Inaugural issue. The research log behind this digest started on Tuesday September 16, so this week’s window — Monday September 14 through Friday September 18 — is thinner than it will be once the daily log has run a full cycle. Thirteen in-window items made the cut. Everything the log currently holds from before Monday — OATI’s September 9 Utility Dive webinar with CAISO, the August 19 five-agency CISA advisory on AI-generated PLC exploitation scripts, OpenAI’s August 27 “collective cyber defense” open letter, Google’s September 2 Fairwind Program, Axios’s September 3 report on OpenAI’s Daybreak pitch at EEI, Anthropic’s September 10 threat-intelligence report, the September 12 Amodei essay, Onapsis’s September 1 SAP shared-responsibility piece and Deloitte’s April survey of 60 IOUs — predates the window and is held for context only, not re-reported. The standing anchor facts (OATI AI Genie in daily production at CAISO; Daybreak; the Amodei/Huang split) are referenced below only where they change the reading of an in-window item.
Two threads run through the week, and they pull in opposite directions.
The first is that the institutions that govern the North American grid spent this week talking about AI as a load problem, not a control-room problem. NERC’s CEO published an op-ed on Tuesday whose entire AI content is about data centers straining the system — not a word about AI inside operations. The House passed a data-center cost-allocation bill 417–3 on Wednesday. A coalition of the largest AI companies and four major utilities launched the same day to make data-center load flexible. Meanwhile PJM is one month out from an executive summit whose premise is the exact opposite: that AI belongs in the RTO. Both framings are live at the same institution-adjacent level, and a utility writing a 2027 AI roadmap has to decide which conversation it is actually in.
The second is that this was a good week for autonomy that inspects and a quiet week for autonomy that decides. A Scottish transmission owner moved a robot fleet from trial into routine service; Itron announced an agent-assisted meter-data platform that ships in sixteen weeks’ time; a Houston utility filed an 8-K over an external API that apparently handed out millions of customer records. None of it involves an agent touching control logic. That gap — between AI that looks and AI that acts — is the one this digest exists to track, and this week it did not close.
And then, on Thursday, a third thing happened that outranks both. The New York Public Service Commission voted to order every investor-owned utility in the state to inventory all of its AI systems within 60 days and to report semiannually thereafter. That is not a panel, a pledge or a pilot — it is a filing obligation with a clock on it, and it is the first real regulatory precedent for utility agent governance in the United States. Every utility outside New York should read the order as a template to prepare against rather than a local matter.
🤖 Control-Room & Operational AI
PJM’s inaugural GridAdvance Summit (October 15) is the first executive-level, on-the-record forum in which a U.S. RTO frames AI as a control-room question rather than a planning or back-office one. PJM’s September 15 Inside Lines reminder confirms the program: a keynote from PJM board member Le Xie (co-founder of Harvard’s Power and AI Initiative), a COO fireside chat pairing PJM’s Stu Bresler with NYISO’s Emilie Nelson, and three panels — “From Queue to Connection: Rethinking Planning With AI,” “Innovation vs. Risk: Deploying AI in a Reliability-First Environment,” and “Control Room 2030: Moving From Situational Awareness to Situational Intelligence.”
The participant list is the substantive part: SPP, MISO, ERCOT, CAISO, NERC, EPRI, Dominion, Constellation, Tapestry, GE Vernova and Siemens. CAISO is the one ISO with a documented production agentic deployment (OATI AI Genie triaging 500+ next-day outage requests nightly, operators retaining final authority). NERC is the entity whose view of AI inside CIP-scoped environments has not yet been stated publicly. GE Vernova and Siemens are the two EMS/ADMS incumbents whose roadmaps determine whether control-room AI arrives as a bolt-on or as a native platform feature.
For a utility building an internal AI roadmap, the summit’s three-panel structure — planning, reliability-risk, control room — is a usable template, and “Innovation vs. Risk” is the panel to watch for the first public articulation of RTO agent-autonomy guardrails. Registration closes October 1; in-person at PJM’s Conference & Training Center, and virtual. (Source: PJM Inside Lines, September 15)
The counter-signal came the same week from NERC’s own CEO, and it is worth reading the two side by side. Jim Robb’s September 15 Utility Dive op-ed, “Change 4 mindsets to build a more resilient US grid,” is the reliability organization’s chief executive writing at length about AI — and treating it exclusively as a demand-side phenomenon. His four mindset shifts are siting and permitting, load as an active system participant, technology-mix diversity, and weighing societal benefit against cost. The AI content is entirely about data-center load: NERC is “working at breakneck pace” on criteria for identifying the most consequential large computational loads and the performance requirements that go with them, and EPRI is developing a standardized load-flexibility framework called Flex MOSAIC to bring consistency to curtailing large loads when reliability requires it.
What is absent is the tell. The CEO of the ERO wrote a full-length AI piece with no mention of AI in operations, no mention of agent autonomy, and no mention of what AI inside a CIP-scoped environment would require. That is not an oversight so much as an accurate reflection of where NERC’s institutional attention sits: AI is a load to be managed, not a tool to be governed. Utilities should read that as the current regulatory baseline — and read PJM’s “Innovation vs. Risk” panel as the venue where it might first shift. (Source: Utility Dive opinion, September 15)
The week’s one vendor announcement that puts agents inside a utility system of record came from Itron on Wednesday, and it is a 2027 product. Itron introduced IEE Cloud AI on September 16 — the next generation of Itron Enterprise Edition Meter Data Management, rebuilt cloud-native and integrated with Itron’s Intelligent Edge Operating System, with what the company describes as “AI-trained, natural-language and agent-assisted workflows.” The operational claim is a real one and worth reading carefully: utilities can choose when data loads, validate readings as they arrive, and clear exceptions throughout the day rather than depending on an overnight processing window — near-real-time ingestion with dynamic scaling, in place of the batch cycle that has defined MDM for two decades. Named grid-edge use cases are outage insights, grid visibility, forecasting, anomaly detection and revenue assurance. Itron Managed Services is offered as an operating model in which Itron runs the platform, ships AI enhancements and handles security patching on the utility’s behalf.
Two qualifications belong on this before it enters anyone’s roadmap. First, general availability is January 2027 — this is a forecast item, not a deployment, and no reference customer is named. Second, it is worth being precise about where the agents sit: MDM is a billing-and-data-foundation system, not a control room. An agent clearing VEE exceptions or flagging a meter anomaly is nowhere near dispatch or switching, which is exactly why it is a plausible early production target. The more interesting question the announcement raises is the managed-services one — a vendor continuously shipping AI changes into a system that feeds revenue assurance and outage analytics is a CIP-013 supply-chain and CIP-010 change-management conversation, and “ongoing AI-enabled enhancements” as a service term needs a change-control answer before signature. (Source: Itron press release via GlobeNewswire, September 16; T&D World)
🏭 In Production vs. 🔮 Forecast / Marketing
In production this week: one, and it is robotics rather than agents. SSEN Transmission has deployed seven new inspection robots across its HVDC estate in northern Scotland — at the Blackhillock, Spittal and Kergord converter stations and the Noss Head DC switching station at Wick — bringing its fleet to eight and explicitly marking, in the company’s own words, “a move from innovation trials towards routine use of robotics across its network.” The machines carry out autonomous inspections and monitor electrical equipment, feeding asset-condition data to operational teams; they are expected to complete more than 5,000 autonomous inspections and travel over 1,700 km in the next year. They have captured thermal and ultraviolet imagery from energised HVDC halls — areas engineers cannot safely enter while assets are in service — helping identify potential faults earlier. The deployment follows three years of work with Ross Robotics.
Label this precisely: this is autonomous inspection and data capture, not an agent making or recommending operational decisions. No control logic is touched, no dispatch or switching recommendation is generated, and the AI content is in the perception and anomaly-detection layer rather than in any reasoning loop. That said, it is the only item this week that crossed the trial-to-routine line with a stated volume commitment attached, and the pattern — autonomy earning its place first in the observation layer, where a wrong answer costs an unnecessary truck roll rather than an outage — is the one most utilities will follow. (Source: Energy Live News, September 16)
Forecast / agenda this week, in descending order of how checkable the claim is:
- Itron IEE Cloud AI — a named product with named capabilities and a stated GA date (January 2027), no reference customer. The most substantive of the four because there is a shipping predecessor (IEE MDM) and a defined delivery date to hold it to.
- PJM GridAdvance Summit (October 15) — agenda-setting; PJM has announced no production AI of its own.
- AI Energy Management Alliance (September 16) — a coalition launch with a load-flexibility thesis and no deployed capability attached to the announcement itself. See AI-Provider Posture.
- The UK’s “AI in Energy Strategy” — a plan to publish a plan next year.
Kilowatt Intelligence Issue #18 (September 15) is an aggregation of prior-week items — Daybreak, the Amodei/Huang split, the OATI/CAISO proof point — rather than a source of new deployment evidence.
The week’s best number on the pilot-to-production gap came from National Grid Partners on Friday morning. NGP’s third annual Utility Innovation Survey — 134 U.S. utility innovation leaders, fielded May 20–July 7, released September 18 at the NextGrid Alliance Summit in Boston — reports that 78% are deploying or operationalizing at least one AI application to manage interconnection demand, while 74% say AI-driven data-center load growth is itself degrading reliability. That circularity is the industry’s position in one line: AI is both the load problem and the tool being bought to manage it. The figure to carry into every vendor conversation, though, is the friction one: 84% now take more than a year to move a project from pilot to full rollout, up from 78% in 2025 and 66% in 2024 — getting slower, even as 72% report larger innovation budgets. The priority inversion is the strategy-deck number (reliability in the top three for 73%, against 43% a year ago; net zero 54% → 16%), and the constraint numbers are the regulator’s (87% say rate-case frameworks written before the AI demand boom cap the return on innovation; 83% say data-center costs are flowing to residential bills; only 16% of innovation spend goes to transformational work). Two caveats keep this a benchmark rather than evidence: it is a corporate-VC publication issued alongside two of its own portfolio investments (Terragrit, LineVision), and “deploying or operationalizing” is a single bucket, so it is not a production count. Read against Itron’s January 2027 GA date, it says the honest answer to “when does this reach operations” is later than the slide claims. (Source: National Grid Partners via PRNewswire, September 18)
Scoreboard for the inaugural week: 1 production (inspection robotics, transmission asset monitoring, SSEN), 0 new production agentic-AI deployments in operations, 4 forecast/agenda. The standing production benchmark for agentic control-room AI remains OATI AI Genie at CAISO, which predates this window. This digest will keep the count explicit each week so the ratio is visible over time.
🛡️ AI & Grid Cybersecurity
CenterPoint Energy filed a Form 8-K on September 14 confirming that an unauthorized third party obtained personal information for “a portion” of its customers through one of the company’s externally accessible systems. The filing followed a forum post by a threat actor claiming roughly 7.49 million customer records pulled from a CenterPoint-controlled API that allegedly lacked adequate authentication and rate limiting; the claimed fields include names, phone numbers, email and service addresses, account and premise identifiers, billing amounts, payment information, autopay status and the last four digits of Social Security numbers. CenterPoint has not confirmed the record count, the API description or the authenticity of the published dataset — those remain the threat actor’s claims. The company activated its incident-response protocols, engaged external cybersecurity specialists, notified law enforcement and certain regulators, and stated that electricity and gas delivery were not affected. It expects incident-response costs, partially offset by cyber insurance, and does not currently expect a material financial impact.
There is no AI in the reported facts of this incident, and it should not be dressed up as an AI story. It belongs here for a narrower reason: an under-authenticated, un-rate-limited, internet-reachable API holding millions of customer records is the precise target class that the past month’s AI-threat literature says gets cheaper to find and drain at scale — the mechanism being a drop in the cost and skill per campaign rather than any new technique. The actionable read is an inventory of externally reachable endpoints with authentication, rate limiting and data minimization checked per endpoint. That is a non-AI control that the AI-threat picture directly motivates. (Sources: CenterPoint 8-K, September 14; Reuters, September 14; CyberInsider, September 15)
The week’s second security item is a deadline, not an incident. DOE’s Office of Cybersecurity, Energy Security, and Emergency Response (CESER) held its industry-engagement webinar on Wednesday September 16, 3–4 p.m. EDT, walking through the Request for Information it published in the Federal Register on September 9 under Executive Order 14421, “Securing the United States Bulk-Power System” (signed August 26). The RFI seeks input across eight areas — how to define “bulk-power system electric equipment” and which transaction types, equipment classes and risk thresholds to prioritize; how to identify covered foreign entities; supply-chain governance practices; methodologies for evaluating cybersecurity, national-security and supply-chain risk; inventories of existing installed equipment; and candidate mitigation measures. Written responses are due October 9, 2026.
Nothing in the RFI mentions AI. That is the reason to file. The definitions being settled now — what counts as covered equipment, what supply-chain governance evidence looks like, what methodology establishes risk — are the definitions that AI-bearing equipment, vendor-hosted inference and continuously-updated managed AI services will be measured against later, whether or not anyone intended that. Itron’s managed-services model in the section above is a concrete example of the class of arrangement these definitions will eventually have to address. (Source: DOE/CESER, September 16; Federal Register 91 FR 57322, September 9)
Context worth carrying alongside both (published pre-window, August 27 and updated September 3, so not counted as a week item): DOE/CESER and Sandia National Laboratories have fielded C2E2 under CESER’s AI-FORTS program — LLM and generative AI that compresses grid-security data engineering from two months to a few hours and locates the source of a cyber threat with a stated 95% accuracy in near real time. The candid part is what DOE volunteers about the next research phase: it targets hallucination, because “for critical infrastructure, a hallucination could mean missing a real cyberattack or falsely flagging a non-existent threat.” That is a federal lab naming, in its own promotional copy, the same confident-but-wrong failure mode the New York PSC named as “a new risk category” this week. Treat 95% as a research figure and not an SLA — there is no commercial availability or partner list — but note that the data-engineering compression is the more transferable result, and that DOE reaches small and public-power utilities through a channel no gated frontier-lab program does. (Source: DOE/CESER, August 27, updated September 3; Sandia Lab News, July 30)
🏛️ AI-Provider Posture Toward Critical Infrastructure
The in-window primary source is a coalition launch, and its membership is the story. On September 16 Emerald AI, Google and NVIDIA launched the AI Energy Management Alliance (AEMA), a coalition to advance data centers that dynamically adjust electricity draw in response to grid conditions — shifting workloads, discharging storage, or running on-site generation, turning a large static load into a controllable resource. Eighteen launch partners joined the three founders, including Anthropic and the utilities AES, Constellation, National Grid and NRG Energy; GridUnity, which builds interconnection and planning workflow software, joined as a founding board member. The stated purpose is to accelerate interconnection of flexible data centers, strengthen reliability and protect affordability.
The load-flexibility substance of this item belongs to the DER/demand-response research log and is reported there, not here. It earns a place in this digest on one narrow ground: it is the first in-window instance of the frontier labs engaging the utility sector as a counterparty on grid operations rather than as a cyber-defense customer. That is a different posture from the three programs on file from the preceding weeks — OpenAI’s $1B Daybreak subsidy pitched to utility CEOs at EEI, Google’s access-gated Fairwind program routed through security partners, Anthropic’s tiered enterprise safeguards. Those three sell into utility security. AEMA asks utilities to interconnect AI infrastructure on flexibility terms. A utility now faces the same three companies in two unrelated procurement conversations, and the negotiating leverage in one is not the leverage in the other. Worth knowing which table you are at. (Sources: NVIDIA blog and Fortune, September 16; TechCrunch, September 17)
Worth restating as the open question this section exists to track: the two AI vendors selling hardest into the utility sector are giving opposite guidance on agent autonomy — Amodei warning about rogue-agent proliferation on a 6–12 month horizon (pre-window, September 12), NVIDIA’s Huang forecasting utilities running hundreds of thousands to millions of continuous agents. Both are now in the same alliance with four large utilities. A utility drafting agent-autonomy policy is doing so against a forecast its own suppliers cannot agree on, from inside a coalition that contains both of them.
📜 Regulatory, Standards & Policy
The most consequential regulatory item of the week — and the first real U.S. precedent for utility agent governance — is a New York commission vote. On Thursday, September 17, the New York Public Service Commission voted to initiate an inquiry ordering the state’s investor-owned utilities, National Grid among them, to file an inventory of every AI system used in their operations within 60 days, and to file semiannual reports thereafter. The order also reaches the policies, procedures and protocols governing the technology, and asks utilities to distinguish the type of AI in use — machine learning versus generative. Officials describe it as one of the first such efforts in the nation.
The framing on both sides is worth quoting. PSC chairman Rory Christian: some New York utilities are “already using AI, and in some instances, it may result in net benefits to New Yorkers,” but “the potential upside does not come without risk.” Department of Public Service staff member Dennis DiBari named the thing this digest exists to track: “the possibility of a potential failure or misconfiguration represents a new risk category for the Commission to evaluate and consider.” The listed uses are the unglamorous ones that dominate real utility AI today — facility inspections, outage modeling, customer service — and the listed risks are privacy, model misconfiguration and fabricated output. National Grid’s response, via spokesperson Patrick Stella, was that it “is reviewing the … inquiry and will provide the inventory requested.”
Two things make this matter well beyond New York. First, a state commission has established that an AI inventory is a regulatory artifact — which is precisely the asset needed to answer the first governance question anyone will ask (who authorized this agent to exist), and precisely what most utilities do not currently maintain. Second, the PSC says it will use what it collects to decide what future regulation is needed, so the inventories filed over the next 60 days will shape the rules that follow. A filing that covers IT and customer-facing AI but goes silent on anything touching OMS or ADMS will read very differently to a commission than it does to a CIO. Build the inventory before another commission asks for it. (Sources: WAMC / New York Public News Network, September 17; Bloomberg Government)
The UK became the first major market this week to commit to a named, standalone AI-in-energy policy instrument. On September 14 the Minister for Local Energy and Jobs, Martin McCluskey, confirmed that an “AI in Energy Strategy” will be published next year. It follows an independent review led by Lucy Yu, the UK Government’s AI Champion for Clean Energy, which recommends moving the grid toward AI-enabled, risk-based operation and planning and — the more operationally interesting half — establishing clear governance for AI systems specifically in order to speed up deployment of AI solutions that have already been proven to work.
That framing is worth lifting regardless of jurisdiction: the review treats governance as an accelerant rather than a brake, on the theory that the absence of an agreed assurance route is what strands proven tools in pilot purgatory. Any utility that has watched a working model sit unblessed for eighteen months will recognize the diagnosis. The review also warns explicitly that greater AI use changes the electricity network’s attack surfaces and the ways threats propagate, and that security has to be designed in at deployment rather than bolted on — which is the same conclusion the U.S. federal advisories reached from the incident side, arrived at from the policy side.
For U.S. readers the contrast with the NERC op-ed above is stark: the UK is writing an operations-AI governance strategy while the U.S. ERO is writing about computational load. Neither is wrong; they are answering different questions. But the UK document, when it lands next year, will be the first published template for how a grid operator proves an AI system is fit for operational use — and that template will get read on this side of the Atlantic. (Source: Energy Live News, September 14)
On the U.S. side, the in-window legislative action was about who pays for AI, not about how AI is governed. On September 16 the House passed the bipartisan Ratepayer Protection Act 417–3 under suspension of the rules, sponsored by Reps. Kathy Castor (D-FL) and Gabe Evans (R-CO). The bill requires states to consider adopting standards that would shift the electricity cost increases associated with large data centers onto the tech companies driving them — a PURPA-style consider-and-determine obligation, not a mandate to regulate. The margin is the signal: near-unanimous, on a fast-track calendar, on data-center cost allocation. Congress has found the one AI-and-electricity question it can agree on, and it is a rate question. Nothing in the bill touches AI in operations, agent autonomy, or AI-related cyber scope. Utilities should expect cost allocation to keep absorbing the available legislative oxygen through 2027. (Source: House floor action, September 16)
Also in the policy pipeline (pre-window, tracked for continuity, not re-reported): FERC’s July 16 directive creating the Computational Load Entity as a new registered-entity class, with NERC’s mandatory standards for computational loads due December 31, 2026 and a Phase II work plan due March 1, 2027; FERC’s August 10 approval of CIP-015-2 (internal network security monitoring); and EPRI’s Flex MOSAIC load-flexibility framework, referenced this week in Robb’s op-ed. All load-side or monitoring-side; all set the vocabulary AI-related grid regulation will inherit.
🔍 Vendor Watch
OATI — no new product or deployment announcement inside the window. The most recent items on OATI’s news page are the September 3 FERC 881 regional-deployment release and the September 9 Utility Dive webinar with CAISO (“How to Move AI from Pilot to the Utility Control Room”), both pre-window and both already in the research log. The webinar’s production-readiness checklist — legacy-system integration, operational-data grounding, constraint awareness, sensitive-data protection, evidence-cited recommendations, retained human authority — remains the most procurement-ready artifact the vendor has published; the companion “AI in the Control Room” whitepaper is the next pull for the vendor file. OATI, Siemens and GE Vernova will share the PJM GridAdvance stage on October 15, which is the next scheduled opportunity for a public statement.
MCG Energy Solutions — no AI announcement inside the window. Versify OMS / IAM / Control Area Scheduling remain the adjacent incumbent watch-item: the structured, document-heavy outage-scheduling and market-submission workflows that OATI is compressing with agents at CAISO are MCG’s core franchise, and whether MCG ships production AI in Versify — and on what human-in-the-loop terms — is the open question. Still not a shipped-AI proof point.
Itron — new to the vendor file this week, and now the only vendor in it with a dated AI product commitment (IEE Cloud AI, GA January 2027). Two things to track: whether a named reference utility appears before GA, and what the Managed Services change-control terms actually say about “ongoing AI-enabled enhancements” landing in a revenue-assurance system. Itron’s prior AI collaboration with Microsoft is the relevant context for where the model layer comes from.
GE Vernova / Siemens — no in-window announcements; both confirmed at PJM GridAdvance.
Ross Robotics — new to the vendor file this week on the strength of the SSEN deployment (see Production). A three-year development partnership that ended in a fleet of eight units in routine service across four HVDC sites is a more credible reference than most AI-in-utilities announcements carry, and the reference is checkable with the asset owner.
Emerald AI / GridUnity — new names on the interconnection-and-flexibility side via AEMA, tracked primarily in the DER log. Noted here because GridUnity’s interconnection workflow software is the kind of process-heavy, document-driven system where agentic automation lands first, on the same logic as OATI’s outage-request triage.
SAP (AI Agent Hub) — a tracked watch-item moved inside the window, quietly and without an announcement. As of September 18 SAP’s AI Agent Hub product page lists all six capabilities with no “planned” or “coming in Q3” qualifier — AI registry, evaluation/verification workflow, identity and access control, AI observability, agent mining, and SuccessFactors org-and-skills mapping — where the position described across the ecosystem in May was two of six generally available and four scheduled for Q3 2026. The line that matters for agent governance is the access-control one: “only verified MCP servers can be called in production workflows, and when verification status changes, access is revoked” — the Verification Seal described as runtime-enforced, and the nearest thing SAP has published to an agent kill switch. Discovery is claimed across SAP, Microsoft, Google, AWS, Databricks and ServiceNow. The caveats are heavy enough to keep this in the GA-claimed column rather than shipping: a marketing page is not a dated release note, SAP News Center carried no September GA announcement, no customer is named, nothing states what happens to in-flight agent sessions when access is revoked, and there is no NERC CIP mapping and no answer to where agent authority stops at the OT (ADMS/SCADA/OMS) line. Still, agent identity via SAP Cloud Identity Services plus verification-driven revocation is recognizably the CIP-004/CIP-005 shape a utility security team already knows how to audit — and an agent registry with risk ratings is exactly the artifact New York’s utilities will wish they had when that 60-day clock runs out.
Onapsis — unmoved. The Agentic Gateway / MCP Gateway for SAP Security remains at its March 17 preview, not GA, with no named customer and unstated scope; nothing new on the blog or press page since the September 8 patch-day cluster. The Onapsis Research Labs “SAPMAP” advisory is scheduled for September 21 — next week’s item.
Aggregator note: Kilowatt Intelligence Issue #18 (Todd Durocher, September 15) was the in-window vehicle that surfaced the OATI/CAISO, Daybreak and Amodei/Huang items to this log. It is cited as a secondary source; every item it carried was traced to a primary source in the daily log.
⚡ Utility-Sector Relevance Flags
- Build the AI inventory now, before a commission asks. New York’s order gives its utilities 60 days to produce a complete inventory of AI systems, with the type of each (ML vs. generative) and the policies governing it, and semiannual reporting after that. Every element of that filing is something a utility should be able to produce for its own governance regardless of jurisdiction — and the ones who start after the order arrives will be reconstructing it from memory under a deadline. Note also what a good filing has to include: anything AI-adjacent touching OMS, ADMS or work management, not just the customer-facing chatbot.
- Assume “a new risk category” language spreads. New York’s staff framed failure or misconfiguration — not misuse, not bias — as the novel regulatory exposure, and DOE’s own grid-security researchers named hallucination as their open problem in the same week. That is the failure mode regulators are converging on, and it is the one an AI governance program should be able to demonstrate controls for: what happens when the model is confidently wrong, who notices, and how fast.
- Hold vendor timelines against the industry’s own number: 84% take more than a year from pilot to full rollout, and that figure has gotten worse three years running. Any AI roadmap promising operational impact inside twelve months is betting against the modal outcome in its own sector.
- The governing institutions are having two different conversations about AI. NERC’s CEO wrote about AI as load; the House voted 417–3 on who pays for AI’s load; a lab-and-utility alliance launched to make that load flexible; PJM is convening on AI as an operating tool. All in the same week. Before a utility’s AI roadmap goes to its board, decide which conversation the roadmap is in — because the reliability-compliance questions, the rate questions and the operational-benefit questions have almost no overlap and very different owners.
- New-production count: one, and it is an inspection robot. No agentic-AI deployment entered operations this week. Keep the vendor-claim discount high: the only production agentic deployment on file remains a single bounded use case at a single ISO. The SSEN pattern — autonomy proving itself in observation before decision — is the realistic sequencing for most utilities, and it is a defensible answer to a board asking why the agent pilot hasn’t shipped.
- “Agent-assisted” is arriving in systems of record before it arrives in control rooms. Itron’s announcement puts agents in MDM — billing, VEE exceptions, revenue assurance — sixteen weeks out. That is the likely shape of most utilities’ first production agent: adjacent to operations, inside a system with an audit trail and a human approver, where a wrong answer is a re-read rather than an outage. Plan the governance for that case, not the control-room case, first.
- A managed AI service is a change-management question. “Ongoing AI-enabled enhancements” delivered continuously by a vendor into a system that feeds outage analytics and revenue assurance needs a CIP-010 configuration-change-management and CIP-013 supply-chain answer before signature — not after. Ask for the change-notification and rollback terms in writing during evaluation.
- October 9 is a live deadline. DOE’s EO 14421 RFI closes then. The definitions being set — covered equipment, supply-chain governance evidence, risk-evaluation methodology — will frame how AI-bearing equipment and vendor-hosted inference get treated later, even though the RFI says nothing about AI. Filing is cheap; inheriting someone else’s definition is not.
- External-facing APIs are now a board-level exposure. CenterPoint’s 8-K is a reminder that customer-facing IT surfaces — billing APIs, portals, outage maps — sit outside NERC CIP scope and are exactly where AI-accelerated attackers look first. An inventory of externally reachable APIs, with authentication, rate limiting and data minimization checked per endpoint, is a concrete non-AI action that the AI-threat literature of the past month directly motivates.
- Governance as accelerant, not brake. The UK review’s core claim — that a clear assurance route makes proven AI deploy faster — is the most useful argument in this week’s material for anyone inside a utility trying to unstick a model that works but has no one empowered to approve it. Worth borrowing the framing ahead of the strategy’s publication next year.
- Know which table you are at. The same three AI companies are now selling utilities cyber defense in one conversation and asking for interconnection on flexibility terms in another. Leverage earned in one does not transfer. Make sure the two negotiations know about each other internally, because the counterparties certainly do.
- Contract language to lift. The frontier labs’ own commitment to “ensure agentic identities are traceable and accountable” (pre-window, now the operative norm) is language a utility can require of any vendor whose agent touches OMS, EMS, MDM or market systems. Pair it with the OATI/CAISO checklist as RFP evaluation criteria.
📚 Sources
Primary (in-window, September 14–18, 2026):
– PJM Inside Lines — Reminder: PJM GridAdvance Summit One Month Away (September 15) · https://insidelines.pjm.com/reminder-pjm-gridadvance-summit-one-month-away/
– Jim Robb (NERC President & CEO) — Change 4 mindsets to build a more resilient US grid, Utility Dive opinion (September 15) · https://www.utilitydive.com/news/grid-infrastructure-transmission-data-center-ai-nerc-robb/829592/
– Itron — Itron Introduces IEE Cloud AI, Enabling Grid Edge Use Cases from a Trusted Data Foundation (GlobeNewswire, September 16) · https://www.globenewswire.com/news-release/2026/09/16/3363187/0/en/itron-introduces-iee-cloud-ai-enabling-grid-edge-use-cases-from-a-trusted-data-foundation.html
– T&D World — Itron Launches Cloud Computing Platform, Boosting Utility Meter Data Management · https://www.tdworld.com/smart-utility/article/21283110/itron-launches-cloud-computing-platform-boosting-utility-meter-data-management
– Energy Live News — Robots take over Scottish network monitoring (SSEN Transmission / Ross Robotics, September 16) · https://www.energylivenews.com/2026/09/16/robots-take-over-scottish-network-monitoring/
– Energy Live News — ‘AI in energy strategy’ plan released next year (Lucy Yu independent review; Minister McCluskey, September 14) · https://www.energylivenews.com/2026/09/14/ai-in-energy-strategy-plan-released-next-year/
– NVIDIA — Emerald AI, Google and NVIDIA Launch Alliance to Advance Flexible AI Data Centers (AEMA launch, September 16) · https://blogs.nvidia.com/blog/ai-energy-management-alliance/
– Fortune — Data centers can be good citizens. It’s why we’re partnering with Google and Nvidia to create the AI Energy Management Alliance (September 16) · https://fortune.com/2026/09/16/data-centers-ai-energy-management-alliance-emerald-google-nvidia-anthropic/
– TechCrunch — Google, Nvidia, and Anthropic want Emerald AI to find space on the grid for more data centers (September 17) · https://techcrunch.com/2026/09/17/google-nvidia-and-anthropic-want-emerald-ai-to-find-space-on-the-grid-for-more-data-centers/
– CenterPoint Energy — Form 8-K, U.S. Securities and Exchange Commission (September 14) · https://www.sec.gov/Archives/edgar/data/1130310/000110465926107560/tm2625326d1_8k.htm
– Reuters — CenterPoint Energy discloses customer data breach in SEC filing (September 14) · https://www.reuters.com/legal/litigation/centerpoint-energy-discloses-customer-data-breach-sec-filing-2026-09-14/
– CyberInsider — CenterPoint Energy confirms data breach after hacker claims 7.49M records (September 15) · https://cyberinsider.com/centerpoint-energy-confirms-data-breach-after-hacker-claims-7-49m-records/
– WAMC / New York Public News Network — New York to begin monitoring utilities’ use of AI (Samuel King; September 17) · https://www.wamc.org/2026-09-17/new-york-to-begin-monitoring-utilities-use-of-ai
– Bloomberg Government — Cyber Fears Drive New York’s Push to Audit Utilities’ Use of AI · https://news.bgov.com/bloomberg-government-news/cyber-fears-drive-new-yorks-push-to-audit-utilities-use-of-ai
– National Grid Partners — 2026 Utility Innovation Survey: Industry leaders turning more to AI as data-center boom reshapes grid planning (PRNewswire, September 18) · https://www.prnewswire.com/news-releases/2026-utility-innovation-survey-industry-leaders-turning-more-to-ai-as-data-center-boom-reshapes-grid-planning-302883063.html
– SAP — SAP AI Agent Hub product page (capabilities as listed September 18) · https://www.sap.com/products/artificial-intelligence/ai-agent-hub.html
– U.S. DOE / CESER — industry-engagement webinar on Executive Order 14421 and the associated RFI (September 16) · https://www.energy.gov/ceser/articles/ceser-holds-industry-engagement-webinar-covering-recent-bulk-power-system-executive
– U.S. House of Representatives — passage of the Ratepayer Protection Act, 417–3, under suspension of the rules (September 16); sponsors Rep. Kathy Castor (D-FL) and Rep. Gabe Evans (R-CO)
Supporting primary (published September 9, discussed in-window; comments due October 9):
– Federal Register — Securing the United States Bulk-Power System, 91 FR 57322, DOE Request for Information under EO 14421 (September 9) · https://www.federalregister.gov/documents/2026/09/09/2026-18370/securing-the-united-states-bulk-power-system
Secondary / aggregator (in-window):
– Kilowatt Intelligence Issue #18 — Todd Durocher (September 15) · https://www.linkedin.com/pulse/kilowatt-intelligence-issue-18-september-15-2026-todd-durocher-n8djc/
Context only (pre-window; not re-reported):
– OATI — How to Move AI from Pilot to the Utility Control Room (Utility Dive webinar, September 9) · https://www.oati.com/events/ai-pilot-to-utility-control-room/
– OATI — Under the hood of CAISO’s AI outage management pilot · https://www.oati.com/news/under-the-hood-of-caisos-ai-outage-management-pilot/
– Anthropic — Detecting and countering misuse of AI: September 2026 (threat-intelligence report with IOCs, September 10) · https://www.anthropic.com/threat-intelligence-report-september-2026
– Google — Proactive cyber defense for governments and enterprises: the Fairwind Program (September 2) · https://blog.google/innovation-and-ai/technology/safety-security/fairwind-program/
– CISA — Advisory AA26-231A, Defending Against an Active Threat to Siemens S7 Series PLCs (August 19) · https://www.cisa.gov/news-events/cybersecurity-advisories/aa26-231a
– OpenAI — A call for collective action on cyber defense (August 27) · https://openai.com/collective-cyberdefense/
– Axios — OpenAI unveils plan to protect critical services from AI cyberattacks (September 3) · https://www.axios.com/2026/09/03/openai-critical-infrastructure-cyber-ai-models
– Dario Amodei — We Must Pace the Frontier (September 12) · https://darioamodei.com/post/we-must-pace-the-frontier
– Onapsis — The SAP Shared Responsibility Model: Who is Responsible for Custom Application Code Security in SAP BTP? (September 1) · https://onapsis.com/blog/sap-btp-shared-responsibility-model/
– Deloitte Insights — From silos to synergy: How utilities are integrating AI and geospatial intelligence for resilience (April 2026 survey of 60 IOUs) · https://www.deloitte.com/us/en/insights/industry/power-and-utilities/geospatial-intelligence-ai-utilities-resilience.html
– U.S. DOE / CESER — CESER and Sandia National Lab are Using AI to Safeguard the Electric Grid (C2E2 / AI-FORTS; published August 27, updated September 3) · https://www.energy.gov/ceser/articles/ceser-and-sandia-national-lab-are-using-ai-safeguard-electric-grid
– Sandia National Laboratories — Fighting AI with AI to keep the lights on (July 30) · https://www.sandia.gov/labnews/2026/07/30/fighting-ai-with-ai-to-keep-the-lights-on/
– IgniteSAP — SAP AI Agent Hub and Agent Governance (Angus Macaulay; May 19, 2026 — baseline for the capability comparison) · https://ignitesap.com/sap-ai-agent-hub-and-agent-governance/
– ControlRooms — Agentic Troubleshooting System for Chemical & Energy Operations (September 2) · https://www.prnewswire.com/news-releases/controlrooms-unveils-first-agentic-troubleshooting-system-for-chemical–energy-operations-302867238.html