Why Utility AI Projects Fail Before the First Prompt

Over the past year, I’ve had a lot of conversations with utility organizations about AI. Most of them start from the same place: how can AI help us predict outages, manage assets better, improve inspections, or do more with the resources we have?

All good questions. But in many cases, the biggest obstacle to AI success has nothing to do with the AI itself. It’s the data sitting behind it.

There’s one question I rarely hear: “can our data actually support AI?” In my experience, that’s the question that matters most — and it’s usually the last one anyone thinks to ask.

The problem isn’t a lack of data, it’s disconnected data

Utilities generate enormous amounts of data: asset records, inspection reports, SCADA events, vegetation management programs, work orders, customer information, engineering documents, drone imagery, LiDAR. The issue isn’t a shortage of information. It’s that most of it lives in isolation.

A typical utility might have detailed asset records in GIS, maintenance history in an EAM system, inspection reports stored as PDFs, and operational events flowing in from SCADA. Each system does its job, but none of them tells the whole story on its own. People bridge those gaps through experience and institutional knowledge. AI can’t.

“People bridge data gaps through experience and institutional knowledge. AI can’t.”

Data silos get exposed by AI

Organizations have been working around disconnected systems for years. An engineer exports a spreadsheet, an analyst manually joins datasets, a department maintains its own version of a report because it doesn’t fully trust the source system. It’s not ideal, but the work gets done.

AI changes that calculus. When a model is asked to identify assets at risk of failure or analyze outage trends, it can only work with what it’s given:

• If asset identifiers don’t match between systems, confidence in results drops
• If inspection records are buried in PDFs, important context gets missed
• If operational data isn’t connected to asset records, patterns become harder to detect

What looks like an AI problem is often a data integration problem underneath.

A simple test for your organization is to ask yourself: which assets in your service territory are at the highest risk of failure over the next 12 months?

To answer that confidently, you probably need data from GIS, asset management, inspection records, work management, SCADA, weather feeds, vegetation management programs, and engineering assessments. Getting all of that together in a way that’s current, trusted, and repeatable — that’s the hard part. No AI platform can compensate for a weak foundation.

What successful utility AI projects have in common

The utilities I’ve seen make the most progress with AI aren’t starting by evaluating AI tools. They’re asking how to make their data more accessible and usable first — which turns out to be a better place to start.

Before launching large AI initiatives, they’re investing in connecting systems, automating data movement, and improving data quality. And those investments pay off well beyond whatever AI project originally motivated them. Once your data is easier to access and trust, new use cases start surfacing across:

• Asset management and predictive maintenance
• Outage analysis and storm response
• Vegetation management programs
• Regulatory reporting and compliance
• Operational planning and capital prioritization

The next challenge won’t be AI, it will be action

The conversation is already starting to shift. Today, most organizations are focused on AI that answers questions. The next wave will be AI that takes action — retrieving information from enterprise systems, analyzing asset conditions, recommending inspections, generating reports, and triggering downstream workflows.

Imagine an AI agent that identifies transformers showing signs of elevated risk, creates a work order, attaches supporting inspection history, and routes it for review. The technology is emerging quickly. The limiting factor won’t be the AI model, it’ll be whether the underlying systems and data can work together.

Organizations that have already invested in getting their data house in order will be in a much better position to move quickly. Those that haven’t will likely find themselves rebuilding the same integrations every time a new opportunity appears.

Where to start: three questions worth asking now

If your organization is evaluating AI right now, start with these:

• Can your data support what you’re expecting AI to do? Can information move easily between GIS, EAM, SCADA, and your enterprise systems?
• Can teams actually trust what they’re looking at? Data governance and quality are prerequisites, not afterthoughts.
• Can operational, spatial, and enterprise data be brought together without manual effort? If the answer requires a spreadsheet export, that’s a signal worth paying attention to.

In my experience, the most successful utility AI projects don’t start with AI at all. They start with data — and that work begins long before the first prompt is ever written.

Want to learn more about how FME can help you connect your utility data? Book a personalized demo with a utility advisor today!