Bombshell Harvard Study Debunks The AI Productivity Narrative

A 14-minute video, uploaded on 7 October 2026, on a job-market paper by two Harvard economics PhD candidates, Fiona Chen and James Stratton: “Artificial Intelligence in the Firm: Bottlenecks in Software Production”. They use data from Jellyfish, which reads companies’ GitHub, Jira and AI tool usage directly, covering 718 firms and about 726,000 workers from 2021 to March 2026. That lets them compare output before and after a company adopts AI, with no one filling in a timesheet.

The video separates two waves. Assistants, the autocomplete kind, reached about 45 per cent of the firms by April 2024 and changed little. Agents, which plan and write across files, reached over 95 per cent by January 2026, and they do change the code. Twelve months in, each engineer ships about 4,000 more lines a month, with commits and pull requests up by about a fifth. What doesn’t move is the work that counts: Jira issues and epics resolved show no significant gain, and the paper rules out anything above about 12 per cent. What does move is review. The share of pull requests needing changes nearly doubles, comments rise by a third and review time by about half, even though nearly 80 per cent of the firms already use AI to review code. Engineers quoted in the paper describe one senior reviewer as the bottleneck and pull requests too big to face.

The video’s conclusion is about money. Employment shows no significant effect, junior or senior, so the savings that are supposed to pay for agentic AI aren’t arriving. Tech layoffs began with 2021–22 overhiring, before agents existed. That matters, it argues, for an industry Bain now says needs about $6 trillion a year in revenue by 2031, and for an Anthropic IPO priced on replacing engineers. Two small slips: assistants did lift commits (by 9 per cent, significant), and Bain’s figure is annual revenue in 2031, not a five-year total.