The AI readiness talent gap in utilities: closing both sides of the workforce equation

Rented is not a failure state — it is where almost every utility rationally starts, because the first pilots arrive attached to vendor products and consultancy engagements. The failure is staying there. A rented capability produces demonstrations, reports and occasionally a working model, but it produces no residue: no internal engineer who understands the feature pipeline, no code the utility can modify, no judgement about what worked that survives the supplier rotating its team. The tenth engagement costs what the first did, and teaches the organisation exactly as little.

The tell is the renewal meeting. At Rented, the conversation about year two of an AI contract contains no utility voice that can challenge the supplier technically — procurement negotiates price because nobody can negotiate substance. That asymmetry compounds: the supplier's people learn the network, the feeder quirks, the data gotchas, and then take that learning to their next client. The utility has paid to train the market's engineers on its own grid.

What makes Rented dangerous rather than merely slow is the outflow side. While the capability is rented, the workforce that holds the network's tacit knowledge — the protection engineer who can read a DGA result like a novel, the control-room veteran who knows which alarms lie — keeps ageing. Every year at Rented is a year of retirements uncaptured, and that loss, unlike a delayed hire, cannot be bought back at any price.

Enclave is the most common stage and the least stable. The utility has done the hard fundraising work — headcount approved, a data science team hired, often at salary bands that required an uncomfortable HR exception — and the team is genuinely capable. What it lacks is a route into the operation. Its members sit in a corporate office, work from data extracts, and present findings to committees. The models are competent; the network carries on exactly as before. Six months of this teaches operations that the team is decorative, and teaches the team that the utility was not serious — which is when the CVs go out.

The attrition mechanics deserve more attention than they get. A data scientist inside a utility enclave is being paid below the market that recruited them, working on problems that never ship, and watching former colleagues at software firms deploy weekly. The utility's traditional retention levers — stability, pension, a thirty-year career arc — are precisely the ones this cohort discounts. The result is a two-year revolving door: hire at a premium, lose before the first production deployment, re-hire at a bigger premium. The enclave does not just underdeliver; it actively burns the employer brand needed to staff stage 3.

The exit is not more hiring — it is dissolution. The enclave dissolves into blended pods: one data engineer and one modeller paired with a protection engineer, an asset manager or a shift engineer, sitting with the operational team, owning one operational number together. Utilities resist this because it feels like breaking up the expensive thing they just built. It is actually the first moment the expensive thing starts to pay, and — the part nobody expects — the first moment the data specialists stop resigning, because their work finally lands on a real grid.

Blended is where the talent programme starts producing compound interest. The pod structure does two jobs at once: it ships operational results, and it manufactures exactly the hybrid people the market cannot sell you. Twelve months of pairing turns a data engineer into someone who understands why the feeder naming convention lies, and turns a protection engineer into someone who can read a confusion matrix. Neither conversion happens in a classroom. The pod is the classroom, and its output — beyond the model — is the first generation of staff who hold both worlds.

The binding constraint at this stage is access, not skill. A data engineer cannot build governed feeds from SCADA and the historian without crossing the OT security perimeter, and utilities are rightly conservative about who crosses it — in North America the NERC CIP obligations make that caution a compliance matter, not a preference. Clearing a data engineer to work OT-adjacent takes months of vetting, training and supervised access, which is why the role has to be hired and cleared long before the programme needs it, and why poaching someone already cleared commands the premium it does.

Blended is also the stage where the outflow side of the gap must be wired in, because pods create the first genuine consumers for captured knowledge. A retiring diagnostic engineer's judgement is worth little as a document in the EAM; it is worth a great deal as a labelled dataset a pod is actively training on. The discipline is co-labelling: sitting the expert with the pod over historical cases — DGA results, thermal images, failure records — and recording their calls and their reasoning while they are still on payroll. Utilities that skip this at stage 3 arrive at stage 4 with pipelines and platforms and nobody left who can tell the model when it is wrong.

Pipelined is the stage most talent strategies claim and few operations can evidence. The test is unglamorous: when a pod loses its data engineer, how long until a replacement of equal usefulness is productive? At stage 3 the honest answer is six months to a year of external search plus clearance plus familiarisation. At Pipelined it is weeks, because the replacement already exists inside the organisation — a conversion-cohort graduate who knows the estate, holds the clearance, and has sat in a pod as an apprentice member. The pipeline's product is not headcount; it is replaceability without loss.

The economics need to be faced squarely rather than managed by exception. A utility cannot and should not match trading-house or big-tech pay for pure ML talent — but it does not need to, because its hybrid roles are not fungible with those markets. An OT-cleared data engineer with three years of grid context has a career moat a generic data engineer lacks, and the utility can pay a defensible premium for the hybrid — benchmarked against what it would actually cost to replace them, not against the engineering band next door. Utilities that benchmark hybrid pay against other utilities' generic bands are running a subsidised training scheme for consultancies.

On the outflow side, Pipelined means capture has become boring — which is the goal. The register of single points of knowledge is reviewed quarterly alongside the retirement forecast; every critical judgement has a capture plan with a date; co-labelling sessions are scheduled work with work orders, not favours extracted in notice periods. The wider industry has begun building shared scaffolding for exactly this — EPRI's Open Power AI Consortium pools utilities' effort on power-sector AI models and practices — but the tacit knowledge of a specific network can only be captured by the utility that runs it, while the people who hold it are still on payroll.

Self-renewing is narrower and more procedural than the name suggests. It does not mean every lineworker writes Python; it means the organisation has stopped treating AI competence as a special commodity held by special people. A planning engineer extending a load-forecast feature, a substation engineer flagging a mislabelled training case through the same system they would use for a defect, an apprentice rotating through a pod as routinely as through a depot — these are the textures of stage 5. The specialist roles still exist; what has changed is that the organisation around them can absorb their departure without a programme review.

The governance scaffolding is what makes this safe rather than chaotic. Management-system standards for AI — ISO/IEC 42001 is the reference point — treat competence the way safety management always has: the organisation must determine what competence each AI-touching role requires, ensure it exists, and keep the evidence. For a utility this is familiar territory; it already runs authorisation regimes for switching, for working near live equipment, for control-room roles certified under NERC's operator-certification scheme in North America. Stage 5 simply extends a discipline the industry has practised for a century — competence as a controlled, audited property of roles — to the model estate.

Sustaining stage 5 is a renewal discipline, and regression is quiet. The signature failure is competence rot: role profiles written in the stage-4 push slowly drifting out of date while the model estate grows, until an incident review discovers that the person who approved a model change held no current competence for it. The countermeasure is the same as for any management system — periodic review keyed to change, not to calendar piety: new model classes, new platform capabilities and reorganisations each trigger a competence-profile review, exactly as a network change triggers a protection-settings review.