Why AI Without Orchestration Is Just Expensive Guessing
A recent Utility Dive piece by Sean Burri, an infrastructure engineer at Dominion Energy, makes a point that deserves more attention than it is getting: deploying AI does not automatically improve utility operations. The projects that deliver results are not the ones with the most sophisticated models. They are the ones embedded within enterprise strategy, supported by continuous feedback loops, and connected to the people who actually do the work.
As someone who spends most of his time helping IOUs build the workflow infrastructure behind System Resiliency Plans, I read Burri’s argument and kept nodding. He is describing, from the inside of a major utility, the same pattern I have been writing about in this series from the outside—and the implications for SRP execution are significant.
But first, before I dive in: full disclosure. At Macedon Technologies, we help IOUs design and implement the kind of workflow orchestration that connects AI outputs to field execution and regulatory evidence. Whether it is for rate case defense, vegetation management, or capital project closeout, our goal is to make your transition to automated, auditable planning fast and effective.
If this resonates, grab time with me and let’s talk through it in person.
The Three Obstacles are Really One Obstacle
Burri identifies three barriers to scaling AI at utilities: fragmented data governance, legacy hardware that cannot handle modern data volumes, and cultural resistance from employees who distrust machine recommendations. He is right on all three counts. But from an SRP execution standpoint, these are not three separate problems. They are three symptoms of a missing orchestration layer.
When a utility deploys a vegetation risk model that identifies high-priority spans, that output needs to flow into a work order. The work order needs to be assigned to a contractor. The contractor needs a mobile tool to capture completion evidence. That evidence needs to feed back into the financial system for cost classification. And the whole chain needs to be auditable for the next rate case.
If any of those handoffs is manual—a spreadsheet export, an email chain, a phone call—then the AI model’s output is disconnected from the operational reality. The model might be brilliant. The execution is still broken. And the documentation trail that regulators need simply does not exist.
This is what Burri is getting at when he writes that reliability does not come from software alone. It comes from integration. The question for utility leaders is not whether to invest in AI. It is whether the operational infrastructure exists to make AI outputs actionable, traceable, and defensible.
The "Glass Box" Problem Gets Bigger
In my earlier post on BCR automation, I argued for a “glass box” approach to investment planning—making the assumptions, data sources, and decision logic behind every resiliency investment fully transparent and auditable. Burri makes a parallel argument about AI models themselves: if engineers cannot see why a model produces a particular forecast or recommendation, they will ignore it.
This is not a hypothetical risk. It is happening today at utilities across the country. A risk model flags a distribution segment for accelerated vegetation management. The operations team does not trust the output because they cannot see the inputs. They revert to the calendar-based cycle they have always used. The AI investment produces a report that sits in a dashboard. The field work continues as before.
The fix is not better algorithms. It is an orchestration layer that makes the model’s reasoning visible at every step—what data triggered the flag, what thresholds were applied, what alternative actions were considered—and then routes that context through to the field crew and back to the regulatory filing. When the foreman on the ground can see why this span was prioritized over another, and when the regulatory team can trace that same logic chain into their rate case evidence, the AI becomes trustworthy because it is transparent.
Data Silos are a Documentation Problem
Burri highlights that many utilities maintain separate databases for operations, maintenance, and customer service, creating silos that prevent AI from seeing the full picture. This is true, and it is also the root cause of the documentation problem I described in my post on byproduct documentation.
When data lives in silos, compliance evidence has to be reconstructed after the fact. Someone in regulatory affairs has to pull outage data from one system, contractor invoices from another, work order completions from a third, and manually stitch together a narrative that justifies the utility’s spending in a rate case filing.
That reconstruction process is expensive, slow, and fragile. It is also exactly the process that breaks down when scale increases—which is precisely what is about to happen as NERC’s proposed FAC-003 expansion from 200 kV to 100 kV brings potentially thousands of additional circuit miles under mandatory vegetation management requirements.
The alternative is what I have been calling byproduct documentation: designing operational workflows so that every field action automatically generates the compliance evidence needed for regulatory filings. The work order captures the risk input. The dispatch records the contractor and scope. The mobile completion tool timestamps the verification. The financial system classifies the cost. No reconstruction needed.
But this only works if the silos are connected—not by ripping out legacy systems and replacing them, but by layering an orchestration platform on top that moves data between them in real time. Burri’s observation about data governance is correct, but the solution is not a multi-year master data management project. It is a pragmatic middleware approach that connects what already exists.
Cultural Resistance is a Design Problem
The most important point in Burri’s article, and the one most likely to be overlooked, is his argument about workforce adoption. Employees who have spent decades making decisions based on experience and judgment are not going to suddenly defer to a machine because a dashboard says so. And they should not have to.
The utilities that have successfully scaled AI, as Burri notes, invested in explaining what the technology does, how it learns, and when human intervention remains critical. But explanation alone is not enough. The tool has to be designed so that the human role is embedded in the workflow, not bolted on as an afterthought.
In my post on the vegetation management execution gap, I described the challenge of building mobile completion tools that field crews will actually use. The same principle applies here. If you want a line supervisor to trust an AI-generated risk prioritization, the interface has to show them the reasoning, let them adjust based on local knowledge, and capture that adjustment as part of the documented decision chain. The human override is not a failure of the AI. It is the feature that makes the whole system defensible.
This is the “AI-assisted, human-approved” framework that regulators are increasingly expecting. The NERC wildfire report endorses AI-driven vegetation monitoring but implicitly requires that human judgment be documented at every decision point. Commissioners do not want to see that an algorithm made a choice. They want to see that a qualified person reviewed the algorithm’s recommendation, applied professional judgment, and signed off—with the full context captured in the record.
What This Means for SRP Strategy
Burri closes by arguing that AI’s real power lies in its integration into the organizational fabric of utility operations—not in autonomous prediction, but in partnership with people. I agree completely, and I would add one more dimension: partnership with the regulatory process.
Every AI-driven operational improvement a utility makes is only as valuable as its ability to be documented, defended, and recovered through the rate case. A vegetation management program that uses AI to prioritize high-risk spans, dispatches crews efficiently, and verifies completions in real time is a better program. But it is also a more expensive program. And if the utility cannot demonstrate to regulators exactly how those costs were incurred, why they were necessary, and how they connect to measurable resiliency outcomes, those costs are at risk of disallowance.
The orchestration layer is what closes that loop. It connects the AI output to the field action to the financial record to the regulatory filing. It turns operational improvement into recoverable investment. And it does so not by adding a separate compliance process on top of operations, but by making compliance a natural byproduct of doing the work.
That is the shift utilities need to make. Not from manual to automated. From disconnected to orchestrated.
Let's Talk
If you’re heading into a rate case in the next 18 months, I’d like to show you how two IOUs we’ve worked with turned their vegetation management workflows into audit-ready evidentiary packages—and what their cost recovery outcomes looked like compared to utilities that filed with reconstructed documentation. Schedule a chat with me.