views

Introduction

A recent news report revealed economic productivity in the UK has remained stagnant since 2008 despite the AI “revolution”. The report explored the idea that AI has failed to support growth in the way it was expected to. Long-run economic data available debunks this stagnation, showing an ever-so-slight increase, but fails to explain the reason behind it. This article asks why AI hasn’t delivered, and what would fix it.

First up, let’s look at the data:

Productivity and technology waves

Dashed lines mark the year each technology first reached commercial or public use. On the log scale, a steeper line means faster growth. Views use different sources and units, so compare shapes within a view rather than levels across views.

The chart shows a long delay between a technology arriving and paying off. Why the delay? What happens in the period between a technology’s introduction and the reaping of its benefits? The term luddite comes to mind: A group of people that weren’t irrational - just unfortunate enough to live in the period between a new technology’s arrival and society learning to use it. This is where the costs land, long before the benefits turn up.

The luddites were skilled textile workers who broke machines from 1811 to about 1813. Their complaints were specific: wage cuts, unskilled workers doing skilled work, cheap low-quality goods damaging their reputation and being turned into factory “hands”. Croppers, for example, lost their work to the shearing frame within about a decade. The state’s response was severe. The 1812 Frame Breaking Act made machine-breaking punishable by death, and the British government sent more troops to the North of England than were fighting Napoleon in Spain. These Luddites weren’t afraid of machines. They were reacting to a real loss.

The economic record backs up The Luddites. The economic historian Robert Allen calls it “Engels’ pause”. Between 1780 and 1840, output per worker in Britain rose about 46%, but real wages rose only about 12%. Workers then caught up and overtook: from 1840 to 1900, output per worker rose 90% and real wages 123%. The gains came, just a generation or two late for the people who paid for them. The above data shows the same delay. UK output per hour was 3.88 in 1760 and 3.91 in 1820, essentially flat for ~50 years after Watt’s steam engine and it didn’t double until 1871.

The Gap Between Introduction and Benefit

There are three reasons for delayed benefit following a technological introduction:

Industry has to reorganise around the technology. Paul David’s classic study of electrification found that electric motors supplied under 5% of US factory power in 1899 and only about half by around 1920. Factories first swapped a steam engine for one big electric motor and kept the old layout. The productivity payoff came only after they redesigned the whole floor around a motor on each machine. US manufacturing productivity growth then jumped by about 5 percentage points in the 1920s, roughly 40 years after Edison’s first power stations.
People have to learn. Skills come from experience and education, and education takes a generation. It runs on a generation’s timescale. Claudia Goldin and Lawrence Katz call this the “race between education and technology”. When schooling keeps up, gains spread widely; when it lags, they go to the few who already have the skills.
Firms make invisible investments first. Erik Brynjolfsson and colleagues describe a “productivity J-curve”. Companies spend years on training, new processes and reorganisation that official statistics don’t count as output. So measured productivity can stall, or even dip, before it rises.

The Luddite Window

I’m going to coin a new phrase here. Let’s call the delay between introduction and benefit from a new technology The Luddite window.

Luddite Window n.

The period of time in which costs are experienced by workers displaced by a revolutionary technology before economic benefits are felt.

At the microeconomic level, resistance is a rational response to a new technology, not ignorance. Education isn’t only productivity policy. It also holds society together through the transition. It’s also what keeps a technological transition from tearing at society. A shorter learning lag brings the payoff forward and narrows the window for resistance.

Applying The Luddite Window to AI

The Luddite Window may be shorter for this revolution. AI is software, so firms don’t have to rebuild factories, and it can even help teach people how to use it. The bottleneck shifts almost entirely to human capability and the early costs are already visible: the Bank of England found vacancies in the most AI-exposed jobs fell 15% over three years. That’s the modern version of the Luddites’ complaint.

Some economists argue that part of the lag is a measurement problem, because statistics miss early gains. Others say the delay is about capital and infrastructure more than skills. Both may be true, but learning is the part policy can speed up.

What AI Means for the UK Economy

AI will not fix Britain's productivity problem on its own. It will amplify the capabilities people already bring to it, and right now those capabilities are unevenly distributed.

After fifteen years of stagnation, UK productivity finally appears to be recovering. The Resolution Foundation estimates that output per hour grew by around 1.1% a year in the two years to June 2026, after falling in the two years before. It is tempting to credit artificial intelligence. Adoption is rising fast, and the Bank of England has found early signs that firms building and using AI are becoming more productive.

Yet nobody can prove AI is driving the recovery. The Resolution Foundation found the improvement was broad-based, achieved by the same workers in the same jobs and sectors. BCG concluded that there is still no clear link in UK data between AI adoption and productivity growth.

This article argues that the missing link is human capability. AI is not a machine you switch on to receive productivity. It is a collaborator that returns roughly what you are able to ask of it. If we want AI to lift productivity across the whole economy, we have to invest in the people using it: their ability to know what they want, express it clearly, and judge whether what comes back is any good.

Who is getting the most from AI

The early gains are flowing mostly to people and sectors that were already close to the technology. That pattern shows up at every level of the evidence.

At the sector level, the Bank of England found that software and IT consulting firms contributed ten times as much to annual UK productivity growth in 2023 to 2025 as in the decade before Covid. The industries that build AI are the first to benefit from it.

At the level of usage, Anthropic's Economic Index found that across US states, the share of workers in computer and mathematical jobs closely tracks how heavily people use its AI model, Claude. The same research found that the education level of a user's prompt and of the AI's response were almost perfectly correlated. Sophisticated questions get sophisticated answers; simple questions get simple ones.

At the level of individuals, the picture is similar. A study of ChatGPT users found that people with stronger computing skills and AI experience used more structured, iterative prompting, and got better results. Novices relied on trial and error. In a graduate data-analysis course, prior technical experience predicted success with AI, while self-reported communication skill and familiarity with AI tools did not close the gap.

The common thread is not access. Anyone can open a chatbot. The difference lies in what people are able to bring to the conversation.

What effective prompting actually draws on

If the advantage were simply technical, the answer would be to teach everyone to code. It isn't. The skills behind effective AI use are mostly old ones, and most of them are not technical at all.

Clear communication. Knowing what you want and stating it with context and constraints. In the ChatGPT study, clarity of intent mattered more than grammatical polish.
Domain expertise. Knowing a subject well enough to ask the right question and to spot a wrong answer. Less experienced users tend to get stuck rewording prompts without getting any closer to a result.
Leadership and delegation. In a Harvard Kennedy School experiment, people who led teams of AI agents well also led human teams well, with a correlation of 0.81. That held at 0.69 after controlling for task-specific skills and fluid intelligence. The best leaders asked more questions and encouraged turn-taking.
Critical judgement. Knowing what you don't know, and checking output rather than trusting it. Experts treat AI output as a draft; novices tend to treat it as an answer.
Structured problem-solving. Breaking a vague goal into steps and defining what "done" looks like. This is where technical workers have a head start, but it can be taught in any discipline.

One finding deserves emphasis. In the Harvard study, the leadership skills that predicted success with AI were unrelated to gender, age, ethnicity or formal education. So the answer isn’t more schooling but building specific skills, and that means the gap can close. It is the deliberate development of specific capabilities, which means the gap can be closed.

A caution: this research is young, many studies are small, and these skills overlap heavily. They are best read as a cluster of related capabilities rather than five separate causes.

The strongest objection

The best argument against this article is that AI levels the playing field. There is good evidence for it.

In a landmark field study of around 5,000 customer-support agents, an AI assistant raised productivity by 14% on average. The least experienced workers resolved 35% more cases per hour, while the most skilled saw almost no benefit. The tool worked by capturing what top performers did and passing it on to newcomers. AI tools are also getting better at asking clarifying questions themselves, so the burden of expressing intent is starting to shift onto the machine.

But look at how that levelling happened. The support agents were not prompting an open-ended chatbot. They used a tool designed for one job, trained on experts' behaviour, and built into their workflow. The expertise had already been encoded by someone else.

So the objection doesn't refute the argument; it moves it. Someone still has to design those embedded systems, decide what good looks like, and judge whether the output is right. The skills don't disappear. They move upstream, to the people who build, configure and supervise AI. Embedded AI can spread capability, but only after capable people have built it.

What this means for the UK

The UK's weak spot has never been invention. It has been diffusion: getting new technology and good practice out of leading firms and into the long tail of ordinary businesses. AI risks repeating that pattern unless policy treats capability as seriously as access.

The prize is large. OECD estimates cited by the UK government suggest AI could add 0.4 to 1.2 percentage points a year to UK labour productivity growth over the next decade. That is modelled potential, not a measured result, and how much of it is realised depends on the people using the tools.

Three priorities follow.

Teach AI literacy as a literacy, not a software course. Skip narrow "prompt engineering" training, whose tricks date quickly. Focus instead on clear briefing, structured problem-solving and critical evaluation of output, embedded in every subject from school to adult retraining.
Treat delegation as a core workplace skill. If leading AI agents draws on the same abilities as leading people, management training is AI training. That opens the door to people outside technical professions.
Fund embedded tools for the long tail. Small firms will not all become expert prompters. Support for well-designed, sector-specific tools, built with domain experts, can bring the levelling effect seen in the customer-support study to businesses that would otherwise be left behind.

None of this is as glamorous as data centres or frontier models. It is, however, where most of the productivity is likely to be found.

Conclusion

AI raises the value of human capability rather than replacing it. The early evidence shows its gains flowing to people who can frame problems, communicate intent and judge results, and those skills are currently concentrated in technical sectors.

That is not a reason for pessimism. The skills involved are learnable, they are unrelated to background, and many of them are the same skills good managers, teachers and writers have always needed. The UK can close the gap by investing in people as deliberately as in technology.

The question is not whether Britain has access to AI. It is whether Britain's workforce knows what to ask of it.

Sources

More

Follow me on LinkedIn for more UX articles.

AI Prompting is The New Literacy