Utilities: keep essential services working

Utilities keep power and water reaching homes and businesses. They run networks, read meters, bill customers, fix faults and answer to regulators. This page maps that work onto the guide's frame. The frame is the layers you own, a lifecycle you run, and the capabilities you add. It shows the fourteen layers of a typical utility estate and the systems in each. It places four worked examples on those layers. The short answer: agents fit around the systems utilities already run. In the deployments found, language-model systems advise and people decide.

Who owns which work

Ownership decides which use case belongs to whom. It is not the same in electricity and water.

• Electricity retailers bill households. In Australia, an electricity distributor does not bill households. Retailers do. Billing exceptions, payment plans and hardship belong to the retailer.
• Distributors run the network. A distributor owns poles, wires, substations and transformers. It runs the control room, connects customers and restores supply. It bills retailers for use of its network.
• Meters are split too. In most Australian states, household smart meters belong to competing metering coordinators. Victoria is the exception. Its distributors own the household smart meters, so they run the meter head-end and meter data.
• New Zealand has the same split. Retailers, called traders, bill customers. Distributors bill retailers for network use. Market parties exchange standard files called EIEPs through the Electricity Authority's registry.
• Water utilities bill households directly. A water utility treats and delivers water and collects wastewater. It bills households itself, usually every quarter. There is no wholesale market in between.

AEMO is the Australian Energy Market Operator. Its business-to-business (B2B) e-Hub carries service orders, meter data and notices between market parties. Its market settlement and transfer system (MSATS) holds the standing data for each meter. The Australian Energy Regulator (AER) enforces the retail rules.

The estate

These are the systems a distributor of about one million customers and a metropolitan water utility typically run. Vendors in brackets are examples, not a claim about any named utility. OTOT (operational technology). The systems that run physical equipment: plants, grids, vehicles. Safety rules keep AI out of the control loop.Read the page means operational technology: the control systems that run the physical network. SCADA is the main one. It is the control-room system that monitors and operates field equipment.

Five points on this estate matter most for agents.

• The OT network stays apart. Control-system data usually may not leave it. That decides where an agent runs and what it reads. The historian is the safe copy for an agent to read.
• Three records must agree. In electricity they are the GIS network model, the asset register and the market's standing data. They often do not, and that gap drives exceptions in every use case below.
• The warehouse is a good source for history. It is a poor source for the current state, because it loads overnight (typical, unsourced).
• Interface monitoring shows that a message moved. It does not show that the business outcome happened. Tracing of model and tool calls is mostly absent.
• Supervision already exists. Billing teams sample released work. Contact centres score calls. Control rooms have authorisation levels. The missing role is a person who owns an agent's results and its changes.

The use cases on this estate

Four worked examples sit on this estate. Each one names its systems by layer, the process today, the process with agents, and who approves where.

Customer journeys

The customer account belongs to the retailer in electricity and to the utility in water. A customer calls, chats or uses the app. Common reasons are a home move, a meter read, a bill question, a payment plan, or trouble paying. Today a service representative works across the CIS, the CRM and the market gateway.

With agents, the agent says it is AI and offers a person. It checks identity, with a one-time code before any change. It drafts home moves through the existing market gateway. A fixed rule records life support at once. The agent explains bills without changing the account. It drafts payment plans within policy, and a trained person approves each one. It flags hardship and hands the customer over warm, with a summary.

ContainmentContainment versus resolution. Containment counts conversations that never reached a person. Resolution counts problems actually solved. Target resolution. counts contacts that never reached a person. It does not count problems solved. AGL's assistant on IBM watsonx contains 36 percent of about 125,000 chats a month (IBM, undated) [vendor]. PolyAI reports 67 percent containment for PG&E's voice agent (2026) [vendor]. The two count different things and are not comparable.

Kraken launched agents with Sierra in June 2026. They run set journeys such as identity checks, home moves and meter reads. One unnamed utility went live on about 1.3 million accounts [vendor]. No results are published, and nothing was found on hardship or payment plans. Origin says its voice agent serves over 100,000 customers (12 February 2026, search excerpt). Hansen claims its AI agent can arrange payment plans [vendor]. That is a product claim, not a named utility letting an agent change a plan alone.

Named deployments

The map places each named deployment by area and by autonomy level. The levels come from the autonomy contract. They run from assisted (A1) through delegated tasks (A2) to supervised autonomy (A3). Most deployments advise at delegated tasks (A2). Kraken's set journeys act at one utility. Nothing was found at managed autonomy (A4).

Essential Energy in New South Wales was preparing an AI that reads crew notes for safety signals. It projects finding signals in 76 percent of cases, against 59 percent for its current rules (iTnews, 16 April 2026). That is a projection, not a measurement. Prefer a utility's own filing to a vendor's claim. Kraken claims A$170 million saved at Origin [vendor]. Origin's half-year deck says A$100 to 150 million, seen only in a search excerpt.

No utility was found using agents for regulatory reporting or compliance evidence. Ofwat says regulatory reporting is one area where UK water companies report some AI use. It names no case (Kurrant, 18 June 2026).

The rules that bind the industry

The hardship precedent. In AER v Origin (Federal Court, 6 July 2022), automated processes changed payment plans without checking capacity to pay. Origin admitted over 100,000 breaches affecting over 90,000 customers. It paid an A$17 million penalty. This was automation, not AI. It still sets the line: a trained person decides any change to a payment plan. That is the rule for decisions about peopleThe rule for decisions about people. When an agent makes or shapes a significant decision about a person, such as credit, work, benefits, health, housing or hardship, a named person with authority to change the outcome decides. The person affected gets the notice the law requires, and the decision is logged with the model's inputs.Read the page. Red Energy paid A$474,600 in February 2025 over alleged failures to help customers in payment difficulty (search excerpt).

The AEMC rule change of 10 September 2026. The Australian Energy Market Commission (AEMC) made a final rule on that date. It strengthens payment difficulty and disconnection protections from 30 March 2027. It covers four of the five requests the AER lodged on 19 December 2025. A second rule, on payment plan standards, was still open on 15 October 2026. It would make retailers weigh what customers say they can afford before asking for documents. Feedback on a draft rule was due by 22 October 2026.

The information path rule for SCADA. Control systems act on the network. Agents stay on the information path around them. They read the historian copy, not the control system, and have no write path to OT. No utility was found letting an AI agent operate switching or protection (October 2026). This puts control actions in the safety actuation zone. Joint government guidance on AI in operational technology says the same (NSA, CISA and others, December 2025).

Telling customers about AI. Ofgem and the UK Energy Ombudsman set out two expectations on 5 February 2026. Firms should say when AI is in use. They should make it easy to reach a person.

Regulators asking for an AI list. The New York Public Service Commission opened Case 26-M-0552 on 17 September 2026. It asks named utilities to file an inventory of their AI uses. The Arizona Corporation Commission opened a similar inquiry on 24 March 2026. The AI use inventory shows how to build one.

Evidence and limits

The estate tables are industry knowledge, not a survey. No Australian or New Zealand utility publishes its full system estate. AEMO's pages returned errors, so market procedure details come from search excerpts.

Almost every result on this page is reported by the vendor or the operator. Few are independently audited. Containment figures count different things and should not be compared across utilities.

The named deployments that act alone run fixed rules or set journeys. No language-model agent was found changing a bill, a payment plan or a control system on its own. This guide recommends keeping it that way until a utility can measure the result and recover from a wrong action.

Next step

Start with the worked example closest to your work, such as billing exceptions. Then choose a first task.