Productivity, checked · Part 6 of 8
Leverage: what multiplies an hour
Archimedes is supposed to have said that with a place to stand and a long enough lever he could move the earth. The line was written down centuries after his death, so nobody knows whether he said it. That makes it a fitting start for this part: like much productivity advice, it pairs a real mechanism with a size nobody checked.
The first four stations are about not wasting input: choosing the right work, keeping it moving, giving it attention, keeping the machine running. This station is about the other half of the ratio. Leverage is anything that makes an hour of input produce more valuable output: a skill built over years, a tool, an AI model, a colleague you trained. Every lever costs something to build or use, and some of them quietly push the wrong way.
The question for this station: what actually multiplies output, and by how much?
AI: a jagged frontier
Ethan Mollick's Co-Intelligence describes AI as a strange new co-worker with a "jagged frontier": brilliant at some tasks, oddly bad at others that look just as hard, and you can't tell which is which in advance. The best evidence for this comes from a study Mollick co-authored. 758 consultants at Boston Consulting Group worked on 18 realistic tasks. With GPT-4 they finished 12.2% more tasks, 25.1% faster, at over 40% higher quality (Dell'Acqua et al. 2026) ~85 %. But on one task built to sit just outside the model's ability, 84.5% of consultants without AI got the right answer, against 60.0% and 70.6% in the two AI groups ~70 %. The only difference was which side of an invisible line the task fell on.
Other studies point the same way inside the frontier. In a randomised trial with 453 professionals doing short writing tasks, ChatGPT cut the time by 40% and raised quality by 18% (Noy & Zhang 2023). Among 5,172 customer-support agents, an AI assistant raised issues resolved per hour by 15% (Brynjolfsson, Li & Raymond 2025). In the working paper, novice and low-skilled agents gained 34%.
That last detail is the most consistent finding in the field: AI levels skill ~85 %. In the BCG study, below-average consultants improved quality by about 43%, top performers by about 17%. Experienced support agents gained little speed and lost slightly on quality. A plausible mechanism: the model supplies a competent average answer, which is a big step up for a beginner and a small one, or a step sideways, for an expert.
The catch is that a good assistant makes people stop checking. In a preregistered experiment, 181 recruiters scored résumés with an AI that was about 99%, 85% or 75% accurate, or with none. Those given the better AI put in less effort, followed it more blindly and judged less accurately (Dell'Acqua 2022). Mollick calls this falling asleep at the wheel. The better the assistant, the more tempting it is to stop reading its work.
Code: the stopwatch and the feeling disagree
In the most-quoted trial of AI for programmers, 95 freelance developers recruited on Upwork wrote a small web server in JavaScript. Those with GitHub Copilot finished in 71 minutes instead of 161, 55.8% faster (Peng et al. 2023). The result is solid for that task (~85%) and a weak guide to real work ~30 %: one bounded job, written from scratch, with the less experienced gaining most. Field trials at Microsoft, Accenture and a Fortune 100 firm, covering 4,867 developers, found about 26% more completed tasks, again mostly for juniors (Cui et al. 2025).
Then METR looked at a different population: 16 experienced open-source developers working on 246 real tasks in codebases they knew well. With early-2025 AI tools they took 19% longer. Before starting they had forecast a 24% speed-up, and afterwards they still believed they had been about 20% faster (Becker et al. 2025). The slowdown belongs to one setting and one generation of tools. METR's 2026 follow-up was abandoned as unreliable, partly because developers refused to work without AI, and secondary reports put the new estimates near zero. The gap between feeling and stopwatch is the part worth keeping, and Feedback returns to it.
Teams show a similar puzzle. Google's DORA survey for 2024 found that 25% more AI adoption went with about 1.5% lower delivery throughput and 7.2% lower stability, even as code quality and review speed improved slightly (~70%). The 2025 report found AI now linked to higher throughput but still to more instability, with code review as the new bottleneck. Both are survey associations, not experiments. The mechanism is familiar from the factory floor: speeding up one step only moves the pile to the next one (see Flow).
None of this would have surprised Fred Brooks. In "No Silver Bullet" (1986) he argued that no single technology would deliver even a tenfold gain in software productivity within a decade, because most of the difficulty is essential (the complexity of the problem itself) rather than accidental (friction from tools and languages) ~85 %. AI tools are the toughest test his claim has faced, and none of the controlled studies comes close to tenfold. To be fair to the enthusiasts, Mollick's advice to assume "this is the worst AI you will ever use" also has data behind it: the length of task AI agents can complete doubled roughly every seven months from 2019 to 2025. Brooks is still right, but the margin is shrinking.
Skill: the slow lever
Skill is the oldest lever, and the one most distorted by its popularisers. Anders Ericsson's research on violinists at a Berlin music academy found that the best students had practised alone far more than the merely good ones: about 7,410 hours by age 18, against 5,301 (Ericsson, Krampe & Tesch-Römer 1993). Malcolm Gladwell turned the study's average for the best group at age 20 into the "10,000-hour rule". In Peak, Ericsson rejects the rule himself: there is no magic number, and the hours differ by field ~5 %.
The bigger problem is with Ericsson's own claim. A double-blind replication with 39 violinists found that practice explained about 26% of the differences in skill, not 48%, and the "good" group had practised more than the "best", though not significantly (Macnamara & Maitra 2019). A meta-analysis of 88 studies found that deliberate practice explained 26% of performance differences in games, 21% in music, 18% in sports, 4% in education and under 1% in professions (Macnamara, Hambrick & Oswald 2014) ~85 %. Practice matters. It is not the whole story, and in most jobs it is a small part of it.
What survives is the mechanism. Practice builds mental representations: chess masters recall real game positions at a glance but lose most of that advantage when the pieces are placed at random (Chase & Simon 1973) ~85 %. And experience alone is not practice: without feedback, years on the job can even make people worse (see Feedback).
Knowledge workers rarely have what deliberate practice requires: an established training method and a coach who designs exercises. So Cal Newport's suggestion in So Good They Can't Ignore You that deliberate practice will let most knowledge workers leap ahead rests on thin evidence. The famous "10× programmer" has the same problem: the 1968 study behind it mixed languages and methods, and a re-analysis attributes only about 5:1 to the programmers themselves. Large differences between people are real. Where they come from is less clear than the books suggest.
AI now sits right on this lever. Adam Grant's Hidden Potential argues for scaffolding: temporary support, like a tutor or worked example, that gets a learner past an obstacle and then fades. A randomised trial with about 1,000 Turkish high-school students shows what happens when it doesn't fade. Plain GPT-4 access raised practice scores but cut exam scores by 17%; a version built as a tutor, with guardrails, did no harm (Bastani et al. 2025) ~85 %. Used as an answer machine, AI borrows today's output against tomorrow's skill.
Other people: the manager's lever
Andy Grove, Intel's long-time chief executive, built High Output Management on the same ratio as this series. A manager's output, he wrote, is the output of their organisation plus the neighbouring groups they influence, and leverage is an activity's effect on that output divided by the time it takes. High-leverage work affects many people, lasts a long time, or uses knowledge only you have.
His favourite example is training: a few hours of preparation pay back across a team's whole year ~70 %. The evidence that training works is good. A meta-analysis found an effect of about d = 0.62 on on-the-job behaviour (Arthur et al. 2003). That it beats a manager's other uses of time is Grove's arithmetic, not a finding. In a randomised trial in Indian textile plants, introducing basic management practices raised productivity by about 17% in the first year (Bloom et al. 2013) ~70 %. That was a factory, though, and moving the result to an office is an extrapolation.
Grove's people advice ages less well. His "task-relevant maturity", which matches management style to a person's experience with the task, is a version of situational leadership theory, which has fared poorly in tests. His warning that "delegation without follow-up is abdication" is untested but sensible. It also reads well as a guide to delegating to AI agents: a delegate who has never done the task needs checking, whoever or whatever it is. Adding people has its own coordination cost, which belongs to Flow.
Where the work happens
Remote work is a lever many people now argue about and few have measured. The best evidence is two randomised trials at the same Chinese travel company. In the first, call-centre volunteers allowed to work from home for nine months performed 13% better: 9% from working more minutes, 4% from more calls per minute. Quitting halved, though promotion rates fell for the same performance (Bloom et al. 2015). The result is solid for routine, individual work ~85 % and much less certain beyond it (~50%). One co-author co-founded the company.
The second trial, with 1,612 engineers, marketing and finance staff, tested two days a week at home over six months. Quitting fell by a third, with no effect on performance reviews or promotions (Bloom, Han & Liang 2024) ~85 %. Managers had predicted a productivity loss beforehand. The honest summary: where the work happens is a weak output lever and a strong retention lever. The debate is louder than the effect.
What doesn't work
- The 10,000-hour rule. No threshold exists; the number was one group's average at age 20, and Ericsson called it arbitrary ~5 %.
- "Practice, not talent." In a study of about 10,500 Swedish twins, identical twins who practised different amounts did not differ in musical ability (Mosing et al. 2014) ~15 %.
- Persona prompts. Telling an AI "you are an expert" did not raise accuracy across 162 roles and 2,410 questions (Zheng et al. 2024) ~30 %; context and examples do help (~70%).
- Lab speed-ups as real-world gains. The 55.8% Copilot result came from one bounded task; experienced developers in their own codebases were slower ~30 %.
Try this
- Map your own frontier with a stopwatch. Pick a recurring task, do it a few times with AI and a few without, and time both. In the METR study, felt and measured speed were about 40 percentage points apart.
- Check hardest where AI helps most. It helps beginners most, and beginners are worst at spotting its errors. In unfamiliar areas, verify facts and sources before using the output.
- When learning, make AI the tutor, not the answer key. Ask it for hints, questions and explanations; do the final step yourself.
- Practise the part of your work that has feedback. Pick one skill where you can see results quickly, work at its edge for short, focused sessions, and find someone who can design exercises for it.
- Spend an hour on something that lasts. A checklist, a training session or a written answer to a recurring question pays back every time someone uses it.
16 claims checked
The evidence behind this article
Every claim above, rated by the chance it holds up against the research. Open one for the sources and their exact words.
- Arthur, Bennett, Edens, Bell (2003). Effectiveness of training in organizations: A meta-analysis of design and evaluation features. · Meta-analysis
- Bastani, Bastani, Sungu, Ge, Kabakcı, Mariman (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics · Study
- Bloom, Eifert, Mahajan, McKenzie, Roberts (2012). Does Management Matter? Evidence from India * · Study
- Bloom, Han, Liang (2024). Hybrid working from home improves retention without damaging performance · Study
- Bloom, Liang, Roberts, Ying (2013). Does Working from Home Work? Evidence from a Chinese Experiment · Study
- Brooks (1987). No Silver Bullet Essence and Accidents of Software Engineering · Source
- Brynjolfsson, Li, Raymond (2025). Generative AI at Work · Study
- Chase, Simon (1973). Perception in chess · Study
- Cui, Demirer, Jaffe, Musolff, Peng, Salz (2024). The Effects of Generative AI on High Skilled Work: Evidence from Three Field Experiments with Software Developers · Study
- Dell'Acqua (2022). Falling Asleep at the Wheel: Human/AI Collaboration in a Field Experiment on HR Recruiters · Study
- Dell'Acqua, McFowland III, Mollick, Lifshitz-Assaf, Kellogg, Rajendran, Krayer, Candelon, Lakhani (2023). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality · Study
- DORA (2024). Accelerate State of DevOps Report 2024 · Source
- Ericsson, Krampe, Tesch-Römer (1993). The role of deliberate practice in the acquisition of expert performance. · Study
- Ericsson, Pool (2016). Peak · Book
- Gobet, Simon (1996). Recall of rapidly presented random chess positions is a function of skill · Study
- Grant (2023). Hidden Potential · Book
- Grove (1983). High Output Management · Book
- Joel, Rush, Barnes, David (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity · Study
- Macnamara, Hambrick, Oswald (2014). Deliberate Practice and Performance in Music, Games, Sports, Education, and Professions: A Meta-Analysis · Meta-analysis
- Macnamara, Maitra (2019). The role of deliberate practice in expert performance: revisiting Ericsson, Krampe & Tesch-Römer (1993) · Replication
- Mollick (2024). Co-Intelligence · Book
- Mosing, Madison, Pedersen, Kuja-Halkola, Ullén (2014). Practice Does Not Make Perfect · Study
- Newport (2012). So Good They Can't Ignore You · Book
- Noy, Zhang (2023). Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence · Study
- Peng, Kalliamvakou, Cihon, Demirer (2023). The Impact of AI on Developer Productivity: Evidence from GitHub Copilot · Study
- Zheng, Pei, Logeswaran, Lee, Jurgens (2024). When ”A Helpful Assistant” Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models · Study