Buoy

The biggest efficiency gain ever (terms and conditions apply)

· 6 min read · the Buoy humans

Every few years somebody announces the biggest productivity gain in history. Usually it's a press release. This time there are randomized trials, and they mostly agree: generative AI makes a lot of knowledge work faster, sometimes dramatically. We built Buoy because of what's in the footnotes of those trials, so let's read them together.

Adoption first, because it's absurd

Two years after ChatGPT launched, about 39% of American adults aged 18 to 64 had used generative AI, and roughly a quarter had used it at work in the prior week. The economists who ran that survey, Alexander Bick, Adam Blandin and David Deming, compared it with the arrival of the personal computer and the internet and concluded that adoption had been faster than either.1 (Their 2025 revision softens that to "as fast as the PC" for workplace use. Still absurd.)

So the tools are everywhere. Are they working?

The gains, in order of how much you'd brag about them

Programmers, 55.8% faster. In a 2023 controlled experiment by Sida Peng and colleagues, developers with GitHub Copilot built a small HTTP server 55.8% faster than developers without it.2 That's one bounded task, and the confidence interval runs from 21% to 89%, which is a polite way of saying "a lot, but we're not sure how much."

Writers, 40% faster and 18% better. Shakked Noy and Whitney Zhang gave ChatGPT to half of 453 college-educated professionals doing short writing tasks. Time fell 40%, and independent graders rated the output 18% higher.3 Published in Science, which does not publish press releases.

Consultants, 25% faster and 40% better. Fabrizio Dell'Acqua and a team at Harvard, Wharton, MIT and BCG ran an experiment on 758 BCG consultants. On tasks inside the AI's capability, those with GPT-4 finished 12.2% more tasks, 25.1% faster, with quality rated 40% higher.4 Hold that "inside the AI's capability" clause. We're coming back to it.

Developers at real companies, 26% more work. Three field experiments at Microsoft, Accenture and a Fortune 100 company, 4,867 developers in total, found that those given Copilot completed 26% more tasks.5 Noisy (the standard error is about 10 points), but it's the largest real-workplace study anyone has run.

Customer support, 14% more resolved. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,179 support agents at a software company given an AI assistant. Issues resolved per hour went up 14% on average.6

If you squint, that's "the biggest efficiency gain ever." We're not going to argue with it. We're going to read the next paragraph of each paper.

The fine print says the same thing five times

Here is the sentence that appears, in some form, in nearly every one of those studies: the gains go to the people who were behind.

In the support study, the 14% average hides a 34% gain for novice and low-skill agents and roughly nothing for the most experienced ones.6 In the writing study, the gap between weaker and stronger writers narrowed; ChatGPT lifted the bottom more than the top.3 In the three corporate experiments, less experienced developers adopted the tool more and gained more from it.5

Why would that be? Brynjolfsson's team has the clearest explanation: the AI assistant had been trained on conversations from the company's best agents. What it actually did was hand the top performers' habits to everyone else. The productivity didn't come from the machine. It came from people who already knew what to do, compressed and distributed.

Which raises the obvious question. If the gain comes from borrowing know-how from someone better, what happens when the AI is the thing you don't know how to use?

When the frontier is jagged, you fall off it

Back to the consultants. The Harvard team also gave them a task deliberately chosen to sit just outside what GPT-4 could do well. On that task, consultants using the AI were 19 percentage points less likely to get the right answer than consultants without it.4 They called the boundary between "AI helps" and "AI hurts" the jagged frontier, because it's invisible and it zigzags. Tasks that look identical can sit on opposite sides.

Then in 2025, METR ran the most careful developer study yet: 16 experienced open-source maintainers, 246 real issues in their own repositories, each issue randomly assigned AI-allowed or not. With AI tools, the developers took 19% longer. They believed, before and after, that AI had sped them up by about 20%.7 The researchers were careful to call it a snapshot of early-2025 tools in one setting, not a verdict. But the gap between what people felt and what the clock said is the part worth keeping.

And at the scale of a whole economy, the effect is so far small. Anders Humlum and Emilie Vestergaard matched Danish surveys to payroll records for 25,000 workers in exposed occupations and found average time savings of about 3%, with no measurable effect on earnings or hours two years in.8 The lab results are real; most people just aren't getting them yet.

What we take from all this

Put the two halves together and you get something less exciting and more useful than a headline:

  1. The gains are large and real, and they are largest for people who are newer at the task.
  2. The gains depend on knowing where the frontier is: what to ask, what to approve, when to stop trusting the output.
  3. That knowledge is, today, mostly carried around in people's heads. It got into the support bot by being extracted from humans.

In other words, the biggest efficiency gain ever has a prerequisite, and the prerequisite is a person who already climbed the first rung. That's the whole idea behind Buoy: when you're stuck on the tool, the fastest route isn't a longer prompt. It's ten minutes with someone one level up, looking at the same window you are.

We'll keep updating this as the studies do. If you think we've misread one, the sources are right here.

Sources

  1. Bick, A., Blandin, A., & Deming, D. J. (2024, rev. 2025). The Rapid Adoption of Generative AI. NBER Working Paper 32966. nber.org/papers/w32966
  2. Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. arXiv:2302.06590. arxiv.org/abs/2302.06590
  3. Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192. doi:10.1126/science.adh2586
  4. Dell'Acqua, F., McFowland, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2023). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. Harvard Business School Working Paper 24-013. ssrn.com/abstract=4573321
  5. Cui, Z. K., Demirer, M., Jaffe, S., Musolff, L., Peng, S., & Salz, T. (2025). The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers. Management Science. ssrn.com/abstract=4945566
  6. Brynjolfsson, E., Li, D., & Raymond, L. R. (2023). Generative AI at Work. NBER Working Paper 31161. (The 2024 revision reports 15% on average.) nber.org/papers/w31161
  7. Becker, J., Rush, N., Barnes, E., & Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR. arXiv:2507.09089. metr.org · arxiv.org/abs/2507.09089
  8. Humlum, A., & Vestergaard, E. (2025). Large Language Models, Small Labor Market Effects. NBER Working Paper 33777. nber.org/papers/w33777