AI adoption risks in energy and utilities — and how to mitigate them at every stage

Stage 1 is not the absence of AI — it is the absence of a map. Most utilities at this stage are already consuming model output daily without calling it that: the APM suite scores transformer health with a vendor's proprietary model, the OMS vendor has quietly enabled a predictive module in the last upgrade, and a protection engineer maintains a dissolved-gas-analysis spreadsheet whose thresholds came from a machine-learned fit. None of it is on any register, so none of it has an owner, a validation record or a fallback.

The distinguishing risk is that exposure cannot be enumerated. Ask the leadership team how many models influence operational decisions and the answer is a guess that is usually out by a factor of three. That matters because every later control — validation, drift monitoring, audit evidence — operates per model. A risk framework applied to the four models you know about does nothing for the nine you do not.

This stage is cheap to leave and dangerous to stay in, because unmapped risk compounds silently. The maintenance deferral informed by an unvalidated health score does not fail loudly; it shows up eighteen months later as a transformer failure with no record that a model was ever in the loop — which is precisely the reconstruction a regulator will ask for.

Stage 2 is where risk becomes visible but controls remain paper. The pilots are real: a wind or load forecasting trial, an asset-health proof of value, an outage-prediction bake-off. To run them, operational data starts moving — historian extracts emailed to a vendor's data scientist, AMI interval data loaded into a cloud tenancy under a contract clause nobody technical has read. The controls that exist are legal ones, and legal controls do not stop a feed, classify an extract or notice that a export contained customer data.

The second stage-2 risk is claim inheritance. Vendor models arrive with accuracy figures earned on another operator's network — a different climate, a different DER penetration, a different mix of overhead and underground circuits. A storm-outage model trained on a coastal utility's history transfers badly to an inland one, and the difference does not surface in a demo. Utilities that skip acceptance testing on their own network data discover the gap after the model is coupled to operations, which is the most expensive possible moment.

What makes this stage treacherous is that its failures are quiet and its successes are loud. A leaked extract or an over-fitted demo costs nothing visible this quarter, while a good-looking pilot earns headlines and budget. The programme accelerates toward stage 3 carrying unexamined data paths and untested claims — which is exactly the cargo you do not want at the moment output first couples to the grid.

Stage 3 is the moment the risk class changes physically: a wrong output stops being a wasted analysis and starts being a wrong pre-staging decision, a deferred maintenance visit, a mis-set switching plan. The programme's risk profile now moves with the network. A distribution network is not a stationary system — DER connections grow monthly, reconfigurations change flow patterns, an AMI rollout changes the very telemetry the model was trained on — and every one of those changes is a drift event that arrives on the network's schedule, not the data science team's.

The controls that matter here are operational, not analytical. Drift monitoring keyed to network events, not just statistical thresholds. Freshness alerting on every feed the model consumes, because a historian interface that silently stops updating is more dangerous than one that visibly fails. And above all a fallback: the previous method — the deterministic storm matrix, the seasonal load profile, the manual health index — kept alive, one switch away, and drilled. An undrilled fallback is a document, and documents do not operate networks.

The subtle stage-3 risk is human: trust miscalibration in the control room. Operators either over-trust the new field (accepting recommendations without the scrutiny they would apply to a colleague) or under-trust it (quietly ignoring it, so the programme's value silently evaporates while its risk remains). Both are measurable — override rates, time-to-accept — and both respond to the same treatment: showing operators the model's failure cases and its bounds, not just its wins.

Stage 4 is where individual-model risk is largely tamed and a new class quietly replaces it: systemic risk. A portfolio of eight models serving load forecasting, asset health, outage prediction and vegetation management does not carry eight independent risks — it carries shared ones. Most of the portfolio consumes the same weather feed, the same historian, the same feature pipeline. A single upstream failure now degrades five models at once, in correlated ways, during exactly the weather event when all five matter most. The failure mode nobody designed is the one the portfolio inherited from its own efficiency.

Concentration risk arrives the same way. Standardising on one vendor's platform, one cloud tenancy or one forecasting supplier is operationally sensible and creates a single point whose commercial failure, security compromise or model regression propagates across the estate. The mitigation is not to abandon standardisation but to know precisely what depends on what — a dependency register for models, feeds and vendors — and to have exercised the loss of each critical dependency the way transmission planners exercise the loss of a line: as a contingency with a rehearsed response.

The other stage-4 discipline is portfolio-level evidence. By now the regulator conversation has changed: it is no longer 'do you use AI?' but 'show us how you manage it'. NERC CIP audits, Ofgem's licence-condition conversations and — for European operators — EU AI Act conformity all want the same artefacts: the inventory, the review cadence, the incident record, the reconstruction capability. At stage 4 those artefacts either fall out of the framework automatically, or the framework is theatre.

Stage 5 in a utility is narrower than the phrase 'autonomous grid' suggests, and correctly so. It is an enumerated set of contained, reversible actions — volt/VAR optimisation within band, DER curtailment inside connection agreements, battery dispatch within market and thermal limits — executing inside envelopes an engineer signed, with everything else escalating to a person. Switching plans, protection settings and safety-adjacent decisions stay human forever, not because models cannot rank options but because the consequence-reversibility calculus says advisory is where they belong.

The risk that defines this stage is envelope validity. An envelope is a claim about the world — these bounds are safe given this network — and the network keeps changing underneath it. New DER connections, a reconfiguration, a new interconnector: each quietly erodes the assumptions the envelope encodes. The leading indicator is the escalation rate. When the share of decisions falling outside bounds trends up, the world has moved and the envelope needs review before an incident forces one. Utilities already run this discipline for protection settings; the transfer of that habit to AI envelopes is the whole trick.

The second defining risk is complacency — months of correct automated operation recalibrate humans. Operators stop shadowing the system, the fallback drill slips, the reversion switch goes untested through two reorganisations. The mitigations are deliberately artificial: scheduled hours operating on the fallback method, injected exercise scenarios, and treating a missed drill with the seriousness of a missed relay test. Stage 5 is not a destination where risk work ends; it is a permanent operating discipline, and the utilities that hold it treat the envelope, the drill calendar and the decision log as living operational artefacts.