AI in the energy transition, from the operator's chair

Sitting on the operator side of the energy transition, the picture on AI is different from the one presented by the vendors. AI helps in a few real places, and unhelpfully in many more. The pattern of where it helps and where it does not is worth writing down, because the split is not what most marketing materials suggest.

Where AI genuinely helps

Reading long, standard-shaped documents faster. Grid-code compliance dossiers, DNO application packs, planning submissions, HSE reports, permit conditions. These are documents where the structure is well-defined and the reader is looking for specific things. An AI that reads them well is a real time-saver for a stretched team — provided the output is source-bound, so a colleague can verify what was extracted without re-reading everything.

Finding contradictions across a document pack. The single genuinely-hard problem for a human is holding hundreds of documents in mind well enough to notice that one contradicts another. An AI that flags — with citations — where two documents disagree on a fact is doing work that a human cannot do at scale, and that a horizontal search tool cannot do at all.

Draft assembly against a known standard. Not "write the report." Assemble a draft where every claim is tied to a source in the operator's own pack, ready for a person to review, adjust and sign. The AI is doing the tedious cross-referencing; the human is doing the judgement.

Where AI unhelpfully does not

Anywhere the authority is a live engineering judgement. Deciding whether a specific asset is fit for a specific duty, in a specific loading regime, is not a document-reading problem. It is an engineering problem where the document is one input among many. AI that pretends otherwise adds risk without adding value.

Anywhere the value is speculative writing. Regulator-facing artefacts are not creative documents. Producing fluent English that "sounds like" a strong regulatory response is not what makes a regulatory response strong — accuracy and evidence do. Fluent hand-waving is worse than an honest gap.

Anywhere the underlying data is disputed. If two of the operator's own sources disagree about a fact, no AI can resolve which is right. It can flag the disagreement — which is genuinely useful — but the reconciliation is a human call. Systems that quietly pick one and move on are hiding the important problem.

The traceability problem bites hardest here

Regulated energy estates are made of decades of documents, revisions, superseded procedures, drawings that have quietly stopped matching the plant, and standards that keep moving. The single most useful AI application in an operator's world is not a chatbot — it is a system that answers which revision is authoritative, with the reason.

Nothing about a general-purpose enterprise search product answers that question. The tools retrieve; they do not reason about supersession. And in a setting where the wrong revision of a procedure can cause real harm — switching errors, LOTO mistakes, isolation failures — knowing which document is current is not a nicety. It is the whole point.

This is the piece of the AI opportunity that we think is most undervalued in energy-sector marketing right now. It is not glamorous. It does not demo well in a two-minute video. And it is precisely the thing every operator we have spoken to needs.

What follows for an energy operator considering AI

Two useful filters when a vendor is in the room:

• Ask where the trail lives. Not "does the system retrieve sources" — can a colleague, a year later, verify how a specific output was produced? If the answer is a hand-wave, the system is not built for a regulated setting.
• Ask what the system does when two of your sources disagree. The honest answer is "flags it for a human to decide." Any other answer means the system will occasionally hide the important problem — and you will not know when.

The technology is genuinely useful. The buyer's job is to insist on the traceability that makes it defensible. That is a discipline the vendor does not always volunteer.