an expert assesses how far this revolution still has to run

Every week brings fresh claims about AI transforming the workplace. A CEO declares a revolution. A think piece predicts millions of jobs vanishing overnight. The noise is relentless.

But strip away the hype and there is a simpler question. In developed economies, what has AI actually changed about work so far? The answer turns out to be more interesting, and more uneven, than either side suggests.

What’s real

Let’s start with what the evidence supports. AI is delivering genuine productivity gains in specific kinds of knowledge-based and service work. An experimental study^[1] found that professionals using ChatGPT for writing tasks took 40% less time to complete them, with an 18% improvement in quality (as evaluated by their colleagues in blind testing).

And another study of more than 5,000 customer service agents^[2] found a 15% increase in issues resolved per hour. An industry experiment^[3] involving realistic, complex tasks done with management consultants found they completed the work 25% faster and produced results that were deemed to be 40% higher in quality (again, judged by experts in blind tests). Randomised trials^[4] involving nearly 5,000 software developers documented a 26% increase in completed tasks.

AI has long been discussed as a threat to jobs and livelihoods. But what’s the reality? In this new series^[5], we explore the impact it is already having on different occupations – and how people really feel about their AI assistants. These are not small numbers. And adoption is moving fast. A US survey^[6] found that nearly four in ten workers were using generative AI at work by mid-2025. This pace of adoption outstrips the early years of both the personal computer and the internet. Across countries in the Organisation for Economic Co-operation and Development (OECD), firms report that AI integration^[7] into business functions is accelerating. So the productivity story is real, particularly in text-heavy, codifiable tasks across legal, finance, marketing and customer service. That much is not hype. What’s overstated But the apocalyptic predictions have not yet materialised. Employment across OECD countries remains historically robust^[8]. A review of the research-based evidence^[9] produced in the US in early 2026 found that despite rapid adoption, AI has so far caused little in the way of widespread job losses or pay cuts. And a study^[10] (as yet unpublished) that tracked AI chatbot use in Danish workplaces found essentially zero effects on earnings or recorded hours, even among heavy users and early adopters. Why? Because many jobs still require tacit knowledge, physical presence, sound judgement and the kind of contextual awareness that AI cannot yet replicate. And adoption is far more uneven than the headline numbers suggest. While AI use among firms in the US^[11] soared between 2023 and 2025^[12], a report found^[13] fewer firms had actually embedded it into their operations. The information sector, for example, adopted it at roughly ten times the rate of hospitality. One economic modelling exercise^[14] estimates that AI might add somewhere between 1% and 1.6% to US GDP over the next decade. This is significant, but it is far short of the transformative claims^[15]. The gap between productivity gains in controlled studies and real transformation inside organisations remains enormous. The revolution, for most workplaces, has not yet arrived. What’s under-reported Here is where the story gets more consequential and where the commentary falls short. The distributional effects of AI within developed economies deserve far more attention. Not everyone is experiencing this transformation the same way. The evidence on who benefits is strikingly consistent. Less experienced workers see the biggest gains from AI tools. A study^[16] found that AI narrowed the gap between the most and least productive staff, with the largest improvements among lower-ability workers. In customer service^[17], novice agents benefited most. The most experienced staff experienced little improvement and, in some cases, slight quality declines. The industry experiment^[18] mentioned above found below-average performers improved by 43%, while top performers gained 17%. So the biggest gains go to the least experienced workers, narrowing the gap between top and bottom performers within firms. That sounds like good news. But there’s a catch. While AI may compress skills inside firms, the broader labour market is telling a different story. Entry-level roles are shrinking^[19] in AI-exposed occupations. The routine tasks that once justified hiring juniors – jobs which provided learning opportunities for those on the bottom rung – are the first to be automated. Economic theory^[20] has long warned that automation displaces workers from tasks, and the creation of new tasks to counterbalance this is neither automatic nor guaranteed. An estimated 60% of jobs^[21] in advanced economies face some AI exposure.

In most realistic scenarios^[23], inequality worsens without deliberate intervention – partly because higher-income workers hold more capital assets and stand to gain from rising returns^[24] on AI-related investments.

The pattern that is emerging is this: AI helps those already inside the door while quietly narrowing the door for those trying to get in.

Paying attention to the right question

Sector matters. Firm size matters. Job type matters. The AI transition is not one story. It is many – unfolding at different speeds, with different consequences, depending on where you sit in the economy.

The debate has been stuck between breathless optimism and existential dread. Neither is useful. The evidence points somewhere more uncomfortable: a transformation that is real but partial, fast in some corners and stalled in others – and distributing its costs and benefits in ways that are shaped by existing inequalities.

If the productivity gains are genuine, the question is: who captures them? If entry-level work is disappearing, what replaces it? And if the gap between firms that adopt and those that cannot is widening, the focus should be on what we are building in response. Just talking about it won’t be enough.

References

1. ^{^} experimental study (www.science.org)
2. ^{^} 5,000 customer service agents (academic.oup.com)
3. ^{^} An industry experiment (www.hbs.edu)
4. ^{^} Randomised trials (pubsonline.informs.org)
5. ^{^} this new series (theconversation.com)
6. ^{^} US survey (www.stlouisfed.org)
7. ^{^} AI integration (www.oecd.org)
8. ^{^} historically robust (www.oecd.org)
9. ^{^} research-based evidence (laweconcenter.org)
10. ^{^} And a study (www.nber.org)
11. ^{^} firms in the US (www.census.gov)
12. ^{^} 2023 and 2025 (www.mckinsey.com)
13. ^{^} report found (www.anthropic.com)
14. ^{^} economic modelling exercise (academic.oup.com)
15. ^{^} transformative claims (www.goldmansachs.com)
16. ^{^} A study (www.science.org)
17. ^{^} customer service (academic.oup.com)
18. ^{^} industry experiment (www.hbs.edu)
19. ^{^} roles are shrinking (theconversation.com)
20. ^{^} Economic theory (www.aeaweb.org)
21. ^{^} 60% of jobs (www.imf.org)
22. ^{^} Everett Collection/Shutterstock (www.shutterstock.com)
23. ^{^} realistic scenarios (www.imf.org)
24. ^{^} rising returns (www.computerweekly.com)