The AI readiness framework for utilities: scoring a candidate before you fund it

Stage 1 is rarely ignorance, and it is almost never a shortage of assurance culture. The same utility that cannot say whether it is ready for a machine-learning model will not energise a substation without a signed site acceptance test, will not change a relay setting without a protection review, and will not put an operator on shift without a certification record. The discipline exists in abundance. The problem is that the AI candidate arrived through the innovation door rather than the engineering door, and nothing at that door asks for evidence.

The observable tell is asymmetric confidence. Ask three senior people whether a named candidate is ready and you get three confident, incompatible answers, each correct about the part of the estate that person can see. The asset manager knows the inspection history is digitised; the GIS lead knows the connectivity model has an unquantified error rate; the security lead knows nobody has determined whether the candidate lands in a CIP-scoped system. None of them is wrong. There is simply no frame that can hold all three answers at once and score them.

This is a cheap stage to leave and an expensive stage to stay in, because at stage 1 every AI proposal is re-litigated from first principles and none of the argument amortises. The first instrument does not need to be sophisticated; it needs to exist and be written down. A single page naming one candidate decision, the evidence that would have to exist for it to be ready, and where that evidence would come from, is already a stage-2 artefact — and it takes an afternoon, not a programme.

Stage 2 is the most dangerous stage on this ladder, because it produces an artefact that looks like governance and behaves like a newsletter. The survey has a scale, a scoring rubric, a comparison against last year and a slide that goes to the board. It is genuinely useful for one thing — opening a conversation about which parts of the business feel furthest behind — and useless for the thing it is invariably used for, which is deciding whether a specific piece of work should proceed.

Two structural defects cause this, and neither is fixable by improving the questionnaire. The first is self-report: the team most invested in a favourable answer supplies the answer, and does so without malice, because everyone scores their own house against the standard they know rather than the standard the assessor means. The second is resolution. An organisational readiness score cannot approve or refuse a project any more than an average annual temperature can tell you whether to take a coat out today. The unit of assessment and the unit of decision do not match.

Time at stage 2 is not neutral; it inoculates the organisation against the next attempt. Once the survey has run twice and changed nothing, the third round is completed carelessly, the scores drift upward for reasons nobody can name, and 'readiness assessment' acquires a reputation as an administrative exercise. Utilities that have run three annual readiness surveys are usually harder to move to stage 3 than utilities that have never run one, because the word itself now carries a meaning you have to argue against before you can start.

Stage 3 is the first stage at which a readiness answer can be checked by somebody who was not in the room. The dossier's power is not its score; it is that the score is reconstructable. Six months later a different assessor can pull the same freshness report, count the same reconciliation exceptions, read the same interface record, and either agree or produce a specific disagreement. That property — reconstructability — is exactly what a utility's commissioning records already have, and it is the property that converts readiness from a conversation into an engineering artefact.

The character of the argument changes here, and the change is worth naming because it is the main practical benefit. Stage 1 and stage 2 disputes are about opinion and method, and they do not terminate. Stage 3 disputes are about artefacts, and artefact disputes terminate quickly: either the ninety-day freshness report exists for the named feeds or it does not; either the GIS-to-ADMS reconciliation exception count for the in-scope feeders has been produced or it has not. Meetings get shorter. Teams stop preparing rhetoric and start preparing evidence, which is a much cheaper thing to prepare.

The constraint that emerges at stage 3 is authority, and it is the reason most good frameworks stall here. The dossier is honest, well-researched and advisory. Proposals that fail it still get funded, because funding happens in a different room, on a different calendar, chaired by people who never see the dossier. What eventually kills the practice is not disagreement but timing: the dossier arrives after the money, at which point its only remaining function is to be filed as a risk item and referenced in a post-incident review.

Stage 4 is where the framework acquires the only property that ultimately matters: the ability to say no and be obeyed. Everything before this is diagnosis. The mechanics are trivial — a mandatory field on a form, a required attachment on a change record, a line on the operational readiness review checklist. The difficulty is entirely constitutional: somebody has to be willing to hold the first refusal against a sponsor who has already announced the programme.

What makes it survivable is that a utility already runs gates of exactly this shape, and everyone in the building accepts them. Nobody argues that a substation should be energised without a site acceptance test record, or that a protection setting should change without review, or that an operator should take a desk without a current certification. The AI gate is not a new kind of authority being invented; it is the existing kind, extended to a new class of change. Framing it that way — in the vocabulary of commissioning rather than the vocabulary of digital governance — is worth more than any amount of framework design, because it moves the argument from 'why are you slowing us down' to 'which acceptance test applies here'.

The failure that emerges at stage 4 is condition inflation. Refusal is expensive and visible, conditional passes are cheap and invisible, so panels drift toward passing everything with a list of conditions attached. Within a year the register holds dozens of open conditions, the median age is measured in quarters, and a conditional pass has become an unconditional one with paperwork. The metric that matters at this stage is not the pass rate; it is the condition closure rate and the age of the oldest open condition.

Stage 5 is narrower than it sounds, and the narrowness is the point. It is not continuous assurance of everything; it is a small, enumerated set of live decisions whose readiness is maintained the way a utility already maintains relay settings and operator certifications — valid for a stated interval, re-earned rather than assumed, and revocable by a named person without a committee. Anything outside that enumerated set stays at stage 4, reviewed on a calendar, which is a perfectly respectable place for most of a portfolio to sit permanently.

The engineering is largely already in place by the time an operator reaches here: the feeds are monitored, the decision log exists, the revert route has been exercised. The genuinely hard part is deciding, in advance and in numbers, what 'no longer ready' means — and then accepting that the number will sometimes fire when nobody wants it to. A withdrawal trigger that has never fired is either an extraordinarily stable estate or a threshold set where it cannot reach. The second is far more common, and the way to tell them apart is to look at how the threshold was chosen.

Stage 5 is also the stage most likely to regress, and it regresses in silence rather than with an alarm. Expiries get extended in a governance meeting because nothing appears to have changed. Thresholds get widened after a run of alerts that turned out to be seasonal. The withdrawal authority migrates from a duty manager who is on shift to a committee that does not meet during a storm. None of these is a decision anybody would defend if it were written down as a decision; they happen because standing is administratively inconvenient and nobody is measuring its decay.