ChatGPT changed who can start an AI-named company
Preliminary insights for Jimmy McLoughlin / Jimmy's Jobs of the Future Matt Cortland, Prime Directive AI Ltd 2026-05-28
TL;DR
UK Companies House data shows a structural break in 2023, the year ChatGPT went mainstream. Before 2023, UK AI founders were older than the average UK entrepreneur, started companies in teams of 2+, and looked like PhDs or experienced ML engineers. After 2023, AI founders are younger than the average UK entrepreneur, start solo, and own and run the company themselves. The under-26 share of UK AI founders has tripled since 2018. 95% of UK AI startups are now founder-led: one person on both the directors register and the persons-with-significant-control register. The AI-named ecosystem is about 14x the size of the crypto-named ecosystem on the register today (roughly 9,100 companies versus 655), and unlike crypto it is still growing. The Gen Z entrepreneurship story has a mechanism behind it: ChatGPT collapsed the cost of starting an AI-named company, and a generation of Gen Z founders walked through that door. The 2025 wave of agentic coding tools (Claude Code, Cursor) appears to be amplifying that effect: 2025-2026 AI cos are even more solo, even younger, and even more concentrated in pure software, with marketing/advertising as the only meaningfully expanding non-software vertical.
Key findings
- ChatGPT was a structural break in who founds UK AI-named companies. Formations climbed sharply from 2023 (the apparent flatness of 2015-2021 is partly survivor-undercounting of older years in the live register, which omits since-dissolved companies). On the current run-rate, 2026 will exceed 2025 (about 4,800 versus 4,073 on a survivor basis). The rise holds honestly as roughly a 6x increase in AI's share of all new companies.
- The new founders are younger and solo. The under-26 share of AI founders tripled (4.1% to 12.8%, 2018-2025), the solo-founder rate reached 82.9% in 2026, and 95% of AI-named companies incorporated since 2020 are founder-led: the same person is on both the owners (PSC) and directors registers.
- Gen Z is over-represented in UK AI ownership by 2.2x relative to all UK companies.
- Britain is renaming itself into AI. 300 existing companies changed their name to add "AI" in 2025, up sharply from 2018 (the precise multiple is inflated by survivor-undercounting of older years), with roughly 370 projected for 2026. Companies that pivot in are more cross-sector than companies founded as AI from the start.
- AI-named companies are building for the occupations AI most automates. Every company vertical that maps cleanly to an occupation lands on a high-exposure white-collar job (around 65% average on Anthropic's 2026 exposure scale): marketing, programming, financial analysis, medical records, security. The supply side of AI mirrors the demand side.
- The founder base is increasingly international. 36% of post-2022 AI-named company owners hold a foreign nationality, up from about 31% before 2022, and the foreign founders skew younger than British ones.
- Geographic concentration is real, but only after filtering virtual-office formation agents (about a third of AI-named companies register at one). Cambridge is denser than Oxford.
- The AI-named ecosystem is about 14x the size of the current crypto-named ecosystem (roughly 9,100 companies versus 655 on the register today), and unlike crypto, it is still growing.
1. The headline finding
ChatGPT was the structural break in who founds UK AI-named companies.
Two independent measures show the same break in the same year (2023):
Age of UK AI founders, by company-incorporation year
The under-26 share of AI founders went from 4.1% in 2018 to 12.8% in 2025 (a 3.1x increase). The under-30 share went from 11.1% to 23.9% (a 2.2x increase). Over the same period, the under-26 share of all-other-UK founders grew from 5.9% to 10.6% (a 1.8x increase). Both populations are getting younger, but AI is getting younger faster, and the crossover from "AI founders older than baseline" to "AI founders younger than baseline" happened in 2023.
| Year | AI under-26% | AI under-30% | All-other-UK under-26% | All-other-UK under-30% |
|---|---|---|---|---|
| 2018 | 4.1% | 11.1% | 5.9% | 14.4% |
| 2019 | 5.8% | 16.0% | 6.4% | 15.0% |
| 2020 | 3.6% | 11.9% | 6.5% | 14.8% |
| 2021 | 5.2% | 13.4% | 6.4% | 14.7% |
| 2022 | 5.7% | 17.3% | 6.6% | 14.7% |
| 2023 | 8.6% | 19.7% | 7.3% | 15.5% |
| 2024 | 10.3% | 19.6% | 8.7% | 17.7% |
| 2025 | 12.8% | 23.9% | 10.6% | 20.8% |
Sample: 8,176 AI founders with date of birth across the period, 4.4M control founders.
Solo-founder rate, by company-incorporation year
In 2018, 65.5% of new UK AI-named companies had exactly one founder (one person with significant control, notified within 60 days of incorporation). The rest had teams of 2-3 academic or technical co-founders. By 2026 the solo-founder share is 82.9%. The average AI startup is now a one-person operation: one person owns it, runs it, and is the only name on both the directors register and the PSC register.
| Year | AI solo-founder% | Control solo-founder% |
|---|---|---|
| 2018 | 65.5% | 71.7% |
| 2019 | 67.0% | 71.6% |
| 2020 | 69.1% | 69.4% |
| 2021 | 73.0% | 69.6% |
| 2022 | 69.1% | 70.1% |
| 2023 | 74.7% | 70.9% |
| 2024 | 74.4% | 74.9% |
| 2025 | 79.9% | 78.4% |
| 2026 (partial) | 82.9% | 81.6% |
Both populations are converging on "solo founder is the default." AI got there faster.
The mechanism
Before ChatGPT, starting an AI-named company meant assembling a team that could train or fine-tune a model. That required a PhD, deep technical skills, and usually a co-founder. The data shows it: pre-ChatGPT UK AI founders were older and more team-based than the average UK entrepreneur.
After ChatGPT, a 22-year-old can spin up an AI product alone with a credit card and an API key. The data shows that too: post-ChatGPT UK AI founders are younger and more solo than the average UK entrepreneur. The technology lowered the floor on entrepreneurship, and a generation of Gen Z founders walked through the door.
The 2025 arrival of agentic coding tools (Claude Code, Cursor's wider adoption) appears to be amplifying that pattern. 2025-2026 AI cos are even younger (under-26 share 12.8% in 2025), even more solo (82.9% solo-founder in 2026 partial-year), and even more concentrated in pure software (78.1% in 2026 vs 56.1% in 2018). The same mechanism, accelerated. See section 2.4 for the vertical-mix shift and the candidate explanations.
This is the Gen Z entrepreneurship story Jimmy's audience is asking about, with a mechanism attached.
2. Eight supporting findings
2.1 Gen Z owns AI at 2.19x the baseline, and runs it themselves
Looking at the full UK persons-with-significant-control register (15.5 million records, the entire population not a sample), people born in the 2000s own UK AI-named companies at 2.19x the rate they own other UK companies. 9.4% of AI-named company owners are Gen Z, vs 4.3% of other UK company owners.
| Birth decade | AI owners % | All-other-UK owners % | AI ratio |
|---|---|---|---|
| 1950s and earlier | 2.8% | 10.5% | 0.27x |
| 1960s | 9.6% | 17.5% | 0.55x |
| 1970s | 21.2% | 22.8% | 0.93x |
| 1980s | 30.9% | 26.4% | 1.17x |
| 1990s | 26.0% | 18.5% | 1.41x |
| 2000s (Gen Z) | 9.4% | 4.3% | 2.19x |
Extending the lens to under-36s: 35% of UK AI ownership sits with people born 1990 or later, vs 23% of other UK ownership.
Crucially, the Gen Z signal is not "young owners with hired professional boards." When we matched directors to PSCs by surname and date of birth, 95.4% of UK AI-named companies founded since 2020 have the same person on both registers. Of 816 AI-named companies with a Gen Z owner, 764 (93.6%) have a Gen Z person who is also the only director. Young founders are doing both jobs.
2.2 Britain is in a once-in-a-generation company-naming wave
In 2025, 283 UK companies rebranded their existing business to include AI in the name, up sharply from the 2018 baseline (the precise multiple is overstated because older years omit since-dissolved companies). These are existing operating businesses repositioning into AI in real time, not new entrants. The AI-named ecosystem as a whole (companies with AI-related keywords in their current or historical names) is about 14x the size of the crypto-named ecosystem on the register today (roughly 9,100 companies versus 655), and unlike crypto it is still growing. (We compare current surviving cohorts on both sides; an earlier draft cited ~32x against crypto's eroded 2021 peak, which overstated the gap.)
| Ecosystem | UK companies on register today | Status |
|---|---|---|
| AI-named | 9,137 | still growing |
| Crypto-named | 655 | peaked ~2021, now declining |
| General tech buzzwords | rising | still growing |
2.3 The registration codes say "software"; the websites say "whole economy"
UK companies pick up to 4 SIC (Standard Industrial Classification) codes at incorporation. For AI-named companies, the dominant codes are pure-IT: 38% choose 62012 (business software development), 28% choose 62020 (IT consultancy), 19% choose 62090 (other IT services). About 75% of AI-named companies have at least one of the pure-IT 62xxx codes.
The interesting story is in the long tail: how many AI-named companies also tag a vertical-industry SIC alongside (or instead of) the pure-IT one. This is the data on where AI is being applied across the UK economy.
| Industry section | AI-named companies with a SIC in this section (excludes pure-IT) |
|---|---|
| Media / advertising | 1,045 |
| R&D / other professional | 903 |
| Management consulting | 633 |
| Retail / wholesale | 466 |
| Education | 348 |
| Construction / real estate | 211 |
| Finance / fintech | 200 |
| Healthcare / pharma / biotech | 198 |
| Manufacturing | 117 |
| Transport / logistics | 48 |
| Legal | 38 |
| Energy / agriculture | 10 |
The clearest vertical concentrations are media/advertising (likely AI marketing tools, content generation, ad optimization), management consulting (AI strategy advisors, integration consultancies), and retail (personalization, e-commerce optimization). Healthcare and finance are smaller in absolute count but high-value: the same 198 healthcare AI-named companies includes diagnostics, mental-health tools, and patient-data platforms.
Methodology caveat: SIC codes are self-reported at incorporation and rarely updated. A company that started as "AI consultancy" and pivoted to "AI legal software" probably still files under 62020. This is a snapshot of what founders said they would do, not necessarily what they do today. The cross-industry counts therefore understate the true degree of horizontal AI adoption.
The truer cross-section, from the companies' own websites. Because the SIC codes understate horizontal spread, we also classified each AI-named company by what its website actually describes, not the code it filed. On that basis the "software" concentration largely dissolves. The single largest category, developer tools and infrastructure, is only 13.6% of classified companies; the rest spread right across the economy:
| What the company actually does (website-classified) | Share of classified AI cos |
|---|---|
| Developer tools / infrastructure | 13.6% |
| Services / agency | 12.1% |
| Fintech / finance | 10.7% |
| Marketing / advertising tech | 9.3% |
| Healthcare / pharma | 7.6% |
| Retail / e-commerce | 7.1% |
| Education | 6.7% |
| Content creation | 6.4% |
| Security / cyber | 3.9% |
| Manufacturing / industrial | 3.9% |
| HR / talent | 3.8% |
| Real estate / proptech | 2.8% |
| Legal | 2.2% |
| Gaming / entertainment | 2.1% |
The registration code is where you incorporate; the website is what you actually built. By the codes, AI looks like a software industry. By the websites, it is a horizontal layer spreading into finance, marketing, healthcare, retail, education, and beyond, with no single vertical above 14%. (Classification: website scrape plus an LLM classifier, high-confidence classifications only; about 9,600 AI-named companies have a usable website. This is the basis used for the AI-exposure analysis in 2.7.)
2.4 The vertical mix is narrowing, not broadening, and the post-Claude-Code era amplifies that
We split the SIC-vertical analysis above by incorporation year to test a specific hypothesis: did the arrival of agentic coding tools (Claude Code in early 2025, Cursor's rapid adoption from late 2024) make it easier for domain experts in healthcare, legal, finance, etc. to ship AI products in their fields, broadening the vertical mix?
The data says no.
Share of each year's UK AI cos that tag a pure-IT / software SIC:
| 2018 | 2019 | 2020 | 2021 | 2022 | 2023 | 2024 | 2025 | 2026 |
|---|---|---|---|---|---|---|---|---|
| 56.1% | 56.9% | 59.7% | 47.2% | 59.6% | 64.6% | 70.6% | 72.4% | 78.1% |
The trend is the opposite of the broadening hypothesis. UK AI cos are more concentrated in pure software over time, not less. By 2026, 78% of new AI cos tag a pure-IT SIC, up from 56% in 2018.
Among non-software verticals, the picture splits:
| Section | 2018 | 2022 | 2025 | 2026 | Trend |
|---|---|---|---|---|---|
| Media / advertising | 1.4% | 2.1% | 7.5% | 6.3% | Rising sharply |
| Education | 2.7% | 2.1% | 4.9% | 3.9% | Rising |
| Management consulting | 7.4% | 8.2% | 8.1% | 8.5% | Flat-rising |
| Retail / wholesale | 3.4% | 6.4% | 6.7% | 3.6% | Spiked then dropped |
| Finance / fintech | 2.7% | 4.3% | 2.0% | 1.5% | Shrinking |
| Healthcare / pharma / biotech | 2.7% | 2.8% | 1.5% | 1.4% | Shrinking |
| Legal | 0.7% | 0.7% | 0.4% | 0.3% | Flat-shrinking |
| Construction / real estate | 6.1% | 5.7% | 1.8% | 0.9% | Shrinking sharply |
| Manufacturing | 5.4% | 1.8% | 1.3% | 1.2% | Shrinking |
So: AI is concentrating in pure software, with one major exception (marketing/advertising, which rose from 1.4% in 2018 to 7.5% in 2025). The regulated verticals (healthcare, legal, finance) are flat or shrinking in relative share, despite the technology becoming cheaper to deploy.
Possible explanations (we are not certain which is dominant):
-
SIC labelling effect. Founders default to "62012 business software development" at incorporation regardless of the eventual application, because at day one the product is a piece of software. A legal-AI tool may be SIC-tagged as software, not legal, even though it serves the legal market. Under this explanation, the vertical concentration is real but partially obscured by founder-side defaults.
-
Marketing has no regulatory barrier. Marketing and advertising are uniquely accessible because there is no professional licensing, no procurement gate, and clear pain points (paid acquisition is expensive). So the marginal new AI founder who can now ship software with less effort goes to the easiest paying market.
-
Regulated verticals have non-technical bottlenecks. In healthcare (MHRA approval, NHS procurement), legal (SRA regulation, professional indemnity), and finance (FCA regulation), the bottleneck has never been "we can't build the software." It has been getting the product approved, sold, and adopted. Lower software-building costs don't relax those constraints, so vertical AI startup formation in those sectors doesn't accelerate.
-
Survivor bias in the data. Healthcare/legal/finance AI cos might just be incorporating under different names that don't include "AI" as a keyword. A telemedicine company built around an LLM might call itself "Acme Health" rather than "Acme Health AI." This would understate vertical AI formation in our data.
-
Sample period limitation. Claude Code launched in early 2025 and adoption took months. The 2026 data through April may be too early to show its effect.
We surface these as candidate explanations, not as a settled conclusion. What the data definitively shows: the vertical mix of UK AI startups is narrowing toward pure software, with marketing/advertising as the only meaningfully rising non-software vertical. The "why" needs a follow-up cut (perhaps tracking the same companies via SIC reclassifications over time, or interviewing founders, or scraping company websites to compare self-described vertical against SIC).
2.5 Geographic concentration is real, but you have to filter for virtual offices first
23% of all UK companies (and 34% of UK AI-named companies) are registered at virtual office addresses operated by formation agents like 1st Formations and Hoxton Mix. Any naive geographic map of AI activity will be dominated by these mailbox locations rather than where the work happens.
After filtering 10,364 known formation-agent addresses, the real UK AI clusters are:
| Rank | Postcode area | AI cos per 1,000 cos |
|---|---|---|
| 1 | EC (City of London) | 3.26 |
| 2 | CB (Cambridge) | 2.96 |
| 3 | WC (Bloomsbury / Holborn) | 2.75 |
| 4 | OX (Oxford) | 2.04 |
| ... |
Cambridge is denser than Oxford for AI, and Bloomsbury (WC) is denser than the City of London if you count just first-party-address companies. We have 50 per-constituency story sheets ready, each carrying its MP, top AI-named companies, formation date trend, and how the constituency compares to the national rate. Any local outlet can plug in their patch.
2.6 The naming culture sidebar
A bonus look at how Britain's 8,437 AI-named companies are named. Out of every ten new UK AI-named companies:
- ~7 use the pattern "X AI Ltd" (e.g. GECKO AI LTD). The space-suffix is the workhorse.
- ~1-2 use the pattern "AI X Ltd" (e.g. AI LABS LTD). The prefix form. Tends to indicate consultancies, services businesses, or generic-AI plays.
- ~1 uses the pattern "X.AI Ltd" (e.g. NETIX.AI LTD). The brand-form that early AI darlings (Stability.ai, Character.ai, X.ai) made fashionable.
Trend over time: the X.AI brand form is fading. It was 17.6% of UK AI cos formed in 2018, now down to 7.4% of 2026 cos. Founders are gravitating to the mundane "Foo AI Ltd" rather than the trademark-style dot. Reading: AI naming is normalising.
The "Agent" tell. Until 2024, virtually no UK AI-named company name contained the word "Agent" or "Agents". In 2025, sixteen did. Three more in early 2026. This is the cultural arrival of "agentic AI" showing up in incorporation paperwork inside one calendar year.
The top distinctive words in UK AI-named company names (excluding AI, Ltd, Limited, and common stopwords):
| Rank | Word | Count |
|---|---|---|
| 1 | labs | 117 |
| 2 | tech | 97 |
| 3 | systems | 73 |
| 4 | digital | 58 |
| 5 | global | 56 |
| 6 | data | 53 |
| 7 | business | 51 |
| 8 | London | 43 |
| 9 | studio | 39 |
| 10 | marketing | 38 |
The vocabulary mirrors the era. "Labs" signals research credibility. "Studio" signals creativity. "Systems" signals B2B. "Data" appears 53 times in just AI-keyword names, double-tagging the technical positioning.
Mythology and sci-fi names are alive: Nexus (17 cos), Nova (10), Apex (8), Atlas (6), Athena (6), Titan (4), Genesis (4), Matrix (3), Synapse (3), Orion (3). Naming an AI-named company after a Greek deity or a Star Trek concept is still in fashion.
Animals, in declining order of popularity: Dragon (3), Phoenix (3), Eagle (2), Robin (2), Falcon (2), Tiger (2), Dolphin (2). A pronounced bird-of-prey tilt (Eagles, Falcons, Hawks, Herons, Ravens). Birds suggest watching, soaring, oversight: an iconography fit for AI-named companies signalling "we are above the work."
This is the kind of sidebar a Sunday Times tech feature can lift directly. Datawrapper-ready CSV at data/charts/11-ai-name-words.csv.
2.7 AI-named companies are forming around the occupations AI most automates
In March 2026, Anthropic published a labour-market study ranking occupations by "observed exposure" to AI, how much of each job's work AI is actually being used to do. Computer programming topped it (74.5%), followed by customer service (70.1%), data entry (67.1%), medical records (66.7%), and market research / marketing (64.8%).
We mapped each UK AI-named company to the occupation its product targets (using its website, not its self-reported SIC code) and matched that to Anthropic's exposure score. Of the companies that map cleanly, every single vertical lands on a high-exposure white-collar occupation, and the weighted-average exposure is about 65%:
| What the company builds for | Companies (post-2022, classified) | Anthropic AI-exposure |
|---|---|---|
| Market research / marketing | 568 | 64.8% |
| Computer programmers | 490 | 74.5% |
| Financial analysts | 389 | 57.2% |
| Medical records | 276 | 66.7% |
| Information security | 141 | 48.6% |
The plain-English reading: the new AI-named companies are overwhelmingly being built to do the exact work AI is best at automating. Anthropic measures AI's effect on existing jobs (the demand side); this measures where AI entrepreneurship is concentrating (the supply side). They point at the same occupations. And the young cohort Anthropic finds facing slightly slower hiring into exposed jobs is the same cohort we find over-represented as founders, one door narrowing as another opens.
Caveats: a company's vertical is a rough proxy for an occupation (a fintech AI tool might serve analysts or consumers), so this is a directional bridge, not an exact occupational overlay; 56% of companies do not map to a single occupation and are excluded here; and we use only high-confidence website classifications. The direction, however, is strong and one-way: toward the most-exposed work.
2.8 UK AI is increasingly founded by foreign nationals
Among AI-named companies incorporated in 2023 or later, 35.7% of owners hold a foreign nationality (64.3% are UK nationals, merging British / English / Scottish / Welsh / Northern Irish). The foreign share has risen from about 31% before 2022 to the high-30s since. The leading foreign nationalities are Indian, Chinese, and Pakistani, followed by American, Italian, and Nigerian.
The international founders also skew younger: median birth years around 1991-1993 for Indian, Chinese, and Pakistani founders, versus 1982 for British founders. The youth story and the international story overlap, UK AI is being built disproportionately by young, internationally-born founders.
Caveat: this is the nationality of the owners of UK companies, most of whom are UK-resident. It reflects who is building companies here, not migration flows, and nationality is self-reported on Companies House filings.
3. What this could become
3.1 Podcast angle
The "Gen Z is the most entrepreneurial generation" hypothesis is supported by the data, but the more interesting story is why now. ChatGPT didn't just change what AI-named companies do, it changed who can start one. An episode framed around "the falling cost of starting a tech company, and what that means for the next generation of UK entrepreneurs" lets you use the data as a hook into a forward-looking conversation about which other industries are next.
3.2 Press release angle
Three numbers that hit hard:
- The under-26 share of UK AI founders has tripled since 2018.
- 83% of new UK AI-named companies are now founder-led, up from 66% in 2018. One person owns and runs the company.
- The UK AI-named ecosystem is about 14x the size of the current crypto-named ecosystem, and unlike crypto, it is still growing.
National outlets to target: FT (tech and labor markets), Times (entrepreneurship), Guardian (Gen Z and economy), BBC News (regional angle via constituency sheets).
3.3 Local press angle
The 50 per-constituency story sheets (in docs/constituency-sheets/) are pre-built for distribution. Each one carries the MP, the constituency's AI-named company count, the rebrand wave count, the top three AI-named companies by registered address, and how the constituency compares to the national rate. Pitch to Metro, regional ITV, BBC Local, Reach plc regional press.
3.4 Shareable graphics
- The naming wave. Monthly AI-named company formations 2018-2026 with the ChatGPT launch (Nov 2022) annotated. The kink in the line is the story.
- The structural break. Two-series line chart: AI under-26 founder share vs all-other-UK under-26 founder share, 2018-2026. The crossover at 2023 is the visual.
- AI vs crypto vs tech. Stacked bar by year, 2014-2026. Crypto's bust is visible. AI's growth dwarfs it.
- Solo-founder rate. Two-series line: AI vs all-other-UK, 2018-2026. The convergence-from-above is the visual.
- Geographic map. UK density of AI-named companies per 1,000 companies, formation-agent-filtered, at postcode-area or constituency level.
Datawrapper-ready CSVs for all of these live in data/charts/.
3.5 Interactive one-pager
Optional but high-leverage: a public lookup at jobsindex.co.uk where any user enters their postcode and sees their local AI ecosystem stats (companies, founder ages, top firms). Drives organic traffic. Makes the data feel personal. Approximately one week of build time after the deliverable.
4. Methodology
This section explains the data, the comparisons, the filters, and the limitations, so you can read every number above with full context.
4.1 Data sources
| Source | Coverage | Notes |
|---|---|---|
| Companies House BasicCompanyData bulk export | 5,696,442 active UK companies | 2026-04-01 snapshot. Reproducible from download.companieshouse.gov.uk. |
| Irish CRO bulk export | 812,270 Irish companies | For the cross-border comparison. Smaller role in this deliverable. |
| Companies House PSC bulk snapshot | 15,489,532 persons-with-significant-control records | 2026-05-02 snapshot. Includes all currently-active and historical PSCs. |
| Companies House officer data (via API) | 599,707 officer records | Fetched for 62,000 AI-keyword-matched companies plus a 30,000-company stratified control sample. Full UK officer fetch (5.6M companies) is running in the background under launchd. |
| ONS postcode-to-constituency lookup | 878,800 postcodes | For mapping company registered addresses to current 2024 parliamentary constituencies. |
| Formation-agent address index | 10,364 known virtual-office addresses | Derived from address-occurrence patterns. Any address with 100+ registered companies is flagged. |
All data is reproducible. The infrastructure lives in ~/code/jobs-index/ and can be re-run end-to-end from the bulk files.
4.2 How we define "AI-named company"
We use word-boundary regex matching on the current company name and any historical previous names. The "AI" search captures AI as a standalone token (so "AI Labs Ltd" matches, but "AID International Ltd" does not). We also track 9 related terms (artificial intelligence, GPT, LLM, machine learning, deep learning, generative AI, Copilot, AGI, agent) and combine them into an AI ecosystem total. For the AI keyword specifically, we found 10,159 UK companies; for the broader ecosystem, the count is larger.
This definition has known limitations. It misses AI-named companies with names that don't reference the technology (probably a meaningful share of mature AI-named companies). It also picks up some false positives (companies with AI in the name that aren't operationally AI, e.g. an immigration consultancy named "AI Legal"). We have not corrected for false positives in this deliverable because the magnitudes of the findings are large enough to absorb the noise.
4.3 How we define "founder"
For the PSC analysis (who owns the company), a "founder" is a person with significant control whose notified_on date is within 60 days of the company's incorporation date. This restriction is important because many companies acquire later investors who become PSCs after incorporation. We want the founding ownership team, not the post-investment cap table.
For the director analysis (who runs the company), a "founder director" is a director whose appointed_on date is within 60 days of incorporation.
4.4 The control sample, and why it required apples-to-apples filtering
We initially built a 30,329-company control sample by stratified random selection across all UK companies, with approximately 250 companies per incorporation year drawn at random. The intent was to ensure every era of UK company history was represented. The effect is that the control sample is wildly older than the population, with 78% of control companies being pre-2000 incorporations.
AI-named companies, by contrast, are 73% from the 2024-2026 cohort.
A direct comparison between "all AI directors" and "all control directors" produces a misleading 17x Gen Z over-representation, because most of that ratio is just "AI-named companies are newer companies, and newer companies have younger people in them." For any age-sensitive comparison, we restrict both sides to a matched incorporation window (usually 2020-2026). This drops the director-level Gen Z over-representation to ~1.67x, which is the honest number.
The PSC analysis does not have this problem because PSCs are required to be currently in control, and we compare against the full UK PSC population (8.24 million records). The 2.19x figure is a clean apples-to-apples ratio with no age confound.
4.5 The formation-agent filter, and why every geographic claim has to apply it
23% of UK companies and 34% of UK AI-named companies are registered at virtual office addresses operated by formation agents. These are not the working addresses of the businesses. They are mailbox services that founders use for incorporation, tax registration, and statutory mail. Examples include 1st Formations (71-75 Shelton Street, WC2H 9JQ) and Hoxton Mix.
Any geographic map of UK company activity that doesn't filter formation agents will show concentrations at these mailbox addresses rather than at the places where work actually happens. We identified 10,364 such addresses by flagging any address with 100+ registered companies and joined the resulting list to every geographic analysis. The Cambridge-beats-Oxford finding survives the filter. The earlier "Hereford has unusually high AI density" finding does not (Hereford was a formation-agent artifact).
For transparency: Prime Directive AI Ltd's own registered office (71-75 Shelton Street, 1st Formations) is in our flagged list. The formation-agent finding is a transparency-of-the-data point, not a gotcha.
4.6 What we don't have yet
- Employee counts. Companies House does not publish headcount. We will not be able to say "this AI-named company has 5 employees" from this dataset alone.
- Revenue or funding data. Not in the bulk export. We have not yet integrated Crunchbase, Beauhurst, or Companies House full-accounts filings.
- Company websites. Not in the bulk export. Building this is the next infrastructure milestone.
- Sector / SIC code accuracy. Companies self-report a SIC code at incorporation. Many AI-named companies file under 62012 (software development) or 70229 (management consulting) regardless of what they actually do.
- Saturated director-level demographic comparison. The current director-level Gen Z ratio (1.67x age-matched) uses a 1,701-company control. A larger control sample is currently being fetched in the background under launchd and will be available within 7-12 days. This will tighten the director-level numbers but is not expected to change the direction.
4.7 Anti-bias commitment
This analysis was run before any preferred narrative was selected. The PSC and officer probes were specified ahead of seeing the results. When the initial preliminary officer sample (21% of the priority-10 queue) suggested Gen Z was barely present in AI direction (0.4%), we reported that, even though it contradicted the working hypothesis. When the full director and PSC data later showed a 2.19x Gen Z over-representation in ownership, we updated. We will not back-engineer the analysis to confirm a desired finding, including yours. If the deliverable surprises you, that is the data, not a framing choice.
4.8 Data availability
Source data is Companies House bulk products (company snapshot, persons-with-significant-control snapshot) plus the Companies House public REST API for officer appointments, all under the Companies House open licence. Derived, aggregate, chart-level data (the figures behind every chart in this report) is published in the project repository. Person-level officer and PSC records (names, dates of birth, addresses, nationality) are not redistributed: they are personal data, and bulk republication would exceed the spirit of the open licence and raise data-protection concerns. Any future open dataset (for example on Hugging Face) will contain aggregates only.
5. Next steps
Suggested follow-up call: 45-60 min, anytime in the next two weeks.
Agenda:
- Lock the headline framing (the "ChatGPT changed who can start" angle vs the "Gen Z entrepreneurship" angle; they are compatible but differ in emphasis).
- Lock the press release timing and target outlets.
- Decide podcast structure: single deep-dive episode, or a series.
- Decide which shareable graphics get prioritized.
- Co-bylines and credit handling.
- Local press distribution: national outlets first, or simultaneous push including the 50 constituency sheets.
- Public-site decision: is jobsindex.co.uk worth standing up alongside the press push, or after?
All data, queries, and reproducible artifacts live in ~/code/jobs-index/.
Headline candidates with full scoring: docs/headline-options.md (Candidate C is the current lead).
Auto-regenerated full findings: docs/findings-v2.md.
Datawrapper-ready chart CSVs: data/charts/.
Per-constituency briefs: docs/constituency-sheets/.
30-min check-ins on the running data fetch: docs/p100-progress.md.