Rethinking Ai: Why Conventional Ai Is Failing The Grid, A…

Contributed by Donald McPhail | VP of market development at eSmart Systems

Utilities are being asked to inspect more assets, more often, with fewer specialists, while regulators sharpen severity thresholds and aging infrastructure throws up new failure modes every year. Computer vision AI models were supposed to help close that gap. In many ways, they have. But the deeper anyone gets into operational AI for the grid, the more apparent it becomes that the standard pipeline of large labeled datasets, long training runs, and periodic redeployment is not just slow. It is structurally mismatched with the work.

A new white paper from eSmart Systems, From Months to Minutes: A New Architecture for Domain AI Models, argues that the answer is not more data or more compute, but a different architecture entirely. The paper introduces Adaptive AI, a patent-pending approach designed for environments where conditions change, expert intent matters, and the cases that count operationally are usually the rarest ones in the dataset.

Where the Conventional Pipeline Breaks Down

Traditional deep learning depends on large, balanced labeled datasets. In power grid inspection, that assumption collapses with the real-world operations. Asset types are highly diverse across voltage classes, materials, vintages, and configurations. Operating conditions vary with weather, lighting, vegetation, geography, and mounting orientation. And the definition of what counts as a defect shifts between customers, regions, and regulators.

The result is an uneven accuracy profile. Models perform well on common components and frequent fault types, but poorly on low-frequency, high-severity conditions that matter most: rare insulator damage, early-stage structural degradation, and emerging failure modes with no deep training history. The cases an inspector needs help with the most are the ones the model is least equipped to flag.

Once a conventional model is trained, its knowledge is effectively frozen. Improving it means more data, more labeling, more training, more resources, and more redeployment, and a cycle that takes months to run. When a regulator or industry standard updates a severity threshold or a new asset class comes into scope, the retraining clock resets, and the labeling burden falls back on the domain experts whose time is already in shortest supply. It’s an expensive and time-consuming approach, and it’s holding back the rate of scaling.

A Different Architecture: Adaptive AI

Adaptive AI replaces the retrain-and-redeploy cycle with a continuous, human-in-the-loop learning architecture. It builds on top of large pre-trained foundation models, which act as general-purpose feature extractors, and adds a modular domain intelligence layer that subject matter experts can configure, refine, and extend directly. There is no data science team in the middle of the loop, no relabeling step, no redeployment pipeline.

The practical consequence is that new fault types or conditions of interest can be recognized from a handful of image examples, and that changes take effect immediately. A field engineer who notices an unfamiliar fault pattern in inspection imagery does not file a request that joins a months-long queue. They flag three examples in the system, and Adaptive AI picks up the new condition and begins surfacing it across subsequent imagery on the network.

The white paper is careful to distinguish Adaptive AI from few-shot learning, which has become a common shorthand for any approach that learns from small data. Few-shot methods are a valuable ingredient, but not on their own a complete solution. They typically treat all provided examples as equally informative, offer no built-in mechanism for continuous refinement, and provide no guidance on what the system should learn next. Adaptive AI is the architecture built around and beyond few-shot learning, with two things in particular that set it apart.

The first is that Adaptive AI treats foundation models as evolving infrastructure rather than a fixed dependency. When a newer, more capable foundation model becomes available, performance improves automatically with no retraining or relabeling. In a documented crossarm material classification benchmark, simply switching from one foundation model (OpenAI’s CLIP, 2021) to a newer one (Meta’s PE, 2025) lifted F1 accuracy from 0.72 to 0.89 in under a minute. A traditionally trained model on the same task reached 0.94, but only after one to three days of work plus extensive annotation, and only improves further by repeating that cycle.

The second is a continuous feedback loop built into the system. Concepts are updated by adjusting stored example representations rather than retraining model weights, so each iteration is instantaneous, and unhelpful changes can be reversed just as quickly. That low cost of failure changes the nature of AI development. Teams can experiment, iterate, and converge on high-performing models without the cautious, slow-moving cycles that traditional retraining demands.

What This Looks in the Field

eSmart Systems has applied this approach inside its Grid Vision platform for aerial and drone-based inspection across transmission and distribution networks. In March 2026, the company launched AI Studio, which lets users build and test Adaptive AI models directly, including for use cases well beyond overhead line inspection. The potential extends to any situation where an organization needs detection or classification from imagery, across the energy value chain, and into adjacent infrastructure sectors.

“The biggest bottleneck in operational AI isn’t model accuracy. It’s the time it takes to get a model aligned with what the people using it actually need it to do,” said Erik Åsberg, CTO of eSmart Systems. “We developed Adaptive AI to close that gap, so that when a utility’s priorities change or a new condition shows up in the field, the system can respond in minutes, not weeks or months.”

There is a broader implication for utilities and asset owners thinking about how AI fits into their operations over the long term. When domain experts can build and refine models themselves, specialized knowledge is no longer mediated by data science teams and annotation pipelines. Over time, an organization accumulates a growing library of refined concept detectors, each representing a specific condition, component, or failure mode, which can be reused across applications and composed into larger systems. Foundation models become a platform the domain layer rides on, rather than a ceiling that limits what can be built on top of them.

Minutes, Not Months

Infrastructure inspections, asset management, and condition monitoring operate under fundamentally different constraints than consumer or enterprise AI. Events of interest are rare, but decisions are safety-critical. Physical variability is the norm rather than the exception. There is regulatory accountability and a legal liability behind every classification, and a missed detection has real-world consequences.

The Adaptive AI architecture is designed for those constraints. It is a response to operational problems that utilities and asset owners face today, not a projection of what AI might do tomorrow. Ultimately, this enables operators to increase the speed to value from their investments in AI for safer, more reliable, and more efficient infrastructure.

The full white paper, “From Months to Minutes: A New Architecture for Domain AI Models,“ is available for download from eSmart Systems. It covers the architecture in more technical detail, including the head-to-head benchmark and the strategic implications for organizations building long-term AI capability in operational settings.

---

About the Author

Donald McPhail is vice president of market development at eSmart Systems. He has more than 15 years of experience working with electric utilities and technology vendors across the United States, Australia, the United Kingdom, and Europe, helping organizations integrate asset intelligence, component-level inspection data, and risk modeling into wildfire mitigation, extreme weather resilience, and grid modernization strategies.