UK SIC 2026 Carves AI Model Development Into Its Own Classification for the First Time in 19 Years - and the Data Already Doesn't Fit

Something quietly extraordinary happened this past spring, and almost nobody in the AI engineering community noticed. The ONS published SIC 2026, the first complete overhaul of the UK's Standard Industrial Classification since 2007, and buried inside a 668-class taxonomy was a structural admission: general-purpose AI software and AI model development are now distinct economic activities, separated from broader software categories at the 5-digit level. After nearly two decades of lumping transformer researchers and payroll-SaaS developers into the same statistical bucket, the UK government has drawn a line.

The timing, however, is awkward in a very particular way.

What SIC 2026 Actually Changed in Software Classification

The revision, effective 14 April 2026, is the first complete overhaul of the classification in 19 years. It expands from 618 to 668 classes across 22 sections, introducing 132 new codes and retiring 98. For the software sector, the critical move was architectural: AI-related activities are now explicitly recognised at the 5-digit level, pulling AI model development out of the broad software bucket that previously contained everything from embedded firmware to enterprise CRM. Think of it as the ONS finally installing a proper junction in a road network that had been routing every vehicle, mopeds and articulated lorries alike, through the same single-lane trunk road.

The ONS guidance is unambiguous on the intent: the separation of general-purpose AI software and AI model development from broader software categories reflects genuine differences in economic activity. Future updates to the explanatory notes are already planned, with the ONS noting that AI classification will be considered in future product classification reviews as the sector continues to evolve.

There is, however, a significant catch. Companies House has not yet set a mandatory adoption date. The SIC 2026 framework is not yet in operational use across ONS systems, with the earliest planned adoption appearing in the 2031 Blue Book, the UK National Accounts publication. In practice, the plumbing has been redesigned on paper; the water is still flowing through the old pipes.

The Commercial Data Already Straining the Old Taxonomy

Here is where it gets interesting for anyone who cares about what the numbers actually measure. Beauhurst's updated ranking of the top 100 UK SaaS companies, refreshed on 29 September 2026, shows those companies have raised £34.4bn across 667 equity rounds, the largest total of any sector Beauhurst tracks. By 22 September, 2026 had already recorded £4.13bn in UK SaaS equity, against a peak year of £5.47bn in 2021. FNZ leads at £2.29bn, followed by autonomous driving company Wayve at £2.15bn and Ocado Group at £1.76bn.

Separately, Beauhurst's AI startup ranking, updated the same day, shows the top 100 UK AI startups raised £1.56bn in just the first nine months of 2026, equivalent to 62% of their combined historical total. Concentration is extreme: Ineffable Intelligence alone accounts for £814m, or 32% of the total raised by the entire cohort.

So what is the overlap between those two datasets? Precisely the problem SIC 2026 was trying to solve. A company like Wayve, autonomous driving, deep neural networks, perception stacks, appears in the SaaS ranking because its delivery mechanism is software-as-a-service. But its engineering activities look nothing like a subscription HR platform. Under SIC 2007's legacy codes, both sit in the same statistical neighbourhood. Under SIC 2026, they would be formally separated. But the data collection won't catch up until 2031 at the earliest.

The Trademark Signal Is Already Moving

There is a faster-moving proxy. AIBD analysis of UK Intellectual Property Office (IPO) Trade Mark Database (TMD) filings shows 1,228 applications in Nice Class 38, covering telecommunications and data transmission services, in Q3 2026: a 2% rise on the prior period. Class 38 is where AI companies filing for brand protection around inference APIs, model hosting, and data pipeline services tend to cluster, sitting alongside the more obvious Class 42 (software as a service) and Class 9 (downloadable AI software).

The Nice Classification itself went through its own taxonomic update on 1 January 2026, when the 13th Edition came into force. Artificial intelligence as a service (AIaaS) was formally recognised as a service in Class 42, a structural change that Burges Salmon described as one of the most impactful shifts to trademark workflows in recent years. The dual-class search requirement this creates, where practitioners must now check both pre- and post-2026 classes for conflict, has added measurable analytical overhead to IP clearance processes across the UK and the 150 other WIPO signatory states.

The trademark data is, in a sense, the canary. Companies brand their products before they generate revenue and long before they appear in official statistics. The 1,228 Class 38 filings in a single quarter suggest the communications-layer infrastructure of AI, the pipes connecting models to applications, is being commercialised at a pace the ONS data collection architecture cannot yet resolve.

The ONS Is Using AI to Classify Itself

Perhaps the most quietly remarkable footnote in the SIC 2026 guidance is this: the ONS has developed an open-source AI tool that can classify business descriptions into statistical categories. The tool supports assignment of businesses to industries, flexible classification against alternative frameworks, and improved responsiveness to emerging economic activities. The statistical office is deploying a language model to manage the classification problem that language-model companies are causing. Alan Turing would have appreciated the recursion.

But there is a real engineering point here. Manual classification at the rate that new AI companies are incorporating is not feasible. The ONS received confirmation via FOI in August 2026 that its Business Insights and Conditions Survey was tracking AI use across SIC sectors, but at aggregated industry level, not at the 5-digit granularity that SIC 2026 now theoretically permits. Over half of UK businesses in information and communication (58%) already report using AI. The data exists; the classification plumbing to route it correctly does not yet.

The 6-12 Month Implications

For engineers and infrastructure teams, the immediate practical consequence is nil. SIC 2026 changes nothing about how you build or deploy models today. But the medium-term signal matters in three ways.

First, funding and procurement eligibility increasingly flow through industrial classification. As SIC 2026 notes explicitly, the framework shapes everything from R&D tax relief to government procurement, so the eventual separation of AI model development into its own code will create new qualification categories for grant programmes, Industrial Strategy contracts, and Innovate UK funding streams.

Second, the ONS AI classification tool is a live experiment in automated economic measurement. If it performs well, it becomes the template for near-real-time sector tracking, which would compress the lag between commercial activity and official data from years to quarters. That is a meaningful change for everyone writing investment theses based on sector-level statistics.

Third, the trademark signal, 1,228 Class 38 filings in Q3 alone, suggests that the AI infrastructure layer (inference APIs, model routing, telemetry pipelines) is where commercial brand-building is currently concentrated. The companies filing those marks are building the traffic management system for a road network that the ONS is still drawing on paper. Watch that gap.

The taxonomy always lags the technology. Von Neumann's stored-program architecture predated any formal classification of "computer programming" by years. What is different this time is that the ONS knows it, has said so publicly, and has built an AI to help close the distance. Whether the plumbing gets connected before the next revision cycle is, in its quiet way, one of the more consequential infrastructure questions facing the UK's AI economy.