What Brynjolfsson 2025 means for who you train first

For L&D leaders, integration managers, and functional executives planning their next AI training cohort: the empirics on who actually gains from AI assistance, who doesn’t, and who loses ground are sharper than the conventional pilot heuristic admits.

Brynjolfsson, Li, and Raymond’s 2025 randomized controlled trial in the Quarterly Journal of Economics (n=5,172 customer-service workers) found AI delivered a 30% productivity gain at the bottom of the skill distribution, no measurable gain at the top, and a measurable quality decline among the most experienced.⁷ The result has held under replication scrutiny. It maps onto a clean operational rule: AI compresses the gap between the least and most experienced, but it does it by lifting the bottom, not by extending the top.

The conventional pilot playbook says the opposite. Start with the senior team. Build executive buy-in. Cascade from there. The logic feels right because senior people set tone. The data says the logic is upside down for productivity outcomes.

Three things follow from that.

First, the people with the most to gain are the cohort with the least seniority. New hires, recent rotators, less-experienced practitioners. They’re the cohort where training transfers fastest into measurable output change. Pilots that target this cohort produce the cleanest before-and-after telemetry — which is exactly the evidence procurement teams need to authorize the next phase.

Second, senior people need a different curriculum, not the same curriculum applied later. The Brynjolfsson finding doesn’t say “skip the senior team.” It says automation at the senior level degrades quality without lifting throughput. The training that works there is judgment-layer training: when to override AI suggestions, how to catch quality regressions in juniors who are using AI well, what kinds of decisions AI should not be near. That’s a different course than the productivity-gain training the junior cohort needs.

Third, the cohort sequence shapes the change story. When the bottom-of-distribution cohort shows measurable gain in week six, the rollout has its own evidence. The senior cohort’s training program is justified by what’s happening one layer below. The narrative writes itself, and it’s a defensible one for the board.

This is how we structure the AI training engagements we run. The diagnostic surfaces who’s where on the skill distribution; the curriculum sequences the room-to-grow cohort first; the senior judgment-layer training follows, anchored against the productivity data the first cohort produces. The pilot decides the rollout’s story. The story decides the budget.