BriefingInfrastructure

Assets That Outlive Their Builders

Condition monitoring, digital twins and AI-assisted capital planning for infrastructure owners, and how to buy software for assets built to last a century.

Arclight Research5 min read

Fig. I · Infrastructure

In brief

  1. IStart with the asset register. AI applied to an inaccurate inventory gives confident answers to the wrong questions.
  2. IIA digital twin justifies its cost only when the asset changes faster than inspections, decisions need simulation and someone owns upkeep.
  3. IIIBuy the model, own the data. Contracts should keep condition data in open formats that will outlive any vendor or platform.

When Frontinus took charge of Rome’s water supply in AD 97, he set out to understand the system before managing it. He surveyed the aqueducts, measured what they delivered and compared the results with the official records. The records did not match the water.

Infrastructure owners face the same problem nineteen centuries later. Water systems, bridges, transmission networks, rail and ports are built to serve for fifty years or more, and the records describing them were assembled across generations of engineers, contractors and acquisitions. Before AI can help manage these assets, it has to help establish what is actually there.

The register comes first

Capital planning rests on the asset register: what is owned, where it is, what it is made of, when it was installed and what condition it was last found in. In older systems the register is often incomplete and inconsistent, and the gaps tend to cluster in the oldest and least visible assets, which are frequently the ones that matter most.

This is where AI should start. Document intelligence can read decades of as-built drawings, work orders and inspection reports to fill missing attributes, computer vision on street-level and aerial imagery can confirm what is in the field, and entity matching can reconcile records from systems that were never designed to agree. The result is a register with a confidence level on every attribute, which is more honest than one that presents every field as fact.

Condition monitoring that changes decisions

Inspection is where the technology has matured fastest. Drones and fixed cameras capture bridges, towers, insulators, dams and rail corridors, and computer vision flags cracking, corrosion, spalling and vegetation encroachment across far more imagery than inspectors could review by hand. Acoustic sensors locate leaks in buried water mains, fiber-optic sensing turns a pipeline or rail line into a continuous sensor, and satellite radar measures millimeter-scale ground movement around embankments and dams.

The limits are specific. Vision models detect defects but cannot judge structural significance, which depends on load paths, design and history. Serious defects are rare, so a model can score well on average while missing the one that matters, and imagery collected without reference to the owner’s inspection standard creates a second, unofficial record.

The test for any monitoring program is whether it changes a decision: an inspection interval, a repair priority, a load restriction, a capital plan. Data that changes nothing is a cost.

When a digital twin is worth it

Digital twins are often presented as the destination of every infrastructure program, and the common failure is an expensive three-dimensional model that nobody updates. A useful twin is a working model of an asset’s behavior, kept current with field data, that lets engineers ask what happens if. Water utilities have run hydraulic network models for decades, and that is a twin in all but name. A twin justifies its cost when three conditions hold:

  • The asset changes faster than the inspection cycle. Loads, flows, temperatures or deterioration move on a timescale that periodic inspection misses.
  • Decisions depend on simulation. Operators need to test a closure, an outage, a load increase or a failure before committing to it.
  • Someone owns keeping it current. Updating the twin is a funded operating responsibility with a named owner, not a project deliverable.

Where those conditions hold, physics-informed machine learning is often the right technique, because it constrains models with the engineering the asset was designed to, which matters when failure data is scarce. Where they do not, a good register, a risk model and disciplined inspection deliver most of the value at a fraction of the cost.

The asset will outlast the model, the vendor and probably the company that installed the sensors.

Capital plans, permits and project controls

Capital planning

This is the decision the rest of the program exists to inform. Survival models estimate failure probability across pipe cohorts, bridge decks or transformer fleets from material, age, environment and history, and optimization selects the renewal program that buys the most risk reduction within a budget and a crew capacity. Owners working to the ISO 55000 asset management standards will recognize the logic; AI makes it practical at the level of individual assets and makes the risk carried in the maintenance backlog visible.

Permitting

Permitting is document work at scale. Environmental reviews can run to thousands of pages, and much of the effort goes into assembling the record, cross-referencing it and responding to public comments. Language models with retrieval over the project record can draft sections, check consistency and cluster thousands of comments by issue, provided every statement traces to a source the agency can audit.

Project controls

Models trained on an owner’s past projects can estimate schedule and cost risk at the outset, a practical form of reference class forecasting: judging a project by how similar projects actually turned out rather than by its own plan. They can also read RFIs, submittals and change orders to flag disputes before they become claims. The constraint is the owner’s own history, because projects never closed out with clean data leave little to learn from.

The long-life mindset

Software platforms last a few years, and vendors are acquired or leave the market. The asset will outlast the model, the vendor and probably the company that installed the sensors. That should change how infrastructure owners buy:

  • Own the data. Condition data, imagery and labels should be held in open formats, exportable without fees or the vendor’s cooperation.
  • Record the reasoning. Keep what each model showed and why each capital decision was made, for the engineers who will revisit it in twenty years.
  • Pay for outcomes in the field. Tie fees to verified defects, avoided failures or inspection hours saved, rather than to dashboards delivered.
  • Maintain models like components. Version them, monitor them and retire them deliberately.

The owners who manage this well treat today’s condition data as an inheritance. It will train models that have not been built yet, used by engineers who have not been hired yet, to make decisions about assets that will still be in service.

Questions for your next leadership meeting

  1. How confident are we in our asset register, attribute by attribute, and what would it cost to find out?
  2. Which inspection or monitoring programs changed a maintenance or capital decision last year, and which only produced data?
  3. Do our digital twin investments meet all three conditions, and who is accountable for keeping each one current?
  4. If our main monitoring vendor disappeared tomorrow, could we export our data, imagery and labels in a usable form?
  5. Is our capital plan ranked by risk reduced per dollar at the level of individual assets, or still by age and budget line?

ARCLIGHT · RESEARCH

This analysis is general in nature and does not constitute investment, legal or engineering advice. To discuss how it applies to your organization, request a private briefing.

Begin

Apply this analysis to your own assets.

Tell us about your assets, your board and the decision in front of you. A principal replies within one business day.