
The problem
A regulated utility's research was entirely manual. Across the regulatory team, ten or more analysts at any given time would navigate state commission websites, scan thousands of docketed cases to find the right one, open dozens of documents, and read hundreds of pages to find a single answer.
This wasn't occasional work. It was the foundation of every rate case filing, the mechanism the company uses to recover its costs and earn revenue. Slow or incomplete research meant weaker filings, with millions of dollars on the line.
They tried Microsoft Copilot first. It failed in ways that would have been dangerous in front of a commission.
- Wrong jurisdiction. Asked to summarize a case, Copilot called an Idaho proceeding a Washington one and cited the wrong state's statutes.
- Wrong case number. Asked for a company's most recent rate case, it returned an outdated docket. The real filing was months newer.
- Said the data didn't exist when it did. On a security issuance case, it claimed the costs weren't disclosed. They were: $284,031.70, sitting in the compliance filing with page numbers.
- Fabricated rule language. It quoted an administrative rule word for word. The language does not exist. It was invented.
Where the answers come from
Most AI tools answer by skimming the public internet. Most commission filings aren't there. This tool works from the actual filings instead, so every answer traces back to a real document. It reads the case files, cross-references dockets, and shows its work. Copilot answers in seconds because it skims. This takes a few minutes because it reads. That difference is the whole point.




