New York Ordered Utilities to Inventory Their AI. What Happens When It Changes?
By KEIRSTEN BRAGER
New York now requires 12 named electric, gas, and water utilities to report all artificial intelligence (AI) use cases in their operations and disclose how those uses are governed. Arizona is examining related questions through an open docket. Both proceedings ask utilities to describe their AI use at a point in time, and AI systems do not always stay inside the boundaries documented or approved at that point.
What Regulators Are Asking For
New York: An Inventory Due in 60 Days
On September 17, 2026, the New York State Public Service Commission issued an Order Initiating Proceeding in Case 26-M-0552. It directs 12 named electric, gas, and water utilities, including Con Edison, National Grid, New York State Electric & Gas, and Central Hudson, to submit within 60 days (by November 16, 2026, unless the Commission's Secretary grants an extension) a report covering all AI use cases in their operations.
The reports must describe each AI system, its functional objective, and the level of human oversight. A second filing, due the same day, must disclose the policies, procedures, and protocols that govern each utility's AI use, including "verification methodologies, pre-deployment testing, human oversight protocols, override capabilities, and data privacy practices."
The Commission will review the disclosed governance practices in light of each utility's statutory obligation to provide safe and reliable service. At the session where the order was adopted, Dennis Dibari of the Department of Public Service staff, who presented the order, called it "step zero" for inventorying AI use and said he would propose that the Commission require utilities to update the inventory as new use cases emerge during the proceeding. Chair Rory Christian said utilities will file semiannual reports. The order's own text sets only the two filings due within 60 days.
Arizona: Gathering Input First
On March 24, 2026, Arizona Corporation Commissioner Lea Márquez Peterson opened Docket No. AU-00000A-26-0060 to examine how regulated electric, natural gas, and water utilities use AI in planning, forecasting, storm response, procurement, plant operations, and distribution systems. The docket requests information from utilities, and Arizona Public Service filed a response in April. A Special Open Meeting workshop is scheduled for October 21, 2026, where Commissioners will hear from the Electric Power Research Institute (EPRI), four utilities, and Commission staff.
What the Australian Incidents Show
On June 18, 2026, an OpenAI agent researching public medicine spending during an internal evaluation gained access to non-public files on a Medicare statistics portal administered by Services Australia. Australia's Prime Minister made the incident public on September 24, and OpenAI said its models took actions it did not intend. OpenAI later notified the New South Wales government of a second June incident involving a web application containing historical fire data.
OpenAI operated the models and the evaluation environment, and by its own account it did not become aware of the Medicare activity until August 11, about eight weeks later. A utility running an agent that a vendor controls or updates will usually have less visibility into that agent than its developer has. An inventory filed in November can describe such a system accurately on the day it is filed and miss what the system does the following month.
Keeping the Inventory Current
An AI use-case inventory is accurate until something material changes, such as a model update, new privileges for a service account, a recommendation tool that gains the ability to submit or execute, a new application programming interface (API) connecting the system to operational data, or a human approval step that is removed.
Shared ownership makes these changes harder to detect. When information technology (IT) owns the platform, a business unit owns the use case, operational technology (OT) owns the affected environment, and a vendor controls the model, no single party may be accountable for deciding when the inventory and risk assessment need an update. Without a designated owner for inventory accuracy, material changes may not reach the people responsible for governance or regulatory reporting.
What CIP Programs Already Do
The New York and Arizona proceedings do not create NERC Critical Infrastructure Protection (CIP) requirements. The comparisons below are control analogies, and whether a given AI system falls within CIP scope is a separate determination. They show disciplines NERC-registered entities can use when AI affects Bulk Electric System (BES) Cyber Systems, reliability processes, sensitive operational information, or existing security controls.
| What state oversight examines | Relevant CIP discipline | Existing control practice |
|---|---|---|
| AI use cases and operational scope | CIP-002 | Identify, categorize, and document applicable scope |
| Governance policies and accountability | CIP-003 | Maintain approved security policies and delegated responsibility |
| Human and nonhuman privileges | CIP-004 and CIP-007 | Authorize personnel access, and manage accounts and permissions |
| Connections and access paths | CIP-005 | Control and monitor applicable electronic access |
| Verification and applicable testing | CIP-010 | Maintain baselines and assess applicable changes before deployment |
| Sensitive information handling | CIP-011 | Identify BES Cyber System Information (BCSI) and protect and securely handle it |
| Vendor and product changes | CIP-013 | Address supply-chain risks in procurement and vendor relationships |
CIP-010 ties an approved baseline to a controlled change process, which is the discipline an AI inventory needs once it is filed. It does not apply to every AI system or every model update, and as an earlier post on AI changes outside the change process explains, a vendor-side model update can change how a tool behaves without changing any element of a CIP-010 baseline.
Many AI tools used for planning, forecasting, procurement, and customer operations can sit in corporate IT outside CIP scope. State oversight may reach systems that CIP does not, and those systems may not inherit the same configuration-management discipline.
Which Changes Should Trigger Review?
Reviewing every change would add a heavy administrative load, so material triggers should be defined by changes to a
| Trigger | Change that prompts review |
|---|---|
| Permissions | New roles, scopes, or service accounts |
| Integrations | New systems, APIs, or data feeds |
| Operational authority | Advisory to autonomous action |
| Model behavior | Retraining, new model, or new version |
| Human approvals | Approval steps added or removed |
| Monitoring | Logging or alerting changes |
| Override | Kill switch or fallback changes |
| Data classification | New BCSI or sensitive data in scope |
| Vendor updates | Vendor-pushed model or feature changes |
Questions to Ask Now, by Role
CIP Senior Manager or Delegate
Which AI use cases touch systems or processes inside existing CIP scope?
Could an AI capability change alter a baseline, access path, security control, or BCSI repository?
What evidence would show that an AI-initiated action was approved?
Governance, Risk, and Compliance (GRC) and Regulatory Teams
Who owns ongoing inventory maintenance after the filing?
Which changes require the filing, internal inventory, or risk assessment to be revisited?
Can governance claims be supported by technical evidence?
Capital Allocators
What discovery, logging, identity, monitoring, and evidence-correlation capabilities are missing?
Which vendor contracts must be revised?
Are we funding a one-time inventory or a sustainable control process?
Utility AI Inventory and Change-Control Checklist
For each AI use case, the inventory should record the items in Table 3.
| Area | What the inventory records |
|---|---|
| Use and ownership | Approved use and operating objective. Business and technical owners. |
| Model and vendor | Model, provider, and deployed version. Vendor notice requirements for model or feature changes. |
| Connections and data | Connected systems, APIs, tools, and data feeds. Sensitive-data and BCSI repositories. |
| Authority | Authorized read, recommend, submit, approve, write, and execute actions. Human approval and escalation points. |
| Identity and access | Accounts, roles, service identities, and privileges. |
| Change control | Material changes that trigger review. Tested override, containment, and rollback. |
| Logging and evidence | Logging sufficient to attribute consequential actions. Evidence of approval, testing, deployment, and monitoring. |
| Boundaries | Applicable regulatory and NERC CIP boundaries. |
A Readiness Test
One test of an inventory process is whether the organization could reproduce its AI inventory six months after the first filing and show who approved each material change in between. If New York's semiannual reports follow as the Chair described, the 12 named utilities will face that test at their second filing.
Sources and Further Reading
New York State Public Service Commission, Order Initiating Proceeding, Case 26-M-0552, issued and effective September 17, 2026.
New York State Public Service Commission, September 17, 2026 session transcript, Item 201.
New York State Department of Public Service, Case 26-M-0552 docket.
Arizona Corporation Commission, Commissioner Lea Márquez Peterson Opens Formal Inquiry into Potential Use of Artificial Intelligence in Utilities' Operations, March 27, 2026, announcing the docket opened March 24.
Arizona Corporation Commission, Docket No. AU-00000A-26-0060, eDocket.
Arizona Corporation Commission, Commissioner Lea Márquez Peterson Hosts Workshop on AI in Utility Operations, September 29, 2026.
ABC News (Australia), OpenAI hacked Medicare portal, Prime Minister Anthony Albanese says, September 24, 2026, including the incident timeline and OpenAI's statement.
ABC News (Australia), Rogue OpenAI agent enters another NSW government website, tech giant says, October 2, 2026.
Ampyx Cyber, When AI Changes Without a Change Request, September 25, 2026.
