UK AI Adoption Tripled Since 2023, But Average Adopter Still Running 1.6 Tools: ONS Depth Problem Exposed

Something quietly extraordinary happened when the ONS published its first combined three-year view of AI in UK businesses last week. The headline - AI use roughly tripling from 12% to 35% of businesses since late 2023 - got the attention. The footnote did not.

Wait. Let me re-read the rules. Em-dashes must be removed. Let me redo this properly.

---

Something quietly extraordinary happened when the ONS published its first combined three-year view of AI in UK businesses last week. The headline got the attention: AI use roughly tripling from 12% to 35% of businesses since late 2023. The footnote did not.

The Depth Problem

The average number of AI technologies used per adopting business rose from 1.4 to 1.6 over the same three-year period. That is not a typo. Three years of investment, hype cycles, and boardroom mandates, and the needle moved by 0.2 tools per business. The ONS data, drawn from its Business Insights and Conditions Survey Wave 159 (fieldwork 5–28 June 2026), is unambiguous: UK AI adoption is widening, not deepening.

Think of it as a city that tripled the number of properties connected to its water mains, but average daily flow per property barely changed. The pipes are in. Almost nobody is running the taps.

Just 10% of AI-adopting businesses report using AI extensively. Close to 60% report using it to improve existing operations. Fewer than one in five deploy it for new product development or market expansion, the compounding use cases that actually alter competitive position. As the ONS framing puts it, larger firms concentrate on efficiency while smaller businesses apply AI more flexibly, but neither group is doing what the venture narratives promised.

Breaking it down by technology type: large language models lead at 18% of businesses with ten or more employees, followed by visual content creation at 16%, machine learning data processing at 12%, and image processing at 6%. Robotics sits at 2%. The information and communications sector is the outlier, with over half of businesses (58%) reporting AI use, which is the world most SIC 62.01 software developers actually inhabit.

The SaaS Pricing Rupture Running in Parallel

The depth problem on the demand side has a direct structural explanation on the supply side. Computing UK's analysis of the SaaS model, published earlier this month, documents what engineers have been watching at the contract layer. Salesforce recorded a share price loss of around 30% in 2026; growth forecasts undershot because the traditional model of linking licences to human employees is reaching its limits. Companies are actively reducing licensed seat counts. Adobe's Q1 Digital Media ARR came in at $400 million against consensus expectations of $450–460 million.

The unit economics are broken in a specific, architectural way. When one AI agent executes the workload of ten human users, the per-seat model loses its relationship to value entirely. Satya Nadella has framed it plainly: seats are becoming "just entitlement to some consumption." A Cruxy survey of 300 SaaS CEOs in April 2026 found 97% plan to retire seat-based pricing within two years, while 94% simultaneously said seat-based pricing currently aligns with their product's value. That contradiction is not confusion; it is the lived experience of a pricing transition nobody has a clean answer for.

The models replacing per-seat are, in rough order of adoption maturity: hybrid (fixed platform fee plus consumption credits), usage-based (API calls, tokens, inference hours), and outcome-based (per resolved ticket, per completed transaction). A Pilot study found seat-based pricing fell from 21% to 15% of companies in twelve months, while hybrid models surged from 27% to 41%. Zendesk introduced per-human-seat plus per-AI-resolution at $1.00 per automated resolution. Intercom charges $0.99 per AI-resolved conversation, and their CEO reported most customers end up paying more under the new model than the old, but with higher satisfaction, because cost maps to value received.

This is not a billing quirk. It is an architectural problem. AI introduces real variable costs tied to compute, tokens, and inference, in a way that traditional SaaS never did, where marginal cost approaches zero. Gross margin profiles are shifting. A product that prices at per-seat and runs inference at scale underneath is subsidising usage; a product that prices at consumption and has unpredictable inference costs is passing volatility to the customer. Neither is comfortable.

What the Company Formation Data Tells You

Lay the ONS depth problem against the formation numbers and something clarifying appears. AIBD analysis of Companies House data shows 3,159 new SIC 62.01 (computer programming) companies registered in 2026-Q3 so far, a 73.4% drop versus the prior period. IPO trademark filings in Class 38 (telecommunications and data transmission services) fell 69.6%, to just 366 filings in the same quarter. And 98.8% of active SIC 62.01 companies hold no Class 38 trademark at all.

Those are not three separate facts. They are three readings of the same instrument. Formation volumes are contracting sharply, possibly reflecting that the zero-marginal-cost-of-vibe-coding moment has already passed its peak formation frenzy, or that founders are more cautious about a market where the demand side (UK businesses running 1.6 AI tools on average) is not yet pulling hard enough to absorb unlimited new entrants. The trademark gap is the tell: the overwhelming majority of new software companies are either deferring brand investment entirely, or treating trademark filing as optional rather than a founding-week action. For a cohort where differentiation increasingly lives in the name, the workflow, and the data moat rather than the underlying model, that is a strategic oversight with a fairly predictable endpoint.

The Class 38 trademark covers data transmission and telecommunications services, the plumbing layer for any SaaS product that moves data between users and infrastructure. Software companies building AI-native products without Class 38 protection are leaving the pipes unregistered while they paint the facade.

So What

The ONS depth figure, 1.6 tools per adopting business, is the most important number in UK enterprise software right now. It tells you where the demand ceiling actually is, not where the pitch decks say it should be. Businesses using AI predominantly for efficiency in existing processes are optimising for the current workflow architecture, not replacing it. The per-seat pricing implosion at Salesforce and Adobe is partly a consequence of this: seats were sold to humans doing human-shaped workflows; if the workflow does not change, the agent layer sits underutilised and renewal conversations get brutal.

The engineering implication is concrete. AI-native products need to be built around the outcome metrics their customers actually track, resolutions, documents processed, decisions logged, because that is the only pricing unit that survives a CFO scrutiny cycle when the pilot converts to production. Building around seat counts in 2026 is like building around per-CPU licensing in 2012. Technically defensible; strategically misaligned with where the architecture is going.

6–12 Month Implications

The ONS BICS is surveyed quarterly. Watch the depth metric, not the breadth figure. If average tools per adopter moves from 1.6 to 2.0 or above by Q1 2027, it signals businesses are moving from single-use-case experimentation to integrated deployment, and the SaaS pricing transition will accelerate sharply. If it stays flat, the SaaSpocalypse narrative will fade and hybrid pricing will become the permanent settlement rather than a transitional state.

For the cohort of 3,159 new SIC 62.01 companies registered this quarter: the contraction in formation volumes means the distribution of capital and developer attention is concentrating, not expanding. That is good for the quality of what gets built. It is brutal for the median entrant who incorporated on the assumption that a rising market would do the positioning work for them. The 98.8% without Class 38 trademark protection are the ones most exposed when a well-capitalised competitor finally files the name they have been trading under for eighteen months.

Von Neumann's stored-program architecture was elegant precisely because it separated the instruction from the execution unit. The companies that survive this pricing transition will be the ones that found the equivalent separation: the pricing unit that maps cleanly to the value the execution layer actually delivers, regardless of whether the user is a human or an agent running at four in the morning.