AI, Utilities, and Solving Key Business Problems: Part 3 – Why We Need to Think About Data Differently

In the last article, we talked about how algorithms are advancing, but your data may not be ready to leverage them. Now, let’s talk a little about what that means in practice, and more positively, what you can do about it.

What Got Us Here Won’t Get Us There

Many utility organizations already have mature data governance, with standards, effective policies, classification frameworks, and decades of established processes. That approach isn’t wrong, and it was built to solve the problems we had at the time: reporting, compliance, operational visibility, and analytics. But AI introduces a different requirement.

We have organized our data around systems, but AI requires us to organize around decisions. And that changes the way we need to think about our governance models. Our old methods allowed us to report on what had happened in the past and create accurate historical pictures. But AI doesn’t work that way. AI is built to perform predictive modeling, and to do that it needs context.

Consider a pole in your GIS system. You’ve got its location, age, and material type. But what your analyst knows is that pole data came from a field survey in 2018, when you didn’t catalog information about pole attachments. So, when your AI model tries to predict which poles might fail in the next year, it can’t see the attachment data. Anything that depends on attachment information becomes inference, not certainty.

This is why our traditional thinking needs to evolve when we are talking about AI; the model doesn’t care WHERE (what source system) the data is from; it needs to know WHY the data matters. Humans naturally understand context. We know where information comes from and what we need to ask to find limitations. AI only knows what we tell it (or what we tell it to find out). The question isn’t “what system should own this data” but instead “what does this data mean”?

Thinking Differently…What Does it Look Like?

Data needs to become an asset, independent of the storage system. For years we’ve treated applications as strategic assets and data as a byproduct of those applications. AI inverts that relationship. The applications still matter, but the data becomes the enduring asset that creates value across systems and use cases. While it seems like a small shift, this changes a lot of what we think about data.

Geospatial data is a good place to start. A GIS contains a treasure of data: assets, locations, materials, potentially even inspection records. But we’ve historically done a very poor job of tracking metadata on that information meaningfully. We know the details about the pole, but not how those details were gathered, or the critical assumptions made in the gathering. When an AI uses data for inference, it won’t have the full picture necessary. The issue isn’t unique to GIS. It simply illustrates a challenge that exists throughout the enterprise.

Operational data has many of the same issues. SCADA systems track real-time grid conditions. Smart meters capture consumption. Sensor networks stream data constantly.

Each system is valuable on its own, and the new opportunity is that AI creates value when those systems can be understood together.

A predictive model evaluating pole failure doesn’t care whether information originated in GIS, SCADA, work management, or weather systems. It only cares whether those datasets provide enough context to make a reliable prediction.

We see the pattern across the utility in work orders, customer data, and, especially, imagery. So much data, all with different owners, fragmented standards, and no comprehensive way to tie that data to the enterprise analysis that AI empowers.

The industry is steadily responding to the change with support – ISO/IEC 5259-3 was published in 2024^[1] to specify what data quality looks like in machine learning. We have historically focused on the idea of traceable, verifiable, complete. AI expands that definition. Now we need to understand where information came from, when it was collected, what assumptions existed when it was gathered, and how confident we are in its accuracy.

The context surrounding the data becomes almost as important as the data itself.

If we go back to our pole example, the ISO standard would require you to document: collected via field survey in 2018, no attachment data cataloged, therefore attachment inference carries uncertainty. That metadata is as important as the data itself for training an AI model.

The path to value with AI isn’t data volume. It’s learning to treat data as an independent strategic asset. Cataloging, documenting provenance, making data available to the whole organization instead of locking it away in source systems. Managing data quality, context, and metadata as core requirements, not afterthoughts^[2].

Why Governance, Context, and Ownership Matter (And What You Can Do About It)

Back in article 2 (link if you’d like to catch up), we talked about how up to 80% of AI projects fail due to poor data quality, not bad algorithms^[3]. And we’ve established that “poor data quality” in this context means siloed and ungoverned data. Data without a shared context that provides meaning, and the underlying documentation that provides authority and confidence.

So what does it look like to step back and start with the data foundation? The good news is you don’t have to fix everything all at once. Begin by asking:

• What’s the business outcome we are trying to achieve?
• What data does that require to be integrated?
• What do we need to know about the data to make our analysis value-added?

If we go back to our pole example, we would start by understanding the use case. The goal is to be able to create a proactive model for storm hardening, meaning we need clear information about the equipment on our poles and how likely it is to fail under specified conditions. So, the context around attachments becomes critical, as evaluating structural integrity and capacity are key factors in that analysis.

Then we can move towards understanding the data and creating governance models that help keep the metadata necessary up to date for the use case. Then our AI algorithm can start to provide meaningful results, and we can generate powerful forward-looking analysis.

AI delivers when the data is trustworthy, documented, and easily interpreted. When you get excited about AI, throw a little excitement the way of data management, because that’s where the fun starts. It’s the hard work and foundation that enables everything that comes after.

For most utilities, an investment in AI is not a question. The question is whether the organization is treating data as a strategic asset capable of supporting the decisions it is looking to improve with AI. Better data isn’t the outcome. Better decisions are.

At UDC, that’s our specialty. We work with utilities to diagnose where data is fragmented, where critical issues are, and how to solve your business problems to get the outcomes you need.

Next in this series: We’ll dive into the better decisions, operations, and outcomes you get when you change the way you think about data.

Missed part one or part two of TJ’s AI series? Read them here:

Part 1: AI Isn’t a Magic Wand

Part 2: AI by Any Other Name