Why Prescriptive Maintenance Programs Stall After Pilot

 
Maintenance determines how much of an asset-heavy operation is available to run. In manufacturing, energy, and rail, the same constraint holds: revenue depends on equipment being in service, and every hour of unplanned downtime is capacity that cannot be recovered later.

Most of the past decade of improvement has gone into seeing failure earlier, and the tools are now widely deployed. Detection is no longer the constraint.

Deciding is. A predictive model produces a probability. Someone still has to weigh it against maintenance capacity, parts availability, and service commitments, commit the work, and afterward judge whether the call was right. That decision is where maintenance economics are settled, and it is the part of the process that remains largely unsupported.

Prescriptive maintenance is the engineering answer to that gap. It determines the specific intervention, fits it to real operating constraints, and validates the outcome against what the model expected. Whether it works depends on the quality and traceability of everything feeding it.

Prescription Raises the Bar on Data Quality

Prescription raises the stakes accordingly. A false positive from a predictive model costs an inspection. A prescriptive system commits a maintenance slot, parts, a crew, and an asset taken out of service. Every recommendation has to be traceable to source signal, timestamped, and reproducible under review.

Fidelity, rather than inference speed, is what constrains these systems in practice. The question is whether the system’s account of an asset corresponds closely enough to its actual condition to justify committing resources against it. It also determines whether technicians engage seriously enough for their overrides to mean anything.

The Brownfield Reality

Asset-heavy estates are heterogeneous by construction. A single operation may run control, safety, and embedded computing platforms from different manufacturers across procurement cycles spanning decades, with fixed infrastructure accumulated over 30 years or more. In recent research by GlobalLogic and HFS, 49% of industrial companies with $1B+ revenue surveyed identified difficulty integrating new technologies with legacy systems as their greatest barrier to deploying advanced digital technologies.

Three engineering problems have to be resolved before a prescriptive layer can operate reliably on top of that estate:

IT/OT convergence. Control systems, safety instrumented systems, and SCADA networks were engineered for deterministic reliability, and connecting them to analytical infrastructure without compromising it is a safety-critical design problem. A traction control network on a rail fleet and a distributed control system in a chemical plant present the same constraint.

Multi-vendor normalization. Sensor outputs across equipment generations use different sampling rates, units, and failure taxonomies, and reconciling them cannot flatten the distinctions that carry diagnostic meaning.

Edge-to-cloud partitioning. Where latency carries safety consequence, there are hard limits on how much processing can move upstream, and those limits are set by the asset, not architecture.

Operators who skip this find model performance bounded by input quality, and the boundary stays invisible until recommendations start being wrong in expensive ways.

The Loop That Makes Prescription Work

Platforms built for this pattern already exist in the market, including within Hitachi. What determines whether they perform is the engineering beneath them. A prescriptive system earns its accuracy through feedback across four layers.

Perception. Onboard sensing samples the physical state of the asset continuously, which requires calibrated, time-synchronized data across estates never designed to provide it.

Inference. Edge platforms classify anomaly severity and estimate remaining useful life close to the asset, where the physics of the process sets the response window.

Prescription. Outputs become work that fits real constraints: maintenance capacity, parts inventory, crew availability, service impact.

Validation and learning. The system compares what it prescribed against what was found, and divergence becomes training signal.

Where the Knowledge Gets Captured

The validation layer is where human judgment becomes a performance mechanism rather than a compliance step. A technician reviewing a recommendation can confirm it, override it, or act on it and find something the model did not anticipate, and each of those is a labeled example. Confirmation reinforces a pattern, an override marks a boundary the model had wrong, and an unexpected finding identifies a failure mode absent from the training data entirely.

Capturing this requires deliberate design. The system has to prompt for the reason behind an override, not just the fact of it, present recommendations with enough context that disagreement is informed, and record annotations in a structure that later training can use. Systems that treat technician input as an approval step capture almost nothing.

Systems that treat it as the primary source of expert signal accumulate something more durable: an institutional record of diagnostic reasoning that compounds with every intervention, and a data asset in its own right, with the same requirements for ownership, quality standards, and retention as any other.

Governed co-pilots then return that record to the point of work, putting accumulated diagnostic history in front of a technician alongside live asset data and applicable standards. Humans on-the-loop are what keep the system honest, and the architecture has to be built so their corrections land somewhere that changes the next recommendation.

This is knowledge engineering as much as machine learning, and it usually determines whether maintenance teams trust the output enough to act on it at all.

How These Systems Fail Quietly

Prescriptive systems rarely fail in ways that announce themselves. Two patterns are worth naming, and both are failures of capture.

Trust collapses in either direction. When technicians over-trust the system, they stop exercising the judgment that generates the training signal, and confirmations become reflexive. When they under-trust it, they route around it, resolve issues informally, and the overrides that carry the most diagnostic value never get recorded. Either way the loop stops learning. Confidence presentation, override friction, and how recommendations are surfaced at the point of work are engineering decisions with direct consequences for whether the system improves.

The return path becomes a checkbox. A closed loop depends on accurate findings being recorded after every intervention, including findings that contradict the prescription. Maintenance management systems designed for compliance reporting tend to capture that work closed, not what was seen. Without deliberate instrumentation of the return path, the system trains on its own assumptions and grows more confident while getting less accurate.

Even as McKinsey studies show that Gen AI copilots can help reduce unplanned equipment downtime by up to 90%, 35% of industrial leaders told HFS they had no proven framework for how humans and machines should work together. When the model works but the system around it was not engineered to keep it honeds, prescriptive programs will stall after pilot.

Engineering Inside Operational Constraints

Building this in a live operating environment means embedded engineering and IT/OT transformation inside conditions that do not relax: safety certification, regulated change control, assets that cannot be withdrawn for convenience, and OT systems that predate the concept of an API.

GlobalLogic has engineered against this pattern across Hitachi’s Energy, Rail, and Connective Industries businesses, contributing to HMAX digital asset management platforms in each. In rail, HMAX spans fleets, signaling systems, and infrastructure on an edge-to-cloud architecture developed with Hitachi Digital and NVIDIA.

Platforms of that class depend on the foundation beneath them holding: signals that are calibrated and traceable, OT estates that can be read without compromising their determinism, and a validation path that returns what technicians actually found. That foundation is what GlobalLogic engineers, with more than 500 edge AI experts and 10 years of industrial automation and AI experience applied to problems of this shape.

Contact GlobalLogic to scope the data foundation behind your prescriptive program, starting with a single asset class.

With contributions by Yevgenii Kolometskyi, Associate Vice President, Engineering, GlobalLogic