Two-Thirds of Maintenance Teams Are Adopting AI. Most Will Aim It at the Wrong Assets.

More than two-thirds of maintenance teams plan to adopt AI-powered predictive maintenance by the end of 2026. The vendor pitches are working. The procurement cycles are closing. And for a specific slice of the asset portfolio — instrumented rotating equipment, well-monitored compressors, sensor-rich production lines — the returns are real. Machine learning models are now forecasting discrete equipment failures six to twelve months in advance with accuracy rates above 85%.

But here is the problem most AI vendors are not advertising: in the infrastructure sectors that stand to gain the most from predictive maintenance, the majority of assets are not the ones the AI can see.

The 70% Problem

In a typical water utility, more than 70% of the asset base by count — and often by replacement value — consists of buried pipelines: cast iron mains, ductile iron laterals, aging prestressed concrete pressure pipes. These assets have no vibration sensors. Most have no inline condition monitoring at all. Their degradation is slow, largely invisible, and periodically catastrophic.

The same blind spot applies in gas distribution (where much of the network was laid in the 1950s and 1960s), in power transmission (roughly 70% of US transmission lines and transformers are more than 25 years old), and in urban rail (where track degradation is spatial and continuous, not discrete and countable).

These are linear assets. They don't behave like equipment. They behave like networks — connected, topologically dependent, with failure modes that propagate from segment to segment rather than from component to parent.

Most predictive maintenance AI was not built for them. It was built for the pump station at the end of the pipeline, not the pipeline itself.

What Gets Measured Gets Optimised — Into the Wrong Problem

The practical consequence is that organisations investing in AI-driven maintenance are optimising the visible part of their infrastructure while the invisible part degrades on its own schedule.

A water utility that deploys a best-in-class predictive maintenance platform for its treatment works — 200 assets, all instrumented, all feeding telemetry — may have genuinely excellent visibility into those facilities. Mean time to repair drops. Unplanned downtime shrinks. The AI dashboard shows green.

Meanwhile, the distribution network serving those works has 4,000 pipe segments with condition scores that haven't been updated in six years. When a segment in that network fails, the AI model had no warning to give. The cascade — pressure loss, isolation valve sequencing, treatment plant feed interruption — unfolds outside the system's line of sight entirely.

The AI wasn't wrong. It was watching the wrong thing.

Two Asset Types, Two Mental Models

The reason this happens is structural. Linear asset management requires a fundamentally different data model than discrete asset management.

Discrete assets — treatment plants, pump stations, depot facilities, manufacturing equipment — are hierarchical. A pump belongs to a pump skid, which belongs to a process train, which belongs to a facility. Failure modes follow parent-child relationships. ISO 14224 codifies how to classify them. OEE metrics aggregate their performance. Most EAM platforms were designed around this mental model, and most AI vendors built their training pipelines to match.

Linear assets — pipelines, transmission lines, rail track, cable networks — are spatial and topological. A 300mm main is not a hierarchy; it is a sequence of segments with geospatial coordinates, material properties, installation dates, soil conditions, and connections to other mains. Its risk is not calculated per-asset but per-network-state: which segments are at pressure, which are approaching end-of-life, and what happens downstream if any of them fail?

Running a graph neural network over a linear network to predict cascade propagation — research now showing meaningful results as of early 2026 — requires the asset data to exist as a graph in the first place. If the pipeline network lives in a GIS system with no connection to the plant asset hierarchy, the graph is permanently incomplete. The AI cannot reason across a boundary that isn't in the data model.

The ROI Trap

The global predictive maintenance market is growing at a 26.5% compound annual rate, projected to reach $70 billion by 2032. That is an enormous amount of money being spent on AI systems whose effectiveness is, for many infrastructure operators, ceiling-capped by incomplete asset visibility.

The Asset Integrity Management market — the segment that directly addresses the condition of aging linear networks — is on a separate growth curve, from $25.8 billion in 2024 toward $50 billion by 2035. Both markets are growing fast because both problems are real. But they are being addressed with separate tools, separate data models, and separate AI layers.

The operator in the middle — a water authority, a transit agency, an energy company with both generation assets and a transmission network — ends up with two AI systems that cannot talk to each other, applied to the two halves of an interconnected infrastructure. The discrete asset management AI can tell you the plant is healthy. The linear network analysis (if it exists at all) can tell you the pipes are at risk. Neither can tell you what happens at the interface.

That interface is where most unplanned outages originate.

What Unified Visibility Actually Enables

The answer is not another integration project to connect two separate platforms via a nightly ETL. That approach produces a reconciled dataset, not a shared reasoning layer — and AI cannot do useful cross-domain inference on data that arrives 24 hours stale with mismatched asset IDs.

The answer is an asset data model that natively represents both topologies: linear network segments as graph nodes with spatial attributes and network edges; discrete equipment as hierarchical objects with parent-child relationships and telemetry streams. When both live in the same model, the AI agent layer can reason across the seam. A pressure anomaly in a pipeline segment can trigger a review of the pump station at its downstream node. A scheduled shutdown at a treatment plant can trigger an inspection recommendation for the mains feeding it while the isolation window is open.

For more on how network-aware risk scoring works across a linear asset portfolio — and why it differs fundamentally from hierarchy-based RBI approaches — see our knowledge guide on linear asset management.

Gartner projects that 40% of enterprise applications will feature task-specific AI agents by end of 2026. In EAM, those agents will only be as useful as the asset graph they can reason over. An agent that knows your equipment but not your network will optimise one half of your infrastructure while the other half ages in the dark.

---

Most maintenance AI is aimed at the assets that are easiest to instrument. The biggest risk sits in the assets that aren't. Curious how UniEAM handles both? Request a demo.