We asked seven utilities to tell us how AI is impacting rate case discovery for their organizations.
There’s a common term in legal proceedings called bet-the-company litigation. It is what it sounds like: a high-stakes case in which the result can determine the fate of a company. Imagine being involved in bet-the-company litigation, except this lawsuit lasts forever. That’s roughly what regulatory work feels like for a major American utility.
Everything a utility builds, buys, or connects must eventually make its way through a litigation-style regulatory process. Billions of dollars in investments, operating expenses, and infrastructure decisions are scrutinized by regulators and challenged by intervenors. Utilities must explain and defend their choices. The other parties in a rate case issue data requests, which the utility is obliged to answer, under oath, within three to thirty days.
Unlike a bet-the-company lawsuit, which eventually ends, utility regulation never does. After resolving a rate case, it’s usually one to three years until the next one. And other, smaller cases, featuring their own discovery requests, are ongoing almost all the time.
States and the federal government chose to adopt this litigation-style, adversarial process over a century ago, and they had good reason to do so. Most large-scale utilities are private companies, and yet electricity distribution is a natural monopoly. In an industry with no competition, rate cases are a way to balance the interests of customers and investors. Investor-owned utilities make money by earning a regulator-determined rate of return on the infrastructure they build. In rate cases, utility commissions decide what the rate should be and which infrastructure is subject to it.
Discovery serves a valid, necessary purpose in rate cases. It allows intervenors and regulators to collect the information they need to challenge and decide. Data requests can illuminate the topics at issue and help the parties work through their inevitable disagreements about who should be paying for which infrastructure.
How is the proliferation of AI affecting discovery in regulatory proceedings? Last month my HData team convened regulatory leaders from seven of America’s largest utilities for a virtual roundtable. The participants from each utility are responsible for receiving, managing, and answering data requests. We wanted to learn if and how AI is changing the nature of their work, and what they are doing about it. Three major themes emerged.
Data Request Volume is Increasing
One of the first trends our participants discussed is that data requests are multiplying, but the people and resources available to answer them are not. During the roundtable:
- One participant described how a recent rate case involved more than 18,000 data requests.
- Another reported more than 900 questions on an annual proceeding, which is more than double the previous year’s volume.
- A third described a roughly 25% increase between major rate cases.
Combined with short turnaround times, these volumes are punishing.
Depending on the jurisdiction and the proceeding, utilities may have as little as three to five days to respond. Even ten days can be a luxury when answering a single request requires coordinating accountants, engineers, lawyers, economists, and business managers across a sprawling enterprise.
A data request doesn’t simply arrive at someone’s desk and get answered. It must be interpreted, assigned to a subject-matter expert, researched, drafted, reviewed, approved, and delivered. A seemingly straightforward question may require information from several different systems, reconciliation of financial records, or a search through years of previous regulatory proceedings.
Data requests have been part of running a utility for decades. But several of our roundtable participants told us that they are now receiving more requests than ever before, and that the requests are seeking more detail than in the past.
One of our colleagues described a data request that asked her utility to produce every invoice from an enormous infrastructure project. This meant she had to coordinate dozens of people to retrieve documentation from enterprise systems. It took months. Utilities are not small organizations. The participants we spoke with have some of the best legal, financial, and regulatory departments in the industry, and they are struggling to keep up.
Data Request Quality is Getting Worse
One theme surfaced repeatedly in our conversations. AI tools are not just allowing regulators and intervenors to create and send more data requests. They are also making data requests worse. It is easy to see why.
A regulator or intervenor can upload a utility’s filing into a general-purpose AI tool and ask it to generate questions from that filing. Within a few seconds, the tool can produce dozens or hundreds of plausible-sounding requests. Our colleagues said they’re receiving more requests that are ambiguous, repetitive, overbroad, or seemingly disconnected from the issues in the proceeding. Some said they recognized the characteristic language and punctuation of general-purpose AI tools, like ChatGPT, Copilot, and Claude:
- “Right after we filed our rebuttal testimony we got a large number of requests in a very short amount of time. Our rebuttal testimony was over 1,000 pages. To produce those questions quickly leads to the assumption that AI could have been leveraged.”
- “This year we definitely noticed the language in the data requests has changed. We assume that intervenors are utilizing AI a lot now to generate their questions because the language has become very ambiguous to us.”
- “When you’ve seen AI-generated language, you know what it is. We’re seeing em-dashes, we’re seeing bolded language, and then the more ambiguous things.”
- “We’ve taken a topic and asked my Copilot, what questions would you ask? And they were almost word-for-word.”
General-purpose AI tools are great at synthesizing the publicly-available conventional wisdom about something. They are not so great at efficiently retrieving the information you need to participate in a rate case. Yet utilities still have to answer them all.
“It’s easier to generate the questions than the responses. We’re just trying to manage it as best we can when we typically respond, even if we object,” shared one roundtable participant. Another stated that, “It’s easy to produce an AI-generated question but it’s not always easy to answer that with AI. It’s not an even playing field.”
Our colleagues pointed out a fundamental asymmetry: generating a hundred data requests is now almost costless. A general-purpose AI tool can invent questions without understanding the underlying utility operations, accounting conventions, or regulatory history. Producing a hundred accurate and defensible answers is expensive.
AI-Generated Data Requests are Producing Less Useful Answers
Perhaps the most revealing moments in our roundtable came when our utility colleagues discussed what actually happens to the answers they laboriously produce. One utility representative estimated that, out of hundreds of responses, only a small fraction might ultimately be cited in testimony or used during hearings. Another agreed that relatively few responses to intervenors actually find their way into those intervenors’ testimony.
Rate case discovery can serve legitimate purposes beyond informing testimony. It can test assumptions, verify claims, and establish confidence in the evidence. But this does raise an important question. If a utility spends thousands of hours responding to discovery, how much of that effort shapes the case by making its way into the questioners’ testimony later on in the case? How much of that effort ends up improving the quality of the regulator’s ultimate decision?
When data requests and the information produced in response to data requests can help a regulator make informed, better choices, they are constructive. But every hour spent tracking down an unnecessary invoice, deciphering a muddled AI-generated question, or replying to a duplicative AI-generated request is an hour that cannot be spent analyzing the merits of an intervenor’s argument. Who pays for all those hours? Ratepayers. We shouldn’t confuse volume of discovery with the quality of discovery.
Of course, commissions and intervenors in rate cases start out at a disadvantage to the utilities they’re querying. The utility runs the grid, so the utility holds the information needed to inform a rate case decision. The utility collects the rates and typically has greater resources to litigate a rate case. Technology and AI tools can help the regulator and the intervenors participate more meaningfully. But when the AI tools generate less-useful data requests, they also lead to less-useful answers. If an intervenor doesn’t get meaningful information to inform its case and cite in its briefs, then both sides have wasted their time and money.
What Can Utilities Actually Do About (More and Worse) AI-Generated Data Requests?
It is possible to relieve the reporting burden of utility regulatory teams without sacrificing the right of intervenors to receive the information they need to make their case. It is possible to reduce the volume of duplicative or overly broad data requests while still improving the quality of decisions by regulators. The answer is not more AI.
About HData
As the operating system for energy regulation, HData serves the largest customer ecosystem in regulated energy, helping enterprises navigate regulatory complexity. Through centralized data, powerful AI, and workflow automation, HData accelerates the research, analysis, and insights critical to deciding the future of energy. Learn more at hdata.com.
