Data Academy · Tutorial 6 of 10 · Energy & utilities

Governed AI in energy & utilities

Utilities are putting AI agents to work on outage triage, crew dispatch, vegetation risk and customer accounts, where the work is repetitive and spread across many systems. Because the same organisation runs both customer IT and grid operational technology, AI only scales when agents respect that boundary, follow the regulator’s rules and leave a full record of every decision.

01 What is changing

Where AI in energy & utilities is heading

From customer chatbots to operational agents: the clearest early value is in maintenance, outage coordination and field work rather than in conversational front ends.
From fixed cycles to risk-based programmes: vegetation, inspection and asset work are being prioritised by modelled risk rather than by calendar.
From isolated use cases to end-to-end processes: utilities are redesigning whole workflows such as storm response, rather than adding a tool to each step.
From AI near the grid to AI governed at the IT/OT boundary: agents propose and document, while switching and control stay with authorised operators.

02 Use cases

Three use cases on the governed path

Each use case runs the same path: a business question, governed context, a deterministic rule, specialist agents, a policy check, an action and a record. How the path works →

Illustrative: names and figures are invented to show the flow.

Use case 1

Risk-based vegetation work before fire season

Vegetation contact is a leading cause of faults and, in dry areas, of fire risk. Prioritising trims by measured clearance and growth means fewer vegetation faults and better evidence for the regulator.

“Which spans on the Ashcombe circuit must be cut before the dry season, and which cannot wait?”Asked by a vegetation programme manager

Context: GIS / EAM: Ashcombe circuit has 214 spans, 38 of them in a high fire-risk zone
LiDAR survey, 6 weeks old: 11 high-risk spans have vegetation inside the clearance trigger distance
Growth model: 2 of the 11 are fast-growing species projected to reach minimum clearance in about 3 weeks
Contractor schedule: framework tree crew has 4 crew-days free this month, enough for the 11 spans
Rule: Any span in a high fire-risk zone inside the trigger distance gets a trim order due within 30 days; any span projected to reach minimum clearance within 30 days is urgent and due within 7 days.
Decision: Trim 11 spans, 2 urgent Severity: High
Agents: Survey agent Reads the LiDAR and growth data per span and explains which conductors are at risk and by when.
Access agent Checks landowner access, protected sites and nesting-season restrictions for each span before the crew is sent.
Scheduling agent Sequences the 11 spans into the contractor’s 4 crew-days, with the 2 urgent spans first.
Policy: Trim orders under the existing framework contract and within the approved budget may be released automatically. Whole-tree removal, work on protected sites or anything outside the contract needs the vegetation manager. Clearance evidence is kept for the applicable standard, such as NERC FAC-003 for in-scope transmission lines.
Action: 11 trim work orders released automatically in the EAM and dispatched to the contractor, 2 marked urgent, and landowner notices sent through the standard access process.
Data: Network spans and circuits (GIS)Asset and work history (EAM)LiDAR and imagery surveysFire-risk zones and weatherContractor schedules and contracts

Use case 2

Storm outage triage and crew dispatch

In a storm, control rooms are flooded with calls, meter events and alarms, and customers wait for a credible restoration time. Faster triage means shorter outages and earlier help for vulnerable customers.

“Feeder F-17 is out: where do the first crews go, can we backfeed, and what do we tell customers?”Asked by an outage manager

Context: OMS: 46 trouble calls and 3 clusters of smart-meter last-gasp events point to recloser R-17-2 locked out
ADMS / SCADA: R-17-2 open with 2,300 customers downstream; backfeed via open point NOP-17/23 would load feeder F-23 to 410 A against a 420 A rating
CIS: 14 customers on the section are registered as dependent on medical equipment
WFM: 3 crews available, the nearest 25 minutes from R-17-2
Rule: Send the nearest crew to the predicted device first; if a backfeed would leave the adjacent feeder within 20 A of its rating, do not propose it for automatic action and refer it to the control engineer; publish a restoration time only after a crew confirms the fault.
Decision: Dispatch crew, refer backfeed Severity: High
Agents: Triage agent Correlates calls and meter events to the network model and explains why R-17-2 is the most likely device.
Switching study agent Runs the load check on the backfeed and drafts a switching sequence with the margin it leaves on F-23.
Customer agent Drafts outage messages from approved templates and lists the 14 medically dependent customers for priority contact.
Policy: Agents have no write access to SCADA or ADMS. Switching is carried out only by an authorised control engineer under the switching procedure, and access to operational systems follows NERC CIP where it applies. Customer messages use approved wording.
Action: Nearest crew assigned in WFM; the draft backfeed switching plan referred to the control engineer in the ADMS; the 14 priority customers queued for a call.
Data: Outage and trouble calls (OMS)Network model and device states (ADMS / SCADA)Smart-meter events (MDMS)Customer and medical-dependency records (CIS)Crews and locations (WFM)

Use case 3

Arrears action on a vulnerable household

Collections processes can move an account towards disconnection or a remote switch to prepayment faster than anyone notices the customer is vulnerable. Checking first means fewer harmful actions and a defensible position with the regulator.

“Account 7714-220 is five months in arrears: can we move to a prepayment switch or disconnection notice?”Asked by a collections team lead

Context: CIS / billing: arrears of £1,140 built up over 5 months, two payment arrangements broken
MDMS: smart meter able to switch to prepayment mode remotely
Priority Services Register: household added last month for medical equipment dependency
CRM: no contact for 6 weeks; a hardship fund enquiry still open
Rule: If an account is on the Priority Services Register or has an open hardship case, any disconnection or switch to prepayment is blocked and a welfare check and affordable repayment plan are triggered instead.
Decision: Block action pending welfare check Severity: High
Agents: Vulnerability agent Reads the register entry and contact notes and explains why enforcement must stop on this account.
Affordability agent Builds repayment plan options from usage, payment history and the open hardship enquiry.
Communication agent Drafts a plain-language letter that sets out the options and where to get independent help.
Policy: Collections staff may offer repayment plans. A switch to prepayment or disconnection needs the vulnerability checks required under the supply licence conditions and Ofgem’s rules on involuntary prepayment installation, and sign-off by the vulnerable-customers team.
Action: Disconnection and prepayment switch blocked in the CIS pending a welfare check by the vulnerable-customers team; repayment plan options held for an adviser to discuss.
Data: Billing and arrears (CIS)Meter capability and events (MDMS)Priority Services RegisterContact history and cases (CRM)Hardship fund and debt policy

03 The foundation

What the agents need to understand

Core entities in the ontology

CustomerPremiseMeterCircuitFeederSpanAssetOutageWork orderCrew

Systems they come from

CIS / billing: customers, premises, tariffs, bills, arrears and registers
OMS: outages, trouble calls, predicted devices and restoration times
ADMS / SCADA: network model, device states, switching and telemetry
GIS: network topology, spans, circuits and geography
EAM: assets, inspections, condition and work orders
WFM / MDMS: crews, schedules and dispatch; smart-meter reads and events

04 Guardrails

The controls that let it scale

IT/OT boundary held

Agents read operational data but never write to SCADA or ADMS; switching and control stay with authorised operators, with NERC CIP observed where it applies.

Vulnerability checked before enforcement

No disconnection, prepayment switch or debt action proceeds without a Priority Services Register and hardship check.

Approved wording to customers

Outage and debt messages come from approved templates and are logged, so customers receive consistent and accurate information.

Evidence kept for compliance

Clearance surveys, work completed and decisions are timestamped and stored so vegetation and reliability obligations can be evidenced.

Every run recorded

Inputs, rule, agent explanations, approver and resulting transaction are kept together, so any decision can be replayed for the regulator.

05 Rollout

From the first use case to many

1
Start on the IT side of an operational process
Begin with work that creates work orders or customer actions, such as vegetation or arrears, before anything near the control room.
2
Build one network and customer model
Link premises, meters, feeders, spans and assets across CIS, GIS, OMS and EAM, so agents reason on the network as it really is.
3
Write rules and autonomy levels with compliance
Agree with operations, compliance and the customer team which actions release, which wait and which are blocked, and record the regulatory basis for each.
4
Add outage support in advisory mode
Let agents triage and draft switching plans while control engineers decide, and review the drafts after each storm.
5
Scale to whole processes
Extend the same model and audit trail across storm response, asset inspection and customer operations, region by region.

06 What to measure

Outcomes, not activity

Customer minutes lostVegetation-related faultsTime to first crew on siteAccuracy of estimated restoration timesHarmful enforcement actions avoidedRegulatory evidence on time

07 Pitfalls

What usually goes wrong

Treating OT as just another API. Giving an agent a path into control systems creates safety and security exposure; keep agents on the IT side and route switching to operators.
Network model out of date. If GIS connectivity is wrong, triage points crews at the wrong device; fix the model and its update process before trusting the agent.
Collections automated without vulnerability data. Debt workflows that never read the register cause real harm and regulatory action; make the register a hard input to every rule.
Pilots that stay in pilot. Use cases built outside the live OMS and CIS stall at integration; design the first one end to end on production systems with their approvals.

08 Diagnostics

Questions to ask your team

1
Which of our operational decisions already follow written rules, and which rely on the control room’s memory?
2
Is our network model in GIS and the OMS accurate enough for an agent to triage from?
3
Where exactly is the line between what an agent may do and what only an authorised operator may do?
4
Could we show the regulator, for any enforcement or vegetation decision, the data, rule and approver in one record?

09 Keep going

Frequently asked

Will agents operate the grid?

No. In this approach agents read operational data and draft plans, but switching and control remain with authorised control engineers. Agents act only through IT systems such as the EAM, WFM and CIS.

How are vulnerable customers protected?

The Priority Services Register and hardship cases are hard inputs to the rules. If either applies, enforcement is blocked and the case goes to the vulnerable-customers team.

What does the audit record contain?

The data the rule used, the rule outcome, each agent’s explanation, the policy check, the approver and the transaction created. It can be replayed for internal review or the regulator.