Water Utilities Can Predict Pipe Failures. Regulators Are About to Ask How.

The predictive maintenance results for water utilities in 2026 have been difficult to argue with. AI-driven systems flagging failing pipeline segments before a burst occurs — sometimes 30 to 90 days in advance — have driven a 34 percent reduction in unplanned downtime at utilities that have deployed them. The case for AI in water infrastructure is no longer theoretical.

What is becoming urgent, and uncomfortably concrete, is the regulatory side of that equation.

Arizona's Corporation Commission opened the first formal state-level inquiry into utility AI in March 2026 — not to restrict it, but to define what accountability looks like when an AI system makes a decision that affects service reliability. Similar inquiries are expected in other jurisdictions before year's end. Meanwhile, water utilities face mounting pressure around non-revenue water, with regulators pushing toward mandatory audits, public performance scorecards, and funding tied to demonstrated asset intelligence.

The question being asked is no longer "do you use AI for asset management?" It is: "can you prove your AI makes defensible decisions, and can you trace every maintenance recommendation to its data source?"

For utilities running two separate EAM systems — one for the pipeline network, one for treatment plants and pump stations — that question is genuinely difficult to answer.

Why the two-system architecture exists (and why it's breaking)

Linear water assets — distribution mains, transmission pipes, service lines — are managed spatially. Their behavior is governed by network topology: pressure propagates, failures cascade, and segment risk is inseparable from where in the network a segment sits. Effective linear asset management requires GIS-native data models, break history analysis, soil corrosion data, and the ability to reason about upstream and downstream consequences when a segment degrades.

Discrete water assets — treatment plants, pump stations, chlorination systems, reservoirs — are managed hierarchically. A treatment plant contains process systems; each system contains equipment; each unit has a maintenance history, an OEE metric, and a failure mode that propagates up through the hierarchy. Effective discrete asset management requires ISO 14224 classification, component-level telemetry, and parent-child failure reasoning.

These are genuinely different problems. Most water utilities solved them with different systems: a GIS-adjacent network management tool on one side, a traditional CMMS or EAM for facilities on the other. The approach worked tolerably when inspection was manual and maintenance was reactive.

It works poorly now. The AI that predicts a high-risk segment in the distribution network does not know the condition of the pump station sitting at that network node — because that data lives in a different system with a different asset ID schema and no shared maintenance history. The recommendation that reaches the asset manager is half-informed.

The compliance seam

When an auditor asks for a decision trace — why was this inspection prioritised, what data drove the recommendation, what was the outcome — a two-system architecture produces a two-document answer. The network system has one piece. The facility system has another. Reconciling them into a defensible narrative requires someone who knows where everything is and has time to assemble it.

Regulators writing accountability frameworks for utility AI are not designing around that manual reconciliation step. They are assuming the answer is already in the system.

CISA's operational technology asset inventory guidance requires a unified, auditable register of assets across the organisation. Expanded NERC CIP controls extend mandatory requirements down to remote facilities and substations that were previously lightly governed. These frameworks do not carve out exceptions for utilities that manage network and facility assets in separate platforms. They assume a coherent, queryable asset record. For utilities without one, the compliance gap is structural, not merely operational.

What happens at the seam

The operational cost of the boundary between systems becomes visible in specific failure modes.

When a burst water main occurs in a pressure zone served by a single pump station, the immediate repair response is a network problem: isolate the segment, identify alternative routing, assess customer impact. Within hours it becomes a facility problem too — the pump station is compensating for the pressure loss and running at conditions outside its normal operating envelope. How long can it sustain that? What is its current maintenance state? Are any components approaching end-of-life under these conditions?

If the linear asset team and the facility team are drawing from different systems, these questions get answered by phone calls and tribal knowledge. That is a slow and error-prone approach when the degradation clock is running.

An agentic platform that holds both asset models natively can surface this in seconds: network impact traced from the break point, pump station state pulled from the same data graph, a combined recommendation with a confidence score and a full decision log attached. The incident response does not stop at the system boundary. Neither does the audit trail.

The accountability question water utilities should be asking now

Gartner predicts that 40 percent of enterprise applications will feature task-specific AI agents by end of 2026. For water utilities, the more pressing number is the evidence standard regulators will require before accepting AI-generated maintenance recommendations as defensible.

That standard has not been formally set in most jurisdictions. Arizona's inquiry will be watched closely. But the direction is clear: AI decisions affecting service reliability and public safety will require traceable decision logs, not just accuracy metrics. A 97 percent failure-prediction accuracy rate is not a compliance argument on its own. A reproducible, timestamped decision chain is.

Utilities that have deployed AI but cannot produce that chain — because the prediction came from a network model that never saw the facility data — are exposed in a way that will not remain invisible once mandatory audits begin.

Water utilities have earned the right to be proud of what predictive AI has achieved. The next phase is demonstrating that those predictions rest on auditable, cross-asset intelligence. That is where the architecture of the EAM platform becomes the regulatory differentiator.

For a deeper look at how network-aware reasoning applies to linear water infrastructure, see our linear asset management knowledge guide.

---

Curious how UniEAM handles both? Request a demo.