A distribution line inspector covers 40 miles of circuit in a day, carries a tablet loaded with three years of thermal images and vegetation reports, and has a prioritized work list that was last updated four months ago. AI can read every image and sensor record in that backlog before the inspector leaves the yard, triage which assets need eyes on them today, and explain exactly why. What it cannot do, and must not do, is decide that a pole passes inspection. That decision stays with the inspector holding the hammer.

Why Asset-Health Triage Matters in 2026

Distribution infrastructure in most of North America was built in the 1950s through 1980s. Wood poles rated for a 25-year service life are now in their fourth decade. Transformers carrying twice their design load due to EV charging and behind-the-meter solar. Overhead spans with vegetation clearances that were surveyed three years ago in a different precipitation regime. The maintenance backlog is not an operations failure; it is a math problem. There are more assets to inspect than there are inspector-hours to inspect them.

The traditional response to this problem is a fixed inspection cycle: every pole on Circuit 441 gets a ground-level inspection once every five years, regardless of its age, loading history, or location relative to high-wildfire-risk zones. Fixed cycles made sense when data was scarce. In 2026, they are an artifact of the pre-data era.

Modern distribution systems generate continuous sensor data from intelligent electronic devices (IEDs), smart meters, fault indicators, and reclosers. Inspection programs have accumulated years of LiDAR (light detection and ranging) surveys, thermal images from aerial patrols, and condition reports from prior ground inspections. The data exists to prioritize inspections by risk rather than by calendar. What has been missing is the ability to synthesize it all into an actionable triage ranking fast enough to be useful.

AI changes that equation. A classification model trained on historical failure data can read an asset's sensor profile, thermal history, loading record, and vegetation proximity score and produce a risk score that tells a field supervisor which 12 assets on today's patrol route most need close attention. That triage ranking is what replaces the five-year fixed cycle with a risk-based cycle. The field inspector's time is redirected toward the assets most likely to fail, rather than distributed evenly across all assets regardless of condition.

The goal is not to automate the inspection decision. The goal is to give the inspector the best possible starting list so that no high-risk asset goes uninspected because it happened to fall at the end of a fixed cycle.

The Data Inputs to Asset-Health AI

Asset-health AI draws on at least four distinct data streams. Understanding each stream, what it measures, what it misses, and how it fails, is essential to interpreting AI triage output with appropriate skepticism.

Sensor and SCADA Data

Load data from smart meters and SCADA (Supervisory Control and Data Acquisition) points records how hard each asset has been working. A transformer that has been at 95 percent of rated capacity on peak days for three consecutive summers is at materially higher risk of thermal degradation than one running at 60 percent. A recloser that has operated fifteen times in the past twelve months on a circuit that historically averaged two operations per year is telling you something about vegetation encroachment or equipment condition on that segment.

SCADA data is continuous and high-frequency, which makes it valuable for detecting recent changes in asset behavior. Its limitation is that it measures electrical performance, not physical condition. A pole that is structurally compromised but still standing will not show a SCADA anomaly until it fails.

Imagery and LiDAR

Aerial imagery from fixed-wing or drone patrols captures the physical condition of spans, poles, and hardware in ways that SCADA cannot. Thermal imaging identifies hot spots on conductors and transformer bushings that precede failure. LiDAR data measures conductor sag, vegetation clearance distances, and pole lean angles with centimeter-level precision.

Computer vision models can process aerial imagery at scale, flagging spans where conductor sag exceeds a threshold, poles where lean angle is outside acceptable tolerance, or vegetation proximity scores that indicate a potential contact risk. These flags are the AI's contribution to the triage list. The limitations are resolution, inspection recency, and environmental interference: a thermal image taken in cold weather may not show a hot spot that would be visible in peak-summer loading conditions, and LiDAR surveys have a finite update frequency that may not reflect vegetation growth since the last survey.

Inspection History and GIS

Prior inspection records and GIS (Geographic Information System) data provide the static context for the sensor and imagery data. A pole's age, manufacturer, treating method, and prior structural condition rating create a baseline against which the AI model assesses the current sensor and imagery data. A 45-year-old wood pole with a prior condition rating of "marginal" is a different risk profile than a 15-year-old pole with no prior flags, even if their current SCADA loading profiles look similar.

The limitation of inspection history data is its reliability. Manual condition ratings entered into GIS vary by inspector, by year, and by the tools available at the time. A "good" rating from 2011 using visual ground inspection only is not equivalent to a "good" rating from 2023 using ground-penetrating decay probe. AI models trained on historical inspection records inherit whatever inconsistencies exist in those records.

How AI Produces a Triage List

The AI triage process combines these data streams into an asset risk score through a classification or regression model that has been trained on historical data. "Historical data" in this context means past inspection records, past failure events, and the asset characteristics and sensor profiles that preceded those failures. The model learns which combinations of age, loading, thermal history, and vegetation proximity score are most predictive of actual failures or inspection-identified deficiencies.

The output is a ranked list of assets, each with a risk score and, if the model is designed to explain its reasoning, a set of contributing factors. An output might read: "Pole 4417-B, risk score 87/100. Contributing factors: age 42 years (weight 35%), vegetation proximity score 8.2 (weight 28%), loading history peak at 94% rated capacity for 18 of the last 24 peak-demand periods (weight 22%), thermal flag from March 2025 patrol (weight 15%)."

That output is the starting point for the field inspector's work list. It is not the inspection decision.

What the Risk Score Means and Does Not Mean

A risk score of 87/100 does not mean the pole will fail. It means the model believes, based on similar patterns in historical data, that this asset has a higher probability of requiring maintenance or exhibiting a deficiency upon close inspection than a pole with a score of 45/100. The score is a probability estimate, and like all probability estimates, it will be wrong for individual assets some of the time.

A high-scoring asset that passes inspection teaches the model something, provided the negative inspection is recorded correctly. A low-scoring asset that fails catastrophically is an event that should trigger a review of what the model missed. Asset-health AI improves over time only if inspection outcomes are fed back into the training data, and only if the feedback captures both true positives (high score, deficiency found) and true negatives (high score, asset is fine).

The inspector who trusts the risk score completely and gives low-attention to the low-scoring assets on the patrol route is creating a systematic coverage gap. Assets that the model scores incorrectly will go uninspected until the model is corrected. Maintaining a baseline inspection requirement for low-scoring assets at a reduced frequency is the standard risk-management control for this failure mode.

Using AI for Vegetation-Management Triage

Vegetation management is the largest single category of distribution outage causes in most of North America. Trees that grow into or contact overhead conductors cause outages, fires, and in high-wind-fire-risk conditions, can initiate the kind of catastrophic wildfire events that have reshaped utility risk management in the western United States.

AI-assisted vegetation triage uses LiDAR and aerial imagery to identify locations where vegetation clearance is below a defined threshold, where growth rates indicate that clearance will fall below threshold before the next scheduled trim cycle, and where the combination of clearance proximity and wildfire-risk zone classification makes the location a priority for immediate trimming rather than scheduled maintenance.

The triage output maps directly to a work list for vegetation management crews. A segment with a LiDAR-measured clearance of 1.2 meters in a High Fire Threat District (HFTD) gets flagged for immediate response. A segment with 4.5 meters of clearance in a low-risk zone gets scheduled for the next available trim cycle.

The verification discipline for vegetation triage is the same as for physical asset triage: the AI flags, the human decides. A LiDAR reading of 1.2 meters is a distance measurement, not an inspection determination. The field supervisor visiting that location may find that the LiDAR captured a limb that has already been trimmed in a non-scheduled trim, or that the distance measurement reflects a temporary condition (a vine-covered fence near the line, not an actual tree encroachment). The AI flag gets a field verification before it drives a work order.

Keeping the Human Inspection Decision

The asset-health triage workflow has a clear boundary that must not be crossed: the AI provides the priority ranking, the human inspector makes the pass or fail determination. That boundary is not a limitation of current AI capability to be overcome in future releases. It is a fundamental property of how regulated utility infrastructure works.

An inspection decision on a wood pole, a transformer, or a conductor span is a professional engineering judgment that carries legal accountability. A pole that passes inspection has been evaluated by a qualified person who applied training, experience, and physical senses to a specific asset at a specific point in time. That judgment cannot be delegated to a risk score.

The accountability chain matters practically in two contexts. First, in the event of a failure: if a pole that the AI scored as low-risk fails and causes an outage or a fire, the utility must be able to demonstrate that its inspection program applied appropriate human judgment, not that it relied on the AI score to skip the inspection. Second, in regulatory compliance: state PUC requirements for distribution inspection programs specify qualified-person standards that require human inspection, not algorithmic certification.

What AI changes is the prioritization upstream of inspection, not the inspection itself. The inspector still uses a hammer, still probes for decay, still evaluates hardware condition, still makes the call. What AI changes is the order in which assets are inspected and the context the inspector has when they arrive at an asset: a pre-loaded summary showing the asset's risk factors, thermal history, and prior inspection notes.

The AI gives the inspector the best possible briefing before they touch the asset. The inspector gives the asset the best possible evaluation after they have read that briefing. Both steps are required.

Worked Example: When the Triage List Gets It Wrong

Here is how an asset-health AI failure looks in practice and how an informed field supervisor catches it. A vegetation management team receives a triage list for a wildland-urban interface circuit. The top 15 locations on the list all appear in a single 2-mile segment that the LiDAR captured in October of last year. Three other segments on the circuit, including one that was not covered in the October LiDAR flight due to cloud cover, show no flags.

A vegetation supervisor who does not look past the triage list sends crews to the 15 flagged locations. The uncaptured segment, which experienced significant rapid growth following a wet spring, goes unpatrolled. Sixteen weeks later, a tree contact causes an outage that the triage list would have flagged had the LiDAR data been current.

An informed supervisor reads the triage list differently. They check the data provenance: when was the LiDAR data collected for each segment? Which segments had coverage? The triage output should include the source data vintage as part of the risk score explanation, because a score derived from 14-month-old imagery in a high-growth region is materially less reliable than a score derived from 3-month-old imagery. Finding that the triage list has a geographic coverage gap, the supervisor adds a manual patrol of the uncaptured segment to the work plan before approving it.

This worked example teaches two things. First, the triage list is only as good as the input data that fed it. Second, the field supervisor's role in reviewing the triage list is not passive acceptance but active interrogation: where did this list come from, what was included, and what might be missing?

Deploying Risk-Based Inspection: Workflow and Workforce Implications

Deploying risk-based inspection is not just a technology deployment. It is a workflow redesign. The field supervisor's morning briefing changes: instead of "here is the circuit you are walking today," it becomes "here is the circuit, here are the top 12 assets the model flagged, here is why each was flagged, here is the data vintage for each flag, and here is the baseline coverage requirement for the rest of the circuit." That briefing takes five minutes with well-designed AI tooling. Without it, the inspector arrives at the circuit without context and makes prioritization decisions by experience and intuition alone, which is exactly what the fixed cycle was designed to compensate for when institutional knowledge is unevenly distributed.

The Great Crew Change means that experienced inspectors who knew which circuits had historically had problems, where the aggressive tree species were, and which transformer banks had been overloaded in prior summers are retiring. The AI triage system captures some of that institutional knowledge in its training data, provided the historical inspection records are complete enough to train on. A utility that has maintained rigorous inspection records for two decades has better training data than one whose records are patchy, and a better AI model as a result. The incentive to maintain good inspection records is not new, but the AI era has given it a new financial and operational rationale.

High Fire Threat Districts (HFTD) in western states add a regulatory dimension to triage deployment. California's Public Safety Power Shutoff (PSPS) framework and similar programs in other western states set explicit clearance-distance requirements for overhead lines in specified fire-risk zones. For utilities operating in or adjacent to HFTDs, the AI vegetation triage output must be aligned with the state PUC's mandated clearance standards, not just the utility's internal maintenance thresholds. A triage tool that uses internal operational thresholds in HFTD zones may produce a work list that is technically compliant internally but does not meet the more stringent regulatory standard. Building the HFTD overlay and the state-mandated clearance tiers into the triage rule set is the configuration step that makes the AI output regulatory-defensible, not just operationally useful.

Key Takeaways

• Asset-health AI replaces fixed inspection cycles with risk-based prioritization by synthesizing sensor data, aerial imagery, LiDAR, and inspection history into a ranked triage list. The inspector's time is redirected to the assets most likely to have deficiencies.
• The four primary data inputs are SCADA and sensor data (electrical performance), aerial imagery and thermal (physical condition), LiDAR (clearance and geometry), and GIS inspection history (baseline context). Each has specific limitations that affect the reliability of the AI's risk score.
• An AI risk score is a probability estimate, not an inspection decision. High-scoring assets require human inspection; low-scoring assets require inspection at reduced frequency, not exemption. Trusting the score completely creates systematic coverage gaps for assets the model scores incorrectly.
• Vegetation management triage using LiDAR and aerial imagery identifies clearance-deficient locations before outages occur. AI flags drive a prioritized work list for trimming crews, but field verification is required before each flag becomes a work order.
• In High Fire Threat Districts, the AI triage rule set must be configured to the state PUC's mandated clearance standards, not just internal maintenance thresholds. A work list that passes internal thresholds but misses regulatory clearance requirements creates enforcement exposure in addition to operational risk.
• The human inspection decision is not a transitional limitation to be automated away. It is a professional engineering accountability requirement that is enforced by state PUC inspection standards and the legal accountability framework for distribution infrastructure safety.
• Triage list interrogation is as important as triage list generation. A field supervisor who checks the data provenance (imagery vintage, LiDAR coverage date, sensor data completeness) catches the gaps that the AI cannot flag because the gap is an absence of data, not a data anomaly.
• Inspection outcomes must be fed back into the AI model to improve its accuracy over time. Recording both deficiency findings and clean inspection results for high-scoring assets is the quality loop that makes risk-based inspection better than the fixed cycle it replaces.