UTILITIES
Outage Q&A wired to the OMS. Billing dispute resolution that reads the meter data. Field crew copilots that know the asset history and the safety protocols. Asset health narratives that surface real failures before they happen.
Customer-hosted on Kubernetes, in your cloud or on premises. For enterprise teams managing AI across business units.
Or email sales@kosmoy.com.
Utilities — electricity, gas, water — sit at the intersection of three pressures: the customer experience expectations of digital natives, ageing physical assets stressed by climate and load, and a regulatory environment that has tightened sharply on cybersecurity. AI is the obvious productivity unlock and the obvious risk surface at the same time.
Customer-facing AI use cases are the most visible (chatbots, billing self-service) but the bigger value is in operations: field crew copilots, asset health monitoring, outage response, regulatory filings. NIS2 has brought operational technology into scope of cyber incident reporting in the EU. NERC CIP has long imposed equivalent obligations on US bulk power. REMIT regulates trading in EU energy markets. AI tooling has to fit inside all of them.
Kosmoy is the operating layer between the utility's apps and the AI they call. Single-tenant Kubernetes deployment fits the typical utility hosting posture (private cloud, often on the same infrastructure as OMS/AMI/SCADA-adjacent systems). Every AI call is logged, attributed and contained.
Utilities historically score among the lowest CSAT in any industry. Chatbots done badly make it worse — chatbots done well, with the OMS and AMI plugged in, change the dynamic.
Asset failure modes are shifting under heatwaves, storms and floods. Predictive maintenance that reads sensor data and historical context becomes a safety question, not an OPEX question.
AI tooling has to coexist with OT environments without becoming a new attack surface. Air-gapped or DMZ-segregated deployment is common.
NERC CIP reports, ARERA submissions, Ofgem RIIO returns, FERC Form 1 — narrative-heavy reporting cycles that AI is well-positioned to draft, with care.
The regulations that shape AI in utilities — and where each one bites on AI deployment.
Energy, water and waste-water utilities are essential entities. AI infrastructure must meet incident reporting (24h/72h windows), risk management and supply chain security obligations.
Bulk power system applies CIP-002 through CIP-014. AI tooling that touches BES Cyber Systems is in scope of personnel training, change management, security monitoring obligations.
Insider information and market manipulation rules apply to energy trading. AI used in trading desks must not leak inside information through context windows.
Substation automation standard. AI tooling that reads or writes IEC 61850 data must respect the security profile.
Smart-meter and AMI data is personal data. Customer chatbots and billing AI must respect data minimisation and consent rules.
Each national regulator has tariff, conduct and reporting obligations that bound AI use in customer-facing and trading contexts.
Customer asks: 'when will my power come back?'. The chatbot is wired to the OMS and weather/storm centre — it returns a specific ETR (estimated time of restoration) for the customer's address, with the cause if known and the crew status. If the storm scope is unclear, it says so honestly and offers updates by SMS.
What to measure
Storm-day call volume on restoration questions, the share the chatbot answers, and how close its ETR came to the actual restoration time.
Customer calls: 'my bill is double last month, why?'. The agent reads AMI data, weather, tariff changes and any recent appliance billing events, drafts a structured explanation, and offers either an immediate self-service path (revised bill if a meter error is detected) or a hand-off to a human if the dispute needs judgement. Citation-grounded — references the specific meter readings.
What to measure
Days to resolve a typical billing dispute, the share resolved through self-service, and CSAT on bill-related complaints.
Line crew arrives at a substation incident. The copilot on the tablet reads the asset record, the recent maintenance history, the safety protocols (LOTO, working at heights, energised line distances), and the manufacturer's maintenance manual. Crew has the institutional answer plus the safety frame, on-site and offline if needed.
What to measure
Field time per task, and safety incidents on novel-equipment work, before and after the copilot.
Vibration, thermal and electrical sensor data on a power transformer. The agent correlates with maintenance history, weather context and similar-fleet failure modes, and produces a weekly health narrative with predicted failure modes and confidence ranges. Maintenance planners decide intervention timing.
What to measure
Forced outages on monitored assets, and how many predicted failure modes maintenance planners confirm on inspection.
Annual reliability filing or RIIO performance return. The agent reads the operational data, the prior-year filing, regulatory guidance updates, and drafts the narrative consistent with the figures. Regulatory affairs reviews, edits, signs.
What to measure
Prep hours on the standardized sections of each filing, and audit findings on filing quality.
Agent governance
Utility agents fall on a sharp risk spectrum. A customer billing agent is a low-risk Q&A surface; a field crew copilot dispatching switching procedures is a safety-critical agent with NERC CIP obligations. Kosmoy's Agent Registry captures the difference and binds each agent's allowed actions to its risk class.
The Action Capsule shows up most often for OT-adjacent agents — those that read SCADA, ADMS or AMI data. The Capsule's network egress is restricted to the AI Gateway and the specific MCP servers each agent is approved to call. An OT-side agent that tries to reach a corporate model fails at the boundary, and the Insights Dashboard captures the attempt.
Chatbot use cases
Utility chatbots are split between customer-facing (largest call deflection opportunity) and field-facing (largest productivity opportunity). Each carries different governance posture and different deployment topology.
Bill, plan, outage status, payment arrangements. Wired to AMI, billing system, OMS through MCP Gateway. Citation-grounded; never invents a tariff or a payment plan.
Push channel during storms — proactive ETR updates, crew progress, restoration confirmation. Lightweight, structured, multi-language.
Offline-capable manual lookup, switching procedure verification, safety protocol surfacing. Runs on the tablet with periodic sync; no live cloud dependency.
'What did we file last year on outage frequency for region X?'. Citation-grounded retrieval from the firm's filing library.
Utilities benefit from Kosmoy's split-personality deployment model. Customer-facing AI (chatbots, billing, outage Q&A) runs in the corporate cloud and connects to the OMS/billing systems through the MCP Gateway with strict per-tool allow-lists. OT-adjacent AI (asset health, field crew, switching support) runs in a DMZ-segregated install closer to the OT boundary, often on dedicated GPU infrastructure.
Cost economics matter in this sector — utility margins are tight and call volumes are huge. The LLM Router sends simple Q&A to a small fine-tuned SLM running on internal infrastructure and keeps frontier models for genuinely complex cases.
AI infrastructure that touches essential service systems is registered as part of the NIS2 supply chain register. Incident detection on AI-component failures flows through the Insights Dashboard. Reporting templates align with the 24h/72h windows.
Yes. Kosmoy supports air-gapped and DMZ-segregated deployments suitable for BES Cyber System adjacency. The platform itself produces no telemetry to Kosmoy; the customer controls the model image, the credentials, the retrieval set.
Yes — for the document-retrieval and procedure-lookup workflows that field crews depend on. Sensor-correlation features need connectivity to the asset health systems; the tablet caches the latest known state for graceful degradation.
AMI data never leaves the customer's infrastructure. Kosmoy's deployment is single-tenant in the utility's own cloud or data centre. Models that process AMI prompts are either local (vLLM/Ollama) or accessed via Azure OpenAI / AWS Bedrock private endpoints with zero data retention.
Enterprise proof
Kosmoy works with regulated and operationally critical organizations, including Banca d’Italia and Leonardo.
See how Kosmoy runs alongside OMS, AMI and OT systems — with NIS2, NERC CIP and sector-regulator obligations addressed.
Or email sales@kosmoy.com.