Mostly beginners. Digital Applied's September 23, 2026 census of 20 studies finds 12 with larger gains for less experienced workers, such as +36% for the lowest-skill support agents in Brynjolfsson's QJE paper against +15% on average, and three that ran the other way. Naval Ravikant's idea of specific knowledge explains the split: AI diffuses trainable knowledge, so a founder's edge now lives in what cannot be trained.
A founder I coach asked me last week whether she should be using AI for the parts of her job she is best at or the parts she is worst at. I gave her the answer I believed, which was "the parts you are worst at," and then went looking for whether the evidence agreed. It mostly does, with two exceptions that matter more to founders than the headline.
The evidence is a table published by Digital Applied on September 23, 2026, with sources read on September 25: twenty randomised trials, field experiments and administrative-data studies that split the productivity effect of AI by worker skill or by task, each figure taken from the primary paper. Twelve of the twenty find larger gains for less experienced or lower-performing workers, three go the other way or show experts slowed, and no row links a productivity gain to that worker's pay. The named numbers are the useful part. In Brynjolfsson, Li and Raymond's Quarterly Journal of Economics paper on 5,172 customer-support agents, the assistant lifted resolutions per hour by 15% on average and by 36% in the lowest skill quintile, while the most skilled agents got no productivity gain and small quality declines. In Dell'Acqua and colleagues' BCG experiment with 758 consultants using GPT-4, bottom-half performers improved 43% against their own baseline and top-half performers 17%, but on a task outside the model's frontier, AI users were 19 percentage points less likely to be correct. In Merali's Yale trial with 300 translators, slower workers saved 21.1% of their time per tenfold increase in model compute and faster workers 4.9%. And in Cruces and colleagues' NBER working paper 34851, 1,174 adults did a business problem-solving task with or without an assistant: the gap between higher- and lower-educated participants fell from 0.548 to 0.139 standard deviations, closing about three-quarters of it, and partly re-emerged in a follow-up module without AI.
The exception I keep coming back to is the founder trial. Otis and colleagues, published online in Management Science in July 2026, gave 640 Kenyan entrepreneurs a GPT-4 assistant for five months. There was no significant average effect on revenue or profit. Low performers ended up nearly 10% worse and high performers may have gained more than 15%, and the authors trace the gap to which advice each group chose to act on. The tool was identical. The judgment was not.
Here is how I now read the census, with Naval Ravikant's vocabulary. Naval defines specific knowledge as knowledge you cannot be trained for; if it can be taught, someone else can be taught it, and eventually a machine can. The mechanism most of these authors propose for why novices gain most is diffusion of best practice: the support-agent paper describes the assistant as spreading what the best agents already did. That is trainable knowledge being copied to everyone. Naval's other line applies to the Kenyan result: in an age of leverage, judgment matters more than effort, and a language model is leverage that pays out only in proportion to the judgment steering it.
| Study (primary source) | Who gained | What it means for a task on my desk |
|---|---|---|
| Brynjolfsson, Li, Raymond, QJE 2025 (5,172 support agents) | Lowest quintile +36%, average +15%, top: small quality decline | If I am already the best in the company at it, the model copies me to others and adds noise to my own work |
| Dell'Acqua et al., HBS/BCG 2023 (758 consultants) | Bottom half +43%, top half +17%; outside frontier −19 points | The frontier matters more than my skill; a task the model is bad at gets worse for everyone |
| Cruces et al., NBER 34851, 2026 (1,174 adults) | Education gap 0.548 to 0.139 SD; partly returns without AI | Assisted performance is not learning; the gap comes back when the tab closes |
| Otis et al., Management Science 2026 (640 entrepreneurs) | Low performers −10%, high performers +15% | On judgment-heavy work the tool amplifies whatever selection ability I bring |
| Humlum, Vestergaard, NBER 33777, rev. 2026 (25,000 Danish workers) | Precise null on earnings and hours | None of the above shows up as pay two years in; the return has to be built, not collected |
So the sorting rule I gave my client, refined. For work outside her specific knowledge, legal drafting, a first financial model, code she reads but does not write, the census says AI moves her most of the way to the median fast, and she should use it without guilt. For work inside her specific knowledge, the studies say to expect little gain and some quality loss, and to use the model as a red team rather than an author. For judgment work, deciding which advice to act on, the Kenyan trial says the model is a multiplier on her selection ability, which is the one thing worth training. And because the Cruces gap partly came back when the assistant was removed, anything she needs to still know in a meeting without a laptop has to be learned the slow way, with the model as tutor rather than proxy.
The limits are real. Eight of the twenty rows are coding studies, two have vendor-affiliated authors, and most used 2022 to 2024 chat or autocomplete tools rather than agents; the one study that compared task types found agentic tool-use tasks gained least. The census does not average across studies because the outcomes are not on a common scale, and it excludes the withdrawn MIT materials-science paper that once carried the "experts gain more" story, which is worth remembering the next time someone cites it. And the founder result is one five-month trial in one country. What it shares with the other nineteen rows is the shape Naval described years before any of them were run: the trainable part is being given away to everyone, and what is left to compete on is what you cannot be trained for.
Sources: Digital Applied, "Who Gains Most From AI at Work? What 20 Studies Found," September 23, 2026 (digitalapplied.com/blog/who-gains-most-from-ai-at-work-studies-compared); Cruces et al., "Does Generative AI Narrow Education-Based Productivity Gaps? Evidence from a Randomized Experiment," NBER Working Paper 34851, 2026 (nber.org/papers/w34851); Brynjolfsson, Li and Raymond, "Generative AI at Work," Quarterly Journal of Economics, 2025; Dell'Acqua et al., "Navigating the Jagged Technological Frontier," HBS Working Paper 24-013, 2023; Otis et al., "The Uneven Impact of Generative AI on Entrepreneurial Performance," Management Science, online July 2026; Eric Jorgenson, The Almanack of Naval Ravikant (Magrathea, 2020).
Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · AI Coach App — Building It in 8 Hours