The 3-Million-Kilometre Problem: Why AI Can't Fix Aging Pipelines Alone

Approximately 3 million kilometres of pipelines worldwide have now surpassed their expected service life. Leakage accounts for roughly 70% of global non-revenue water losses. And as of 2026, regulators across the US, EU, and Australia have shifted from input-based oversight—did you inspect it?—to outcomes-based accountability: can you prove measurable progress on reducing water loss and improving system efficiency?

These are not abstract statistics for a VP of Operations at a water authority or a utility asset manager. They are the numbers that turn up in a regulator's letter, in a board presentation, in a post-incident debrief. The pressure to move from reactive to proactive asset management is real, and it is arriving at exactly the moment when most organisations' tooling is least equipped to handle it.

Why This Problem Is Harder Than It Looks

The instinct, when faced with aging infrastructure, is to reach for predictive maintenance. Deploy sensors. Ingest the data. Train a model. Forecast failure.

That reasoning is sound as far as it goes—and recent research confirms it: tree-based ensemble models now achieve failure prediction accuracy above 88% for water distribution pipelines, with Random Survival Forest variants excelling specifically at time-to-failure forecasting. That is genuinely useful.

But predicting that a pipe segment will fail in 90 days is only the beginning of the problem. The harder question is: what happens when it does?

Pipelines don't fail in isolation. A pressure drop in one segment propagates through adjacent segments. A failure at a junction affects every pressure zone and pumping station connected downstream. In interdependent urban infrastructure networks, cascading failures—where one failure triggers secondary and tertiary failures in interconnected systems—are the mechanism behind the most costly and disruptive events. Recent modelling work has demonstrated a 31% improvement in cascade volume forecasting using network-topology-aware models over standard asset-level approaches.

An AI model that predicts individual pipe failure without understanding the network topology around it is answering a narrower question than the one you actually need answered. The distinction matters enormously for prioritisation: a pipeline segment in a low-criticality branch of a distribution network might warrant scheduled maintenance. The same degradation profile in a segment feeding a hospital, a pumping station, and 40,000 homes warrants an emergency response today.

Managing linear assets well requires a fundamentally different data model than managing a discrete asset like a compressor or a motor. Linear assets are spatial and relational. Their behaviour is governed by topology—by what they connect to, what flows through them, and what depends on them. The right reasoning layer for a pipeline network is a graph, not a hierarchy.

The Hidden Problem in the Machine Room

The same utility running a water distribution network almost certainly also operates pumping stations, treatment plants, and maintenance depots. These are discrete assets—facilities with equipment hierarchies, parent-child failure modes, and performance metrics measured in OEE and ISO 14224 taxonomies. Discrete asset management requires its own reasoning model: a hierarchy-aware view of how component-level failures propagate up through systems to affect plant-level availability.

Most mixed-portfolio operators have ended up with one tool for each world. A GIS-adjacent platform for the linear network. A traditional CMMS for the facilities. These systems were never designed to share a data model, let alone reason about how a pipeline failure cascades into a treatment plant's operating constraints, or how a pump failure changes the pressure profile of a distribution segment twelve kilometres away.

That gap—between the network view and the hierarchy view—is precisely where the most consequential operational decisions live. It is also where most AI deployments silently fail: the model has access to sensor data from assets managed in one system but not the other, and the dependency graph it needs to reason about exists in neither.

What Outcomes-Based Regulation Actually Demands

The shift to outcomes-based oversight makes this worse, not better. Regulators don't want to know how many inspections you completed. They want to know what happened to non-revenue water loss this quarter, and why, and what you're doing about it.

Answering that question requires a unified asset model—one that connects the pipeline network's hydraulic behaviour to the discrete assets that serve it, under a single audit trail. Every prediction, every maintenance decision, every risk threshold needs to be traceable and defensible. "Our AI recommended it" is not a satisfying explanation in a regulatory debrief. A decision log—timestamped, with the data inputs and model rationale attached—is.

This is the architecture that agentic AI in EAM should produce: not a chatbot layered onto a database, but a reasoning engine that can work across both the network view and the hierarchy view, surface the dependencies that matter, recommend actions with explicit justifications, and produce audit-ready records as a byproduct of normal operation.

The Case for a Unified Platform

Running two systems doesn't just create a reconciliation burden. It creates a reasoning gap. The utility that manages its pipeline network in one platform and its treatment facilities in another cannot ask—and get a reliable answer to—the question: "If this segment fails this week, which facilities are exposed, how quickly, and what does that cost us?"

That question is answerable. The data to answer it exists. The barrier is architectural: it requires a platform that holds both the network model and the facility hierarchy, and can reason across them. For a mixed-portfolio operator—a utility with both distribution infrastructure and treatment assets, a transit authority with both rail corridors and maintenance depots—this is not a theoretical advantage. It is the difference between proactive risk management and reactive crisis response.

The 3-million-kilometre problem will not be solved by more inspections or faster dashboards. It requires a fundamentally different approach to how asset intelligence is structured—one that treats the network and the hierarchy as two views of the same operating reality, not two separate systems bolted together at the spreadsheet layer.

Curious how UniEAM handles both? Request a demo.

---

Further reading: Understanding linear asset management — the data models, reasoning patterns, and operational requirements that set network-shaped infrastructure apart.