UK AI Adoption Tripled Since 2023 - But Depth Metrics Barely Moved, ONS Data Shows

Something quietly extraordinary happened on 20 July 2026. The Office for National Statistics published its first combined longitudinal view of AI in UK businesses, spanning 2023 to 2026, and the headline number looked like a triumph: self-reported AI use among businesses with ten or more employees has nearly tripled, from around 12% to around 35%, in roughly 30 months. The trade press noticed the top line. Most missed the depth measure buried underneath it.

A Mile Wide, An Inch Deep

The ONS tracks AI intensity through a proxy: the average number of distinct AI technologies used per adopting business. Since late 2023, that figure has risen from approximately 1.4 to approximately 1.6. Two years of relentless AI headlines, billions in enterprise sales pitches, and the needle moved 0.2. For context, that is a rounding error, not a transformation.

The survey reinforces the point from another angle. Of businesses that report using at least one AI technology, only 10% say they use AI extensively. The remaining 90% are, in the language of distributed systems, running AI at the edge of their stack: single-node, low-throughput, non-critical. The analogy from plumbing is apt. You have installed one tap. You have not yet connected it to the mains.

The technology mix tells a similar story. Large language models lead at 18% of businesses, followed by visual content creation at 16%, machine learning data processing at 12%, and image processing at 6%. Robotics sits at 2%. These are almost entirely single-layer applications: one model doing one job, not orchestrated pipelines of agents composing outputs across a workflow. The ONS data confirms what engineers already suspected. Most UK businesses have a ChatGPT wrapper somewhere on an intranet. That is not the same as AI adoption in any meaningful architectural sense.

The sectoral split is sharp. Over half of businesses in information and communication, the SIC 62.x cohort of custom software and IT services firms that dominate Companies House filings, report using AI, dwarfing other industries. The very sector building AI products is the one consuming them. The integration gap between tech and the broader economy remains wide.

The Pricing Fracture Underneath

This depth deficit collides directly with a structural crisis in how AI software is sold. Computing UK reported this week that the traditional SaaS model of linking licence revenue to human headcount is reaching its limits, and the stress fractures are visible in the earnings of the largest vendors. Salesforce has shed around 30% of its share price in 2026, with growth forecasts undershooting expectations because companies are consolidating IT estates and reducing licensed seat counts. Adobe's first-quarter Digital Media ARR came in at $400 million in net-new additions, meaningfully below analyst consensus of $450–460 million.

The structural logic runs as follows: if AI allows one developer to do the work of five, or one agent to resolve ten thousand support tickets without a human seat attached, then revenue tied to headcount declines even as the value delivered increases. The seat-based model is, as one analyst framed it, like charging per parking space for a self-driving car fleet.

Gartner projects that 40% of enterprise applications will feature AI agents by the end of 2026. Pilot data shows seat-based pricing fell from 21% to 15% of SaaS companies in just twelve months, while hybrid models, flat platform access fee plus variable AI-task billing, surged from 27% to 41% in the same period. Atlassian is one counterexample: cloud revenue grew 29% and total revenue 32% in recent quarters, suggesting that collaboration tooling with deep workflow integration has more durable pricing power than point solutions.

Three replacement models have emerged. Usage-based pricing charges per API call, token, or compute hour: good margin alignment, brutal revenue unpredictability. Outcome-based pricing charges per result; Intercom's Fin agent at $0.99 per resolved conversation reportedly drove 40% higher adoption by aligning the vendor's upside to the customer's verifiable gain. Hybrid models layer a fixed platform subscription over variable agent-execution fees, preserving some ARR predictability while capturing AI-intensity upside.

The engineering implication is non-trivial. Shifting from seat-based to consumption-based billing requires metering infrastructure that most SaaS platforms were never designed to provide. Telemetry pipelines, per-task attribution, idempotent billing events, usage anomaly detection: these are substantial additions to a stack optimised for flat monthly recurring charges.

What the Formation Data Says

Against this backdrop, the UK software formation numbers are striking. AIBD analysis of Companies House and IPO data shows 2,152 new SIC 62.01 businesses incorporated in Q3 2026, an 81.9% decline versus the prior period. Class 42 UK trademark filings, covering software-as-a-service, AI-as-a-service, and technology services, totalled 1,815 in the same quarter, down 75.5%.

Both curves are falling steeply together. The formation crash is not a sign that AI software is contracting as a sector. It signals that the easy incorporation wave, driven by founders who believed wrapping an LLM API in a frontend constituted a defensible business, has broken. The trademark data corroborates it: 95.3% of active SIC 62.01 companies currently hold no Class 42 trademark. That is not a cohort building durable IP-protected software brands. That is a cohort building prototypes.

The survivors will be the ones who solve the metering problem, own a workflow vertical deeply enough to justify outcome pricing, and build the observability infrastructure to prove their value at invoice time.

The Shallow-Adoption Paradox

Return to the ONS depth figure. If the average AI-adopting UK business uses 1.6 AI tools and 90% of adopters use AI only superficially, the limiting factor is not access to models or compute. The ONS notes that AI is more commonly used to improve existing operations than to support new market entry or product development; larger firms focus on efficiency, while smaller businesses apply AI more flexibly across activities. The DSIT AI Adoption Research published in 2026 found that increasing efficiency and productivity is the most cited motivation for adoption. Businesses are optimising, not transforming.

This has an architectural consequence that goes largely undiscussed. Efficiency-first AI is plugged into existing data flows without redesigning them. The AI layer is thin: a summarisation step here, an autocomplete there. It does not touch the underlying data models, the permission architecture, or the integration topology. So when a vendor wants to upgrade from assistant-mode to agent-mode, from co-pilot to autonomous task executor, the plumbing simply is not there to support it. The tap is installed; the mains are missing.

Six-to-Twelve Month Implications

The ONS depth data and the Computing UK pricing story are the same story told from two ends of the same pipe. Demand is wide but shallow; supply is repricing toward depth.

For UK software firms in the SIC 62 cohort, the next two quarters will stress-test whether their metering and billing infrastructure can support a genuine transition to usage or outcome models. Those still on flat seat pricing by mid-2027 will face margin compression as enterprise customers reduce headcount and argue, correctly, that they are paying for seats no longer occupied by humans.

For buyers, the ONS finding that fewer than one in five AI-adopting businesses use AI to develop new products or reach new markets is the most actionable statistic in the report. The firms still running AI as an efficiency tool in twelve months will have ceded compounding advantage to those who crossed from optimisation into product development. That crossing requires deeper integration, better observability, and pricing models that reward the vendor for delivering it.

The formation crash in SIC 62.01 is, paradoxically, good news for the firms that remain. The winnowing has begun.