Rich countries are running short of workers. AI may help fill the gap.
For about two decades, rich countries have had two worries about labour. Workers were not getting much more productive, and there were going to be fewer of them. The first worry fed a long debate about ”zombie companies”, firms too weak to grow but with just enough credit to stay alive. This ties up workers and capital that could be put to better use elsewhere. The second worry had a blunter answer: people would have to work longer. The rich world needed more out of each worker at a time when the number of workers was about to decline.
How AI is changing productivity
Then came ChatGPT. Most people’s first reaction was to stare at the empty text box and wonder what to type. Early uses were small: tidying up an email, spotting a typo, working out what to cook from whatever was left in the fridge. Before long people were asking it the questions they used to type into Google. As the models improved, the requests became more ambitious. As the price per token fell, people used more tokens, much as more efficient steam engines led Victorian Britain to burn more coal, not less. The demand for machine effort has turned out to be enormous.
All that effort requires hardware. Data centres and the chips inside them have become one of the most visible parts of America’s investment boom, while other countries are scrambling to keep up. However, the economic payoff will depend less on how much capacity gets built than on what companies actually do with it.
Economists have seen this before. In 1987, Robert Solow, economist and Nobel laureate, quipped that the computer age could be seen everywhere except in the productivity statistics. These gains did eventually arrive, but only in the late 1990s, after firms had spent years reorganising around the new technology. If AI turns out to be anything like the personal computer, or electricity before it, adopting it will take more than just paying for a subscription. Firms will have to decide which tasks to hand over, how to verify the results and who will be held responsible when things go wrong. A handful of small open-source projects give some idea of what that could look like.
From chatbots to AI agents
Why now? Largely because of the models. OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5, both released in September, are far better than their predecessors at interpreting code that was never meant to be read. Tools such as Ghidra have long been able to turn a compiled program back into pseudocode, a rough and unlabelled sketch of what it does. Making sense of that sketch used to be slow, specialist work. A model can now go through it, work out what each part is for and write new code that behaves in the same way. The result is a reconstruction, not the original source code. For anyone who just wants to copy a feature, that is usually enough.
Paperclip, an open-source platform, tries to give its user a virtual workforce. You set out a mission and then “hire” AI agents, each with a job title, a boss and a monthly budget. The agents wake up on a schedule, check what they have been given and get on with it. The project’s own pitch is the “zero-human company”. That is still a long way off, since the agents need watching. But it changes what limits a one-person business. The bottleneck becomes how much work the owner can hand off and still keep an eye on, rather than how many hours there are in the day.
REA, built by a developer who goes by morluto, should worry software incumbents more. It has collected more than 280,000 stars on GitHub, and it lets AI agents pick apart existing software, from desktop programs to Android apps, and explain how the software works. The idea is that a user spots a feature in someone else’s product and asks an agent to build something similar. Copying software has always been possible; it was just slow and required expensive engineers. If that cost falls sharply, firms whose main defence has been years of coding invested in their products may find it weaker than they thought. A moat dug mainly with coding hours may prove shallow.
Games are another example. AnyPS5, which appeared on GitHub in August, tries to turn PlayStation 5 games into programs that run natively on Windows and Linux, with no emulator in between. That is possible because the console uses the same kind of x86 processor as most PCs. Only a few titles work so far, and getting more of them running means a lot of fiddly, repetitive fixing, which is exactly the kind of work contributors are now handing to AI coding tools. Whether any of it is legal is a separate question.
Can AI help ease labour shortages?
None of this means autonomous companies are about to take over. These are early, rough projects. What matters is the direction they point in. A chatbot waits to be asked something. An agent can be given a standing brief, a budget and a deadline. If agents become reliable, someone who now does tasks one at a time could supervise several streams of work, and a founder working alone could do what used to take a department. In economies short of workers, that would effectively expand the labour supply, though software will not fill the empty posts on a hospital ward or a building site.
It would not be painless. The work AI does well is not necessarily the work employers are struggling to find people for. Some people will find their skills are worth more; others will find that what they sell has suddenly become cheap. Higher output would leave room to compensate those who lose out, but nothing guarantees that anyone will. An ageing population makes a labour-saving technology easier to welcome. It does not change who wins and who loses.
What AI means for investors
Investors have a different question: who keeps the gains? Cheap coding can help a start-up take on an incumbent, but it also makes the start-up easier to copy, and much of the benefit may simply go to customers through lower prices. For a fund such as Espiria Global Innovation, the job is to distinguish between companies that merely use AI and those that can build a sustainable competitive advantage with it. As machine effort becomes more abundant, the scarce complementary assets, such as proprietary data, distribution and customer trust, may become more valuable.
In 1930, John Maynard Keynes, economist, predicted that his grandchildren’s generation might need to work only 15 hours a week. It did not work out that way. Rich societies spent most of their extra wealth on consuming more rather than working less. AI may give them the choice again. For countries that had resigned themselves to ever longer working lives, simply having that choice would be progress.
