Key takeaways:
- The biggest barrier to AI adoption is fragmented, untrusted data spread across GIS, asset management, engineering, and field systems.
- Legacy systems have real business value, so aim to connect AI to them (e.g. through MCP) rather than replace them.
- Design security and governance from the start, because a single breach can destroy trust in an AI program.
- Shadow AI is already in your organization, so channel it through approved tools, a clear approval process, and human review.
- Modernization should start with a well-understood business problem and clear data ownership, with technology selection coming last.
Almost every organization wants to put AI to work, and so many run into the same wall: data spread across systems that were never designed to talk to each other. Asset records, operational data, field notes, spreadsheets, and more, all in siloed systems. Utilities companies face this problem at enormous scale, which makes them a useful case study for anyone working with spatial and operational data.
In a June 2026 survey of 87 utilities and energy professionals across 11 countries and regions, 86% said their organizations have put AI to the test. Yet only 23% have it running in production. That gap is the modernization challenge in a nutshell. The organizations moving from pilot to production are the ones fixing the foundations first: their data, their governance, and their people. Let’s discuss what that looks like in practice and how FME can help.
To listen to a panel discussion and dive deeper into this topic, watch our webinar recording, AI, MCP, and Modernization in the Utilities Industry: A Panel Discussion.
Why AI pilots stall
Often, teams experiment with AI, hear about security risks, and stop. Without a top-down strategy covering governance, security, and training, nobody knows how to move a promising experiment into production safely.
Underneath that hesitation is a more basic problem: trust in the data. Utilities companies hold decades of records spread across systems that each hold only part of the picture. When people don’t know which sources an AI is drawing from, or what they’re allowed to share with it, they’re understandably reluctant to act on its answers.
What data fragmentation looks like on the ground
In the same survey, 77% of respondents named data quality and standardization as the main obstacle to modernization. “Is your data AI-ready?” is a new way of asking an old question.
Take a typical asset replacement program. GIS says a pipe was replaced in 2018, but the asset management system says 2020. Engineering drawings show a different configuration, field crews keep their own notes, and a few spreadsheets were never integrated with anything. When a decision has to be made, the team isn’t sure which data to trust.
This is where data integration comes in. With FME, you can build workflows that read from GIS, asset management, engineering records, and spreadsheets, compare and reconcile them, and flag conflicts for review. Then, run the workflows on a schedule or in response to a trigger so the reconciliation happens continuously.
Tip: The hardest silo is between people
Technology can link systems fairly quickly. The harder silo is between departments, each with its own priorities and measures of success. Agreeing on data ownership, governance, and shared processes usually takes far longer than any integration project, so start those conversations early.
Connect, don’t replace, legacy systems
Only 22% of respondents named legacy infrastructure as their top challenge, the least-cited option. Many legacy systems still provide real business value, and replacing them can be complicated. The better question is how to bring AI to the systems you already have.
That’s where the Model Context Protocol (MCP) comes in. MCP is an open standard that lets AI models connect to external tools and data. Think of it as a general-purpose integration protocol, i.e. a standard way to connect anything to anything.
Instead of copying data out of every system into a new central store, you expose specific, well-described operations that an AI can call when someone asks a question in plain language. (This doesn’t mean the AI gets free rein. Existing credentials and permissions still apply.)
FME supports MCP in both directions. FME Flow can act as an MCP server, turning published workspaces into tools that AI clients can discover and call, with FME Flow handling authentication and logging. The MCPCaller transformer works the other way, letting an FME workspace use tools from other MCP servers. Because every tool is a workspace you design, you decide exactly which records and attributes the AI sees, and you can mask or obfuscate the rest. FME reads thousands of formats, including systems that predate modern APIs, so even older systems can be made available to AI without being replaced.
Security and governance come first
31% of survey respondents named security, cybersecurity, and data privacy as their biggest barrier to AI adoption. For a regulated industry, that number arguably should be higher. Security should be foundational, i.e. you don’t build an AI capability and then bolt security on afterward, because the first breach is usually the end of trust.
The fear that AI means “your data is flying out somewhere” is legitimate because anyone can paste a spreadsheet into a public chatbot. But established architectures can keep data risk at the level you accept today. On-premises models keep sensitive data behind the firewall, and tool-based access through MCP exposes only what you choose. FME can be deployed on-premises, in any cloud, or in hybrid setups, and works with whichever AI model you pick, so your governance doesn’t depend on a single vendor.
Put governance in place before AI rolls out, and communicate clearly which agents are approved and what data can be shared with them.
Shadow AI is already here
Even if your organization has no official AI program, someone has already asked an AI tool to help with a work problem. Assume AI has touched any work that crosses your desk, and make sure a qualified person reviews it before it reaches a customer or drives a decision.
Practical steps include a single point of contact that approves AI tools against security and data standards, an enterprise AI assistant so people have a sanctioned option, and asking why people work around official channels. Often the answer is a backlog: requests for data from multiple systems become IT tickets that wait in a queue. FME’s no-code approach lets engineers, analysts, and operations staff build and test their own integrations and reports in FME and share them under IT’s governance, instead of waiting or working around the system.
Start with business outcomes, not technology
Only 11% of respondents named an unclear AI strategy as their biggest barrier. Some of the rest likely have a clear strategy for certain workloads: not yet. Nobody is ready to hand an agentic system control of an organization’s operational technology (OT) environment, and that’s sensible. Start with lower-risk use cases, learn, and save critical systems for later.
A blanket “no AI” policy, though, is likely to fail. Instead, make technology selection one of the last conversations. First ask, as a whole organization: What are we trying to achieve? What’s in our way (disconnected systems, fragmented data, inconsistent processes)? Which priorities create the most value? Who owns each data set?
From reactive to proactive: integrating GIS with operations
59% of respondents identified integrating GIS and spatial data with operational systems as a top data integration challenge, and it’s where much of the payoff lives. Combining GIS with outage management (OMS), advanced metering infrastructure (AMI), and SCADA data lets utilities see what’s happening on the grid without waiting for customers to call. Ahead of a storm, a utility can combine network data, imagery, and weather forecasts to pre-position crews where damage is most likely.
FME is built for this kind of spatial-to-operational integration, connecting GIS platforms with operational and enterprise systems and automating the data flows between them. FME Realize takes it a step further, letting teams experience that data in real-world context, in real time.
Get started
AI won’t fix fragmented data, but connected, well-governed data makes AI far easier to adopt. If your organization is sitting on legacy systems full of valuable information, that’s a starting point, and FME can help you build from it.
Learn more:
