Almost every leadership team I see is quicker than it was a year ago, and being quicker hasn’t made most of them better. I’ve sat in enough rooms over the last twelve months to know the pattern: more work going out of the door, produced faster, by people who feel busier than ever, and none of it adding up to a team that’s calmer, clearer, or deciding any better. That gap, between the speed everyone can see and the value that hasn’t shown up to match it, is what this piece is about. I don’t think many organisations have noticed it’s happening to them.
Let me give AI its due first, because the argument depends on it. One part of the job has genuinely become cheap, and that part is producing the work: drafts, options, analysis, the first version of almost anything. The studies are not subtle about it. In a controlled experiment with professional writers, ChatGPT cut the time taken by 40% and raised quality at the same time (Noy and Zhang, 2023); developers given an AI assistant finished a coding task around 56% faster (Peng et al., 2023); consultants using GPT-4 produced work rated 40% higher on the tasks that suited it (Dell’Acqua et al., 2023). A capable person now does in twenty minutes what used to fill a morning. That bit is real, and pretending otherwise just makes you look frightened of it.
But a cost that falls in one place doesn’t disappear, it moves. When producing the work gets that cheap, you simply get more of it, and every piece still has to be judged before anyone can safely act on it. Is this actually right, or does it only look right? Would it survive contact with a real client, a real market, a real decision? What, out of all of it, should we decide not to do? That judging and deciding is now the expensive part of the job, and it’s precisely the part the machine hasn’t done for you.
So the saving you thought you’d banked on production quietly reappears on the other side of the ledger, as the cost of checking, correcting and steering everything the machine now makes. Sometimes it nets out smaller. In high-stakes or heavily regulated work, where being confidently wrong is expensive, it nets out larger. Either way you haven’t removed the cost, you’ve moved it, from making the work to vetting it. And making scales far more easily than vetting does.
Whether AI has saved you anything, then, comes down to a single question almost nobody asks before buying the tools: how much judgement does a given piece of work need before you can trust it? Where the task is well-defined and a mistake is cheap and obvious, the cost has moved and shrunk, and you’re ahead. Where the work is open-ended, or a quiet error is expensive, the cost has moved and grown, and you may well be behind. The same technology, pointed at two different kinds of work, gives opposite answers.
The research, taken honestly, points the same way, though I’d treat all of it as supporting the argument rather than settling it. Gallup’s State of the Global Workplace: 2026 Report, published in April and so already ageing, found that the strongest single predictor of whether an AI rollout works isn’t the technology but whether managers actively get behind it, ahead even of technical integration, and that fewer than one in three US employees in organisations already rolling out AI strongly agree their manager does (Gallup, 2026). That’s a correlation drawn from people describing their own workplaces, not proof of cause, so I’d hold it as direction rather than mechanism. But the direction is the point: what decides whether this works sits with management, not with the tools.
Deloitte’s State of AI in the Enterprise tells the same story from the other end. Access to sanctioned AI tools jumped by around half in a single year, from under 40% of workers to roughly 60%, yet among the people who have that access, fewer than 60% use it day to day, a figure that’s barely shifted. More telling still, about 84% of organisations haven’t redesigned a single job around AI, even as most of them expect serious automation within three years (Deloitte, 2026). The tools went in. The work around them didn’t change to meet them.
There’s one finding worth handling with care, because it’s usually quoted badly, and the honest version is more useful anyway. It’s tempting to say AI widens the gap between strong and weak performers, but the evidence cuts both ways. Several of the best-known studies found close to the opposite: the call-centre work by Brynjolfsson, Li and Raymond (2025), where the biggest gains went to the least experienced agents, the writing experiment by Noy and Zhang (2023), which compressed the gap between weaker and stronger writers, and the BCG consultant study by Dell’Acqua and colleagues (2023). On well-defined tasks, AI tends to level people up. Then Otis and colleagues (2024), in a field study with Kenyan entrepreneurs, found the reverse on open-ended work: no average benefit, gains for the strongest, and an actual fall in performance for the weakest, who couldn’t tell when the model’s advice was wrong and acted on it anyway. The way to reconcile the two is the part that matters. AI levels people up where the task is contained and errors are easy to spot, and pulls them apart where the work is open-ended and you need judgement to catch a confident mistake. Most of management is the second kind.
Which brings me to what I think organisations have actually misread. Almost all of them still reward the half of the work that just got cheap. We measure utilisation, output and delivery. We promote the people who visibly produce. We write job descriptions around getting the work made. And the scarce, decisive skill, knowing which of ten plausible options is the right one and having the nerve to bin the other nine, sits on no dashboard anywhere. You can see the strain of it in how stretched senior people have become, though we tend to misread that too. The extra load isn’t the visible work. It’s the checking, the integrating and the sense-making piled on top of the work that still decides how their performance gets judged. Nobody designed for it, nobody counts it, and so people absorb it until something gives.
None of which makes managers replaceable, whatever the headlines say. If anything it does the opposite. It makes the one thing managers are for, judgement, the thing the whole system now waits on, at the exact moment we’d spent thirty years training and rewarding everyone to produce more, and faster.
There’s a sting in the tail, and it’s where I’ll pick up next time. The judgement we’ve all come to depend on wasn’t downloaded from anywhere. It was built slowly, by people doing the junior work and learning, often painfully, from getting it wrong. Which is to say it was built by doing more or less exactly the work AI is now quietly absorbing. I’ll come to what that means, and why “just protect the junior roles” isn’t the answer it first looks like, in the next piece.
References
- Brynjolfsson, E., Li, D. and Raymond, L.R. (2025) ‘Generative AI at work’, The Quarterly Journal of Economics, 140(2), pp. 889–942.
- Deloitte (2026) State of AI in the Enterprise. 8th edn. Deloitte AI Institute. Available at: https://www.deloitte.com/global/en/issues/generative-ai/state-of-ai-in-enterprise.html (Accessed: 22 June 2026).
- Dell’Acqua, F., McFowland III, E., Mollick, E.R., Lifshitz-Assaf, H., Kellogg, K.C., Rajendran, S., Krayer, L., Candelon, F. and Lakhani, K.R. (2023) Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Harvard Business School Working Paper 24-013. Available at: https://ssrn.com/abstract=4573321 (Accessed: 22 June 2026). [Subsequently published in Organization Science, 37(2), 2026, pp. 403–423.]
- Gallup (2026) State of the Global Workplace: 2026 Report. Washington, DC: Gallup. Available at: https://www.gallup.com/workplace/349484/state-of-the-global-workplace.aspx (Accessed: 22 June 2026).
- Noy, S. and Zhang, W. (2023) ‘Experimental evidence on the productivity effects of generative artificial intelligence’, Science, 381(6654), pp. 187–192.
- Otis, N.G., Clarke, R., Delecourt, S., Holtz, D. and Koning, R. (2024) The Uneven Impact of Generative AI on Entrepreneurial Performance. Harvard Business School Working Paper 24-042. Available at: https://ssrn.com/abstract=4671369 (Accessed: 22 June 2026).
- Peng, S., Kalliamvakou, E., Cihon, P. and Demirer, M. (2023) The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. arXiv preprint arXiv:2302.06590. Available at: https://arxiv.org/abs/2302.06590 (Accessed: 22 June 2026).