelligence. The operational upside comes from reducing the volume of non-revenue water, lowering repair costs through early detection, and optimizing maintenance crew deployment. Implementation requires robust data governance to handle sensor drift and ensure model inputs are reliable.
Architecturally, the system ingests live telemetry into a time-series database, where an anomaly detection agent, built on libraries like Prophet or custom LSTMs, flags deviations. These triggers parameterize and execute an EPANET hydraulic model via an API to isolate the leak segment and estimate loss volume. High-confidence events auto-generate work orders in systems like Maximo or ServiceNow, while low-confidence cases route to an engineer dashboard for review. Controls include confidence scoring, approval gates for major interventions, and a feedback loop where repair data recalibrates the detection models, creating a continuous calibration cycle against model drift.
Buyers evaluating automated leak detection and water loss modeling need clarity on implementation risks, data requirements, and operational control. These answers address the practical concerns of deploying a production-grade AI workflow in a critical utility network.
A real implementation requires a validation and imputation layer before analysis. Ingested pressure and flow data passes through agents that flag anomalies, gaps, or instrument drift. For critical gaps, the system can impute values using hydraulic models (like EPANET) or neighboring sensor correlations, but will log all imputations for audit. Low-confidence data triggers an alert for field verification, preventing the workflow from generating false-positive leak alerts based on bad inputs. This gatekeeping is essential for maintaining operator trust in the system's outputs.
Building a production-grade leak detection workflow requires coordinated delivery across technical and operational teams, followed by ongoing management by utility engineers and field crews.
Key Role: Define the technical requirements and validate model outputs. They provide domain expertise on network topology, pressure zones, and EPANET model calibration. During delivery, they work with data scientists to translate hydraulic principles into detection logic. In operation, they review anomaly alerts and approve model retraining triggers based on seasonal changes or network updates.
Key Role: Build and maintain the data pipeline. They integrate SCADA/telemetry feeds, historical maintenance records, and weather data into a time-series data lake. Implementation involves setting up streaming ingestion (e.g., Apache Kafka, TimescaleDB) and ensuring data quality checks. Post-launch, they monitor pipeline health, manage schema changes, and support the scaling of data volumes as more sensors come online.
Key Role: Develop and deploy the core detection agents. They build the anomaly detection models (e.g., isolation forests, LSTMs) and the orchestration logic (using LangGraph or similar) that sequences data validation, model inference, and confidence scoring. Delivery includes creating the API layer for the work order system and implementing human-in-the-loop review queues. In operation, they manage model performance drift and deploy updated detection logic.
Key Role: Primary consumers of the workflow's output. During the pilot, they provide feedback on alert precision and work order clarity to reduce false dispatches. In steady-state operation, they receive prioritized work orders via mobile CMMS (e.g., Maximo, Fiix) integration, dispatch crews, and close the loop by reporting actual leak findings and repair details, which feeds back into the model for continuous learning.
Key Role: Govern system access and data governance. They enforce security protocols for the integration between OT (SCADA) and IT (cloud/analytics) systems, ensuring no impact on critical control functions. They also establish audit trails for all leak alerts, model decisions, and work order generations to meet utility regulatory and reporting requirements for water loss accounting.
Key Role: Own the business case and rollout sequencing. They define KPIs for non-revenue water (NRW) reduction and ROI. During delivery, they manage the phased rollout—starting with a pilot pressure zone—and coordinate change management for field crews. Post-implementation, they track operational savings and commission expansions to additional zones or integration with advanced metering infrastructure (AMI).
A controlled rollout strategy is critical for deploying automated leak detection workflows into live water networks, balancing operational risk with the urgency of reducing non-revenue water.
The governance model begins with a sandboxed pilot on a non-critical network segment, using historical SCADA data to validate detection accuracy and false-positive rates against known incidents. This phase establishes baseline performance metrics and refines the anomaly detection thresholds within the EPANET hydraulic model integration. Operational controls, including mandatory human review of all high-confidence leak alerts before work order generation, are enforced to prevent unnecessary crew dispatches and build operator trust in the AI's recommendations.
Following a successful pilot, the rollout expands to secondary zones, integrating live pressure and flow data from IoT sensors and SCADA systems like OSIsoft PI. The workflow architecture incorporates approval gates within a platform like LangGraph, where high-probability leaks auto-generate draft work orders in a CMMS like Maximo, but require supervisor sign-off. Full network deployment is sequenced by pressure zone, with continuous monitoring of model drift and exception routing for complex scenarios, ensuring the system delivers sustained reductions in water loss without compromising network safety or maintenance efficiency.
Enabling Efficiency, Speed & Accuracy
We build AI systems for teams that need search across company data, workflow automation across tools, or AI features inside products and internal software.
Comparison of manual vs. custom AI workflow for detecting leaks and modeling water loss in distribution networks, showing operational and economic tradeoffs.
| Metric | Manual / Legacy Process | Custom AI Workflow |
|---|---|---|
Mean Time to Leak Detection | 14-21 days (from anomaly to work order) | 2-4 hours (continuous real-time analysis) |
Non-Revenue Water (NRW) Reduction Potential | Baseline: 20-30% of system input volume | Target: 8-15% via predictive intervention |
Hydraulic Model Calibration Cycle | Quarterly or bi-annual manual updates | Continuous, automated calibration with SCADA/sensor data |
False Positive Alert Rate | N/A (reliance on customer calls) | Controlled to <5% via multi-signal anomaly fusion |
Work Order Generation & Routing | Manual dispatch based on crew availability | Automated prioritization and dispatch via CMMS integration |
Audit Trail for Regulatory Compliance | Fragmented logs in spreadsheets & emails | End-to-end digital trail from sensor to closure |
Annual Operational Cost (per 1000 km of network) | $1.2M - $1.8M (labor, repairs, unbilled water) | $650K - $900K (system, cloud, focused labor) |
Human Oversight & Exception Routing | 100% of alerts require engineer review | 18% routed for review; 82% auto-validated & dispatched |
About the author
CEO & MD, Inference Systems
Prasad Kumkar is the CEO & MD of Inference Systems and writes about AI systems architecture, LLM infrastructure, model serving, evaluation, and production deployment. Over 5+ years, he has worked across computer vision models, L5 autonomous vehicle systems, and LLM research, with a focus on taking complex AI ideas into real-world engineering systems.
His work and writing cover AI systems, large language models, AI agents, multimodal systems, autonomous systems, inference optimization, RAG, evaluation, and production AI engineering.
How We Work
One-fit-all AI don't work for modern businesses. At Inferensys, we aim to understand your business & custom requirements; which we use to define most efficient agentic workflows, the data, and the tools for your business.
The first call is a practical review of your use case and the right next step.
