New York orders utilities to inventory AI, buying the map first

Every AI use case in a New York utility will soon have a name on file, along with the policies, procedures and protocols that govern it, under a Public Service Commission order that requires a complete inventory and then tests whether each utility's protocols are adequate and robust against commonly accepted AI governance frameworks. The utilities write the protocols; the commission grades them.

That the order opens by conceding "the actual extent and scope of AI systems in New York utility operations remains unclear" is itself unusual; the commission then enumerates what worries it, from hallucinations and algorithmic bias to transparency problems, data privacy, misconfiguration errors, cyberattacks and "functional brittleness." Its definition of an AI system reaches anything that can make predictions, recommendations or decisions toward human-defined objectives, which covers a great deal of what a modern utility's software already does. PSC Chair Rory Christian called the technology "a double-edged sword" that can deliver cost efficiencies and improved operations while presenting risks that "must be evaluated and addressed appropriately."

Deployments already run ahead of the paperwork, and regulators have named the examples: National Grid's GridCARE system, used to free up interconnection capacity for large-load customers, and the New York Power Authority's use of AI to analyze drone-captured data for vegetation management. Con Edison said in a statement to Utility Dive that it applies AI to improve customer service, identify equipment problems before they reach customers or public safety, and strengthen inspection and mapping. Capacity release and asset health are two of the more consequential places to put a model.

Capacity release matters to investors because grid permission is the underwriting asset, and queue positions and interconnection contracts price before electrons do. GridCARE is the sharpest New York example, a model used to free capacity inside an interconnection queue where the permission to connect has become the binding constraint for developers. An inventory that names it is also an acknowledgement that a learning system sits inside a process the commission reviews for prudence. Once the protocol is on file, it becomes a document the commission can hold the utility to, and probably the hook for whatever performance standard follows.

The disclosure-first sequence matches what California opened in its rate case rewrite this month, where cost transparency and performance-linked compensation share a docket with regulated returns. In New York, the thickness of the filings will say as much about utility self-knowledge as about utility risk: a short list means either the exposure is small or the visibility is. The adequacy review is the expensive half. A protocol judged inadequate is one the utility rewrites, and that rewrite lands on the same operations and compliance staff who assembled the inventory.