The AI competency pyramid.
Where 20.6 million UK workers actually sit. 14 distinct levels, from the people who've never knowingly used AI to the few hundred who build the models themselves.
According to the ONS, 34.31 million people are in employment in the UK. An estimated 20.6 million of them use a computer every day. That's the entire addressable market for AI in the workplace — and it splits across 14 distinct levels of AI competency, from people who've never knowingly used it to the few hundred who build the models themselves.
I mapped every level. The distribution is uncomfortable.
95% of UK computer users sit in the bottom four levels. Everything above Level 4 combined — every automation builder, every agent operator, every AI safety engineer in the country — adds up to less than 5%.
Here's what each level actually means, who's there, and why it matters.
Levels 0–4: The general workforce.
Five levels covering 98% of the UK's computer-using workforce. If you work in an office, sit in front of a laptop, or send emails for a living, you are almost certainly somewhere in this group.
Level 0 — Non-user 25% · ~5.15m people
Doesn't use AI deliberately. May encounter it via autocomplete or spam filters but distrusts or ignores it. Some actively opt out; others simply haven't been given a reason to start.
Who's here: older demographics, sceptics, heavily regulated roles, and anyone who's decided AI isn't relevant to their work. A full quarter of UK computer users.
Level 1 — Passive consumer 40% · ~8.24m people
Uses AI baked into everyday tools without thinking of it as "AI". Accepts Copilot's email suggestions in Outlook, lets Gemini summarise the top of a Gmail thread, hits tab on Grammarly's rewrites. Doesn't seek AI out — but doesn't refuse it either.
Who's here: the average UK office worker. The single largest band. This is where most of the country actually lives.
Level 2 — Casual questioner 20% · ~4.12m people
Uses ChatGPT, Claude or Gemini as a better search engine. Quick answers, definitions, explanations. Single-turn, throwaway prompts. Doesn't save chats, doesn't build on previous conversations.
Who's here: students, curious professionals, the growing mainstream who've tried it and found it useful for quick questions.
Level 3 — Practical user 10% · ~2.06m people
Feeds documents and data into AI for specific tasks. Proofreading reports, rewriting a CV, summarising a long PDF, drafting emails to a specific audience. Treats AI as a capable assistant — useful, but still a tool that does one job at a time.
Who's here: knowledge workers, marketers, admin staff, teachers. Anyone who's realised AI can save them an hour a day if they invest five minutes in a good prompt.
Level 4 — Applied reasoner 3% · ~618k people
The mindset shift. Uses AI to compare, reconcile and analyse across multiple sources. Iterates on prompts, understands context windows, extracts structured outputs. Treats AI as a collaborator rather than a tool.
Who's here: business analysts, finance teams, researchers, journalists. The first level where AI changes how you think, not just what you produce.
The jump from Level 3 to Level 4 is free. No budget, no platform, no IT approval — just a change in how you approach the machine. And yet only 3% of UK computer users have made it.
Level 5: The builders.
Where AI stops being an assistant and starts being infrastructure. Three distinct specialisms, all at the same competency level, each bringing something the others can't.
Level 5a — Automation builder 0.8% · ~165k people
Chains AI into repeatable workflows using n8n, Zapier, Make or Power Automate. Defined inputs, defined outputs, human-in-the-loop checkpoints. AI becomes a step in a pipeline, not a one-off query.
Who's here: operations specialists, technical analysts, automation leads. People who've looked at a 40-step manual process and thought "this could run itself."
Level 5b — Integration engineer 0.5% · ~103k people
Connects AI to enterprise systems via APIs, MCP servers and middleware. Manages authentication, data flow and error handling between platforms. Makes sure the AI can read the CRM, write to the data warehouse, and post to Teams without breaking.
Who's here: integration developers, solutions engineers, platform engineers. The plumbers of the AI era.
Level 5c — Security practitioner 0.3% · ~62k people
Assesses AI tools for data leakage, prompt injection, shadow IT and compliance risk. Writes acceptable-use policies, reviews vendor AI, and works out how to say "no, but..." rather than just "no."
Who's here: InfoSec analysts, GRC teams, CISOs and their reports. Arguably the most under-supplied specialism on this pyramid.
The jump from Level 4 to Level 5 is technical plumbing. APIs, authentication, orchestration. Most analysts can learn it in a quarter if they're given the time. Most aren't.
Level 6: The operators.
Where things get philosophically uncomfortable. At this level you stop telling AI exactly what to do — and start letting it work out the steps for itself.
Level 6a — Agent operator 0.2% · ~41k people
Deploys agents that plan, use tools and act semi-autonomously across systems. Comfortable with non-determinism and multi-step reasoning. Accepts that the agent might take three different paths to the same answer, and that's fine.
Who's here: forward-leaning engineering teams, AI-native startups, small cohorts inside large enterprises.
Level 6b — Red teamer 0.05% · ~8–12k people
Actively probes AI systems for weaknesses. Jailbreaks, prompt injection, data exfiltration, model misuse. Adversarial mindset applied to AI. For every organisation deploying agents at Level 6a, someone needs to be trying to break them.
Who's here: AI red teams, penetration testers, specialist security consultancies. There are barely any of them, and the demand is about to explode.
The jump from Level 5 to Level 6 is cultural, not technical. It requires tolerating non-determinism — accepting that the same input might produce a slightly different output tomorrow. This clashes with forty years of enterprise IT governance, which is precisely why most organisations stall here.
Level 7: The architects.
Designs AI-native products and platforms. AI is infrastructure at this level, not a feature.
Level 7 — AI systems architect 0.1% · ~18–24k people
Designs AI-native products and platforms. RAG pipelines, fine-tuned models, eval frameworks, observability. These are the people deciding which model, which vector database, which orchestration layer, which evaluation suite.
Who's here: ML engineers, AI platform teams, senior engineers at tech firms. A small, well-paid cohort mostly clustered in London, Cambridge, Edinburgh and a handful of scale-up hubs.
Levels 8–9: The frontier.
Fewer than 10,000 UK workers combined. Most of them could fit in a single football stadium, with room left over.
Level 8a — AI safety engineer 0.01% · ~1.5–2.5k people
Builds guardrails, alignment techniques, interpretability tools and evaluation suites. Works out why models behave the way they do and how to make them behave better. Answers the question "is this model safe to deploy?"
Who's here: safety teams at Anthropic, OpenAI, DeepMind and the UK AI Safety Institute. The stakes of their work are enormous. The headcount is not.
Level 8b — Frontier problem solver 0.02% · ~3–5k people
Uses AI to attack previously unsolved problems. Drug discovery, protein folding, mathematical proofs, rare disease diagnosis, climate modelling. Treats AI as a research collaborator on problems that matter at a civilisational scale.
Who's here: Isomorphic Labs, DeepMind's science teams, academic labs, specialist medical research groups.
Level 9 — Model researcher 0.005% · ~800–1.2k people
Builds and pushes the frontier itself. Trains foundation models, develops new architectures, advances capability and alignment research. Writes the papers everyone else reads three weeks later.
Who's here: a few hundred UK-based researchers spread across Anthropic, DeepMind, Meta FAIR, Mistral, and top universities. If you assembled every Level 9 researcher in the UK into one room, it would fit comfortably in a corporate townhall.
What the numbers actually tell us.
Let's return to the statistic I opened with: 95% of UK computer users sit at Levels 0–3.
That's 19.57 million people. It's more than the populations of London, Birmingham, Manchester and Leeds combined. And almost none of them are getting meaningful value from AI beyond auto-complete and summarisation.
Meanwhile, everything above Level 4 combined — the entire builder, operator, architect, safety and researcher cohort — is about 1 million people. Less than 5% of the workforce. Of those, the people actually designing, safeguarding or advancing AI at the frontier number in the low thousands.
Three things jump out when you sit with the distribution.
First, the 3-to-4 jump is free and almost nobody is taking it. The shift from "AI as tool" to "AI as collaborator" costs nothing. No licences, no platforms, no approvals. Just a change in how you approach the machine. And yet 97% of computer-using workers haven't made it.
Second, the 5-to-6 jump is where enterprise transformation projects die. Not because the technology isn't ready, but because the culture isn't. Forty years of deterministic IT governance — "the system must produce the same output for the same input" — collides head-on with the reality of autonomous agents. Most organisations discover this the first time an agent takes an unexpected route to a correct answer.
Third, the safety cohort is dangerously small. For every ten thousand people deploying AI at work, there are roughly three people whose job is to make sure AI behaves safely at scale. That ratio doesn't inspire confidence.
The strategic question.
If you're a leader, here's the question that should keep you awake: which level is your organisation operating at, and which level do you need it to be at in two years' time?
For most UK enterprises, the honest answer is "Level 2 or 3, and we need to be at Level 5 by 2027." That's a two-level leap across a workforce of hundreds or thousands. It's not a training problem. It's a mindset, infrastructure and culture problem all at once.
The organisations that pull it off won't be the ones that buy the most AI tooling. They'll be the ones that deliberately move their median employee one level up the pyramid every year — from passive consumer to casual questioner, from questioner to practical user, from practical user to applied reasoner. Small shifts, compounded, across tens of thousands of people.
That's the game. And right now, most UK organisations aren't even playing it.