AI transformation Learning from the Utilities Experiment

What We Measure, What We Miss

The AI adoption measurement gap — a briefing paper

Prepared to support a decision on impact to the AI Transformation Register · grounded in current UK adoption data · July 2026

The question

The UK’s utility privatisations built a real, enforced measurement system — the Regulated Asset Base and its associated operating-cost controls — around the regulated core of each company. But for at least some utilities, the more consequential boardroom conversation was about growing the unregulated business: diversification, international contracting, adjacent services. The regulator’s instruments never reached that conversation. Individual companies pursued it to very different degrees and with very different results, and the regulatory measurement system could not distinguish, from its own data, which company was doing which — both could show identical RAB compliance while pursuing opposite strategies.

The question this paper asks is whether AI adoption — inside a single organisation, and across the UK as a whole — has the same structure: a real, increasingly sophisticated measurement system tracking one thing, while the more consequential strategic choice happens somewhere that system was never built to see.

The UK measurement landscape, as it currently stands

Multiple credible instruments currently measure UK AI adoption, and they do not agree, because they are measuring different populations against different definitions of “using AI.” The Office for National Statistics’ Business Insights and Conditions Survey put business AI use at 23% as of late September 2025, up from 9% two years earlier. The government’s own commissioned research — 3,500 business interviews conducted for DSIT by IFF Research and Technopolis, fieldwork February to May 2025 — found a tighter figure: around 16% of firms had deliberately deployed an AI technology with a defined business purpose, a further 5% had concrete plans, and roughly 80% had neither. The British Chambers of Commerce reports a considerably higher “active use” figure, around 54%, rising sharply from 35% the year before. Compiled across all published sources, the range runs from about 16% to as high as 71–78% depending entirely on how “using AI” is defined — strategic deployment with a defined purpose, versus any use at all, including AI features embedded in ordinary software.

One further data point matters more than the headline range: across the consultancy benchmarks, only around 7% of organisations report pursuing an enterprise-wide AI strategy with defined governance, even though 85–91% report increasing their AI budgets year-on-year. Spending is nearly universal; strategy, on this evidence, is rare.

The productivity-profit gap

Roughly 75–91% of AI-adopting UK businesses report productivity gains. Only around 12% report increased revenue attributable to AI. The same pattern holds globally: McKinsey’s data shows approximately 88% of organisations say they use AI, only about 1% describe their rollout as mature, and just 6% report meaningful financial returns. Real, measurable local efficiency is widespread; its conversion into business advantage is rare — and no current instrument distinguishes, at the point of measurement, which adopting organisations are on which side of that line.

The capability gap running alongside it

One tracker built from the same official data finds that 97% of organisations report an AI skills gap, and roughly two-thirds of employees have had no formal AI training, even as adoption climbs. Tool deployment and judgment formation are moving at different speeds, and only the first is being tracked as a headline national statistic.

Two adoption philosophies, indistinguishable on the dashboard

The data above is consistent with two genuinely different adoption strategies coexisting under one set of national statistics. The piecemeal strategy is tool-by-tool and unit-by-unit: individual business functions or teams adopt AI to solve a local problem, with no cross-functional view of how the pieces might combine. It is easy to measure — licences, usage, spend — and it is exactly what the productivity-gains figures capture: real, local, and largely non-compounding. The strategic-optionality approach is deliberately cross-cutting: AI is used to connect ideas and capability across business, IS, IT and OT rather than within a single domain, judgment is retained and grown rather than delegated away, and the aim is durable business advantage and future flexibility rather than a single efficiency gain. It is harder to measure, slower to show results, and is plausibly the source of the small minority reporting mature rollouts and real revenue gains.

An organisation, or the UK as a whole, can report the same adoption percentage while pursuing either strategy — exactly as a utility company could show identical RAB compliance while either sitting close to its regulated core or aggressively building an unregulated growth business. No current survey, public or private, asks the question that would separate the two populations: not “are you using AI” but “is this being built as a connected capability or installed as a set of separate tools.”

Why this matters for adoption itself

Boards and governments act on what is in front of them. If the only figures reported upward are adoption percentage, spend, and productivity gain, the rational response — even for leadership that might privately prefer to build strategic optionality — is to optimise for what is visible and rewarded: presence, not advantage. This is a structural bias built into the measurement architecture itself, independent of any individual leadership team’s actual intentions. A board chasing a reportable adoption number will fund the piecemeal path, because it is the path the current instruments can see and credit. The utilities parallel again: the RAB didn’t just fail to see the unregulated strategy — by being the thing regulators and investors could see and price, its existence quietly pulled attention and capital toward what it measured.

Where this differs from what the register already holds

The AI Transformation Register (v0.4) already contains two related mechanisms, and distinguishing this paper’s finding from both is the register-impact question this briefing exists to support. Inversion VII, the metrics inversion, addresses a single organisation’s usage dashboard registering dependency formation as success — an operational-level mismeasurement inside one company. Inversion IX, the value-mechanism split, addresses a single function’s value rating conflating two opposite mechanisms — delegation-value versus frame-expansion value — at the level of one row in one organisation’s capability assessment. What this paper describes sits at a different altitude again: not one dashboard, and not one function, but the entire national measurement ecosystem — ONS, DSIT, the sector surveys — having no instrument anywhere for the piecemeal-versus-strategic distinction, at the population level, across every organisation adopting AI in the UK at once.

Open questions for the recognition function

Three ways this could be treated, offered without a ruling attached:

Option A — a sharpening, not a new mechanism. This is Inversion VII restated at national scale rather than organisational scale: the same mismeasurement, a different altitude. If so, the register entry is strengthened with this evidence rather than extended.

Option B — a genuinely new inversion. The population-level absence of any piecemeal/strategic instrument is a distinct claim from either VII or IX — closer in kind to how the Custody Problem became a new named pattern rather than a restatement of the simultaneity mechanics it sits inside. On this reading it would earn new register content: perhaps an eleventh inversion, or a fourth Kokomo.

Option C — matrix evidence, not register mechanism. Consistent with the uniform-high-2033-risk finding held back at v0.4, this could be treated as evidence that the register’s existing tools (the judgment register, the keep-manual list, judgment-quality metrics) are needed — valuable for the October argument, but not itself a new mechanism the register lacked.

The register’s own Return Loop test (Section 14, v0.3–v0.4) is the applicable discipline: does this teach the register something it did not already hold, or does it confirm what Inversions VII and IX already say, at a different scale? That judgment belongs with the recognition function, not with this briefing.

The question worth holding open Of the 16–78% of UK organisations counted as “adopting AI” depending on whose survey you read, what fraction are building connected, judgment-retaining capability — and what fraction are installing tools in silos? No instrument currently in use, public or private, can answer that question. Given the productivity-profit gap the data already shows, it may be the single most consequential unmeasured variable in UK AI adoption today.

Status Briefing paper, pre-register. Adoption figures drawn from ONS Business Insights and Conditions Survey (Wave 141, fieldwork to late September 2025), DSIT/IFF Research and Technopolis AI Adoption Research (3,500 interviews, fieldwork February–May 2025), British Chambers of Commerce/Atos and University of Essex research (2025–2026), and consultancy benchmark compilations current to Q1–Q2 2026; figures should be treated as indicative given the definitional spread documented above, and re-checked nearer any publication date given the pace of new survey waves. Register impact deliberately left undecided — this paper exists to support that decision, not pre-empt it.