Economic research · Working paper

Who Founds AI-Named Companies? Evidence from the United Kingdom Company Register

Matt Cortland (Prime Directive AI Ltd) and Dr John Fleming (AI researcher, University of Oxford) Preliminary working draft, 31 May 2026

(Academic-register variant of the Jimmy McLoughlin preliminary insights, written for side-by-side comparison with the accessible version. Same findings, formal voice.)

Abstract

We assemble a near-complete record of United Kingdom company formations from the Companies House bulk register and combine it with the persons-with-significant-control (PSC) register and officer appointments retrieved from the Companies House API. We identify companies whose registered names signal artificial-intelligence activity and characterise their founders relative to the broader population of UK formations. We document a structural break around 2023, coincident with the mainstream diffusion of ChatGPT: the age distribution of AI-named company founders shifts younger, the share of single-founder companies rises, and ownership and control increasingly reside in the same individual. We further show that the markets these firms serve concentrate in occupations that an independent measure (Massenkoff and McCrory, 2026) rates as highly exposed to AI, and that the founder population is increasingly composed of foreign nationals. All results are descriptive; we make no causal claims. The analysis is observational and bounded by the coverage and self-reported fields of the UK register.

1. Introduction

The labour-market consequences of generative artificial intelligence have, to date, been studied primarily on the demand side: which existing occupations are exposed, and whether employment or wages have responded (Eloundou et al., 2023; Brynjolfsson et al., 2025; Massenkoff and McCrory, 2026). We examine a complementary margin, firm formation, and ask whether the diffusion of AI has changed the composition of those who found AI-named companies. Company registers are a natural setting for this question: incorporation is a dated, mandatory, near-universal record, and the UK register additionally captures beneficial ownership (PSC) and officer appointments. We treat the population of UK AI-named company founders as an observable and compare it to the contemporaneous population of all other UK founders.

2. Data

Our primary source is the Companies House bulk company snapshot (approximately 5.7 million live UK companies), combined with the PSC snapshot (approximately 15.5 million control records) and officer appointments retrieved company-by-company from the Companies House REST API (collection ongoing at the time of writing; complete for the AI subset). We additionally classify companies' activity from their public websites using an automated scrape and a large-language-model classifier (high-confidence classifications retained; website coverage approximately 76% of the AI subset).

We define an "AI-named company" by registered name: a strict definition requires "AI" as a standalone token, and a broader "AI ecosystem" definition admits nine related terms (artificial intelligence, GPT, LLM, machine learning, deep learning, generative AI, Copilot, AGI, agent). A "founder" is a person with significant control notified within 60 days of incorporation, or a director appointed within the same window. The comparison group is all other UK formations over the same period.

3. Methods

We compute, by incorporation year: (i) the age distribution of founders, using the year of birth disclosed on the PSC register (Companies House publishes month and year only); (ii) the single-founder share; (iii) a founder-led concordance, the share of companies for which the same individual, matched on surname and date of birth, appears on both the PSC and director registers; (iv) rebranding into AI, identified from the dated previous-name fields; (v) an occupational mapping, in which each company's website-derived vertical is matched to an occupation and that occupation's "observed exposure" score in Massenkoff and McCrory (2026); and (vi) founder nationality, as self-reported on the PSC register, with British, English, Scottish, Welsh, and Northern Irish merged to a single UK-national category.

4. Results

4.1 A structural break in founder composition around 2023

The share of AI-named company founders aged under 26 at incorporation rose from 4.1% (2018) to 12.8% (2025), a 3.1-fold increase, against a 1.8-fold increase in the comparison group. The single-founder share of new AI-named companies rose from 65.5% (2018) to 82.9% (2026, partial year). In both series the change is most pronounced from 2023 onward, the year following the public release of ChatGPT. Pre-2023, AI founders were older and more team-based than the comparison group; post-2023, they are younger and more often solo.

The aggregate formation series shows the same break. AI-named company formation rose steeply after ChatGPT, setting a new monthly record in April 2026 (Figure 1). (The apparent flatness of the 2010s is partly survivor-undercounting of older years in the live register, which omits since-dissolved companies; the break holds as a rise in AI's share of all formation.) The annual totals trace the same curve, with the partial 2026 year projected at its current run-rate (Figure 2).

Figure 1. Monthly UK AI-ecosystem company formations, all time. Flat through the 2010s, then a sharp post-ChatGPT acceleration; April 2026 set a new monthly record. Dashed lines mark major model releases.

Figure 2. Annual UK AI-ecosystem company formations, 2015 to 2026. The 2026 bar is part-filed; the faded portion is the run-rate projection. Dashed lines mark major model releases.

Figure 3. Share of founders under 30 at incorporation, UK AI-named companies versus all other UK companies, by year. The two series cross over in 2023. The 2026 point is provisional (partial year, incomplete date-of-birth coverage) and is drawn faded; the dip appears in both series and is most likely a coverage artifact.

Public attention moved on the same timeline. UK search interest in "AI agent" and "Agentic AI" (Google Trends) was negligible before late 2024 and rose steeply through 2025 and 2026 (Figure 4). Global news-article volume for the single term "AI agent" (GDELT) follows the same path: background noise until late 2024, then a steady climb from a few hundred articles a month in early 2025 to nearly 2,000 by spring 2026 (Figure 5). The same one term moves with the step-up in formations across all three signals. We treat these as correlates of attention, not causes; search interest is a relative index, and the news series reflects GDELT's monitored set rather than all coverage.

Figure 4. UK Google Trends search interest. “Artificial intelligence” (grey) is indexed separately; the agentic terms (blue, green) share a scale. Relative interest, read as timing not magnitude.

Figure 5. Global monthly news-article volume for the single term “AI agent” (GDELT), the same term as the search series in Figure 4. Background noise through 2024, then a steady climb to nearly 2,000 articles a month by spring 2026. Raw counts over GDELT's monitored news set.

The same shift appears in the company names themselves. The count of newly incorporated AI-named companies whose name contains "agent" rises to 155 in 2025 from a baseline of 60 to 90 a year, and "agentic", essentially unused before 2024, begins appearing in incorporation paperwork exactly as the attention wave peaks (Figure 6).

Figure 6. UK AI-named companies whose name contains “agent” or “agentic”, by incorporation year. “Agent” jumps to 155 in 2025 from a ~60-90 baseline (software/multi-agent systems is an older term); “agentic” is a net-new coinage appearing as the attention wave peaks. 2026 is a partial year.

4.2 Concentration of ownership and control

Among AI-named companies incorporated in 2020 or later for which both registers are populated, 95.9% are founder-led in the sense defined above: the same individual owns and directs the firm. Founders born in 2000 or later are over-represented among AI-named company owners by a factor of 2.2 relative to the all-company PSC population. The rise in solo founding tracks the rise in formations directly: as AI-named formation rose sharply (about 6x as a share of all new companies; the raw count multiple is inflated because older years omit since-dissolved companies in the live register, so the share is the honest measure), the single-founder share climbed from about 68% to over 80% (Figure 7).

Figure 7. Annual AI-named company formations (bars, left axis) against the solo-founder share (line, right axis). As AI-named formation multiplied several-fold (about 6x as a share of all new companies) from 2018, the solo-founded share rose from about 68% to over 80%. The raw bar counts undercount older years (the live register omits since-dissolved companies), so the share is the honest measure of growth.

4.3 Renaming into AI

We identify 300 companies that adopted "AI" into their name in 2025, up sharply from 2018 (the 2018 count of 35 is depleted by survivor-undercounting of older years in the live register, so the precise multiple is overstated). On the run-rate of the complete months of 2026, we project approximately 370 for the full year. Companies that rebrand into AI are less concentrated in software than those founded as AI (approximately 50% versus 60% with a primary software classification), consistent with rebranding capturing cross-sector incumbents.

Figure 8. Companies that changed their registered name to add “AI”, by year of the change. The 2026 bar is the run-rate projection (faded).

By primary SIC code, roughly half of rebranders were already software or IT, but about one in three came from outside technology (retail, finance, education, healthcare, legal), the genuine cross-sector pivots (Figure 9). A residual "other" category is dominated by generic catch-all codes (other business support, other professional and scientific activities, other service activities) rather than a named sector.

Figure 9. Primary sector (SIC) of companies that renamed into AI. Roughly half were already software; about one in three came from outside tech. “Other” is mostly generic catch-all codes (other business support, other professional/scientific, other service) rather than a named sector.

Naming conventions reinforce the cultural reading. UK AI-named company names draw heavily on a shared vocabulary of mythology and science fiction (and, secondarily, animals), a recognisable startup naming culture (Figure 10). Domain choices show the same signalling: the .ai top-level domain is now widely adopted alongside the conventional .com and .co.uk, with .ai over-represented relative to the wider company population (Figure 11).

Mythology / sci-fiAnimal

Figure 10. Most common mythology/sci-fi and animal words in UK AI-named company names, an indicator of naming culture in the sector.

Figure 11. Domain endings (top-level domains) of UK AI-named companies with a known website: the adoption of .ai relative to .com and .co.uk.

4.4 Sectoral composition: registration codes versus revealed activity

By self-reported SIC code, AI-named companies appear heavily concentrated in software: among companies we can classify from their website, 70.3% carry a 62xxx information-technology code. This overstates concentration, because the code is chosen at incorporation and rarely revised. Classifying instead by website-described activity (Figure 13), only about a quarter of firms are genuinely horizontal, general-purpose (11.8%) or developer-tools and infrastructure (11.2%); the remainder distribute across finance, retail, content, healthcare, marketing, education, security and others, with no single vertical exceeding 12%. On revealed activity, AI is applied across industries rather than confined to a software sub-sector, and the official classification cannot see it.

Figure 12. Sectoral composition of UK AI-named companies over time (self-reported SIC codes). Pure software rises from 56% (2018) to 79% (2026); marketing and advertising is the only non-software sector gaining share; specialist verticals such as healthcare and finance remain small and flat. Companies may hold multiple SIC codes, so shares are independent and need not sum to 100%.

Horizontal / infrastructureIndustry-specific vertical

Figure 13. Revealed activity from each company's website, the counterpart to the SIC code in Figure 12. 70.3% of classified AI-named companies file a generic software SIC, but only about a quarter are genuinely horizontal (general-purpose or developer tools/infrastructure); the remainder build for a specific industry. Website-classified companies only; unclassified excluded.

4.5 Occupational exposure

Mapping the website-derived verticals to occupations and to the Massenkoff and McCrory (2026) exposure scores, every vertical that maps cleanly corresponds to a high-exposure occupation, with a count-weighted mean exposure of approximately 65%: marketing (64.8%), computer programming (74.5%), financial analysis (57.2%), medical records (66.7%), and information security (48.6%). AI-named company formation is concentrated in the same occupations that independent measures rate as most exposed to AI. We interpret this as a supply-side counterpart to the demand-side exposure literature, and note the additional observation that the young cohort which Massenkoff and McCrory find facing modestly slower hiring into exposed occupations is the same cohort we find over-represented among founders. This correspondence is correlational; we do not identify a mechanism.

Figure 14. UK AI-named companies by the AI-exposure of the occupation their product targets, matched to Massenkoff and McCrory (2026). The 56% not cleanly mapped to a single occupation are excluded.

4.6 Founder nationality

Among AI-named companies incorporated in 2023 or later, 35.7% of owners hold a foreign nationality, up from approximately 31% before 2022. The leading foreign nationalities are Indian, Chinese, and Pakistani. Foreign-national founders have younger median years of birth (approximately 1991 to 1993) than UK-national founders (1982). We emphasise that this is the nationality of the owners of UK-registered companies, who are predominantly UK-resident; it characterises who incorporates here and not migration flows.

Figure 15. Foreign-national share of UK AI-named company owners, by company incorporation year.

Figure 16. Leading foreign nationalities among owners of post-2022 UK AI-named companies, with each group's median year of birth.

4.7 Geographic distribution

Geographic concentration is materially distorted by formation agents: approximately one third of AI-named companies register at a virtual-office address. After excluding a curated set of such addresses, AI density is highest in Cambridge (exceeding Oxford) and in central London. Unadjusted geographic counts should not be used.

The distortion is quantified at the level of individual officers: 76.7% of AI-named company directors list the same postcode as their company's registered office, a large share of which are formation-agent addresses. Companies House withholds directors' residential addresses, so we cannot map where owners live. But the subset of directors whose service address differs from the registered office (the only independent location signal available) distributes almost identically by region to those whose address matches it (Figure 18), with London at roughly 43% on both measures. The regional concentration is therefore robust to the virtual-office artifact, even if the clustering at specific postcodes is not. Mapped UK-wide, the same picture holds: London dominates, but real secondary clusters appear in Manchester, Birmingham, Bristol, the Oxford-Cambridge-Reading arc, and Glasgow, Edinburgh, Cardiff and Belfast (Figure 19).

Figure 17. AI-named companies per 1,000 active companies by UK postcode area, after excluding virtual-office formation-agent addresses.

Figure 18. Regional distribution of AI-named company directors, split by whether the director's service address matches the registered office (76.8% do) or differs from it. The two distributions are nearly identical, so the regional concentration is robust to the virtual-office artifact. Service addresses, England-only regions; residential addresses are withheld by Companies House.

Figure 19. UK distribution of AI-named company directors, one marker per postcode area sized by director count, positioned at the area centroid. Director service-address geocodes (residential addresses withheld).

4.8 Company survival and dormancy

Company outcomes are observable from live register status. Among cohorts old enough to have filed accounts (2018-2024), roughly 18 to 25% are dormant, that is, registered but filing dormant accounts and not trading, indicating that a substantial minority of name-signalled "AI-named companies" are not operating businesses (Figure 20). Closure rates (liquidation, dissolution, strike-off) remain modest, highest for the 2024 cohort at 8.2%. Recent cohorts are too young to exhibit failure or dormancy, and dormancy is undercounted until a first set of accounts is filed; these figures are therefore a lower bound for younger years.

Benchmarked against the entire UK register (computed exactly from the Companies House bulk data, not sampled), neither outcome is markedly AI-specific. Age-matched distress (non-active status) tracks the all-UK rate at every vintage (for example 11.5% versus 10.6% for the 2024 cohort), so there is no AI-specific failure wave. Dormancy, measured as the dormant share of companies that have filed accounts, runs only modestly above the all-UK norm, a few points per cohort (for example 22.0% versus 16.9% in 2020 and 26.7% versus 22.0% in 2023), with one cohort below; the all-UK dormancy rate itself rises with recency (from about 13% in 2018 to 22% in 2023), and AI-named companies sit slightly above it rather than at a multiple of it.

(Funding and investor-ownership analyses are deliberately omitted here: the public-filing signals for funding are still being validated and are kept as separate working notes rather than presented as settled findings.)

Figure 20. Outcome mix of each incorporation cohort (live Companies House status), normalised to 100%: trading, dormant (filed dormant accounts), and closed (no longer active). Among cohorts old enough to have filed accounts (2018-2024), roughly 18-25% are dormant; closures remain modest (2024 cohort highest at 8.2%). Recent cohorts are too young to show failure or dormancy, and dormancy is undercounted until first accounts are filed.

5. Discussion

The results describe a change in the composition of AI-named company founders coincident with, but not shown to be caused by, the diffusion of consumer generative AI. The most parsimonious reading is that the technology lowered the fixed cost of founding an AI-named company, and that the marginal entrant under the lower cost is younger, more often solo, and more often the sole owner-operator. The concentration of these firms in high-exposure occupations, and the over-representation among founders of the cohort facing slower hiring into those same occupations, is consistent with entrepreneurship operating as a margin of adjustment to automation, though we caution that this is a correlation between two independent datasets (one UK, one US) and not evidence of a causal channel.

6. Limitations

The analysis is bounded by the UK register and by self-reported fields. Companies House purges dissolved companies from the bulk snapshot, so survival rates cannot be computed from it without bias. SIC codes are self-reported and stale. Dates of birth are disclosed only to month and year, so age is approximate. Nationality is self-reported. AI-named companies are identified by name, which admits both false positives (incidental "AI") and false negatives (AI firms without a signalling name). Website classification covers approximately 76% of the subset and is excluded where low-confidence. The occupational mapping is an interpretive bridge from a firm's market to an occupation, not an exact correspondence. All estimates are descriptive.

References

We cite only the work this analysis directly builds on. The occupational AI-exposure scores used in section 4.5 come from the Anthropic Economic Index; the concept of scoring occupations by exposure to large language models originates with Eloundou et al.

Appel, R., Massenkoff, M., McCrory, P., McCain, M., Heller, R., Neylon, T., and Tamkin, A. (2026). Anthropic Economic Index report: economic primitives. Anthropic.
Eloundou, T., Manning, S., Mishkin, P., and Rock, D. (2023). GPTs are GPTs: An early look at the labor market impact potential of large language models. arXiv:2303.10130.
Handa, K., Tamkin, A., McCain, M., et al. (2025). Which economic tasks are performed with AI? Evidence from millions of Claude conversations. Anthropic.
Massenkoff, M. and McCrory, P. (2026). Labor market impacts of AI: A new measure and early evidence. Anthropic. Primary source for the occupational AI-exposure scores in section 4.5.

Data sources

Companies House (UK): bulk company data product and REST API (company register, officers, persons with significant control).
Companies Registration Office (Ireland): company register.
GDELT Project, DOC 2.0 API: global news-article volume.
Google Trends: worldwide relative search interest.

Data availability

Aggregate, chart-level data underlying this draft is available in the project repository. Person-level officer and PSC records are not redistributed, as they constitute personal data and their bulk republication would exceed the Companies House open-licence terms.