AI coding tools boost code volume but not productivity, Harvard study finds

Research of 700,000 developers shows code review bottleneck offsets AI gains

By LineZotpaper

Published22 October 2026

Read Time2 min

A Harvard study of more than 700 firms has found that AI coding assistants and agents significantly increase the volume of code produced, but fail to deliver proportional improvements in completed work, with slower code reviews and increased rework creating a downstream bottleneck.

The study, based on data from the Jellyfish analytics platform, analysed approximately 300 million work events from 700,000 workers across more than 700 firms between January 2021 and March 2026. Researchers examined data from HR systems, version control, issue tracking, scheduling tools and generative AI tools including GitHub Copilot, Cursor and Claude Code.

AI coding assistants were associated with a 12% increase in lines of code added, a 9% increase in commits and a 5% increase in pull requests, with only the effect on commits reaching statistical significance. AI agents led to larger gains: 30% more lines of code, 20% more commits and 23% more pull requests.

However, when measured by completed units of work, Jira issues and epics resolved per worker per month, the picture changed. Coding assistants produced no increase in resolved issues, while agents brought a 12% improvement, far below the 30% increase in code output.

Researchers attributed the gap to bottlenecks in later stages of development. Code review time increased by 49%, the number of comments per pull request rose by 53%, and the share of pull requests requiring changes doubled. The proportion of workers performing code reviews also grew by 14%, with no evidence that AI helped with reviews.

The authors estimated that engineers spend only 30-40% of their time writing code, with the remainder going to planning, review, testing and deployment. When AI raises code-writing productivity, those complementary steps become choke points.

Commenting on the findings, Reddit users highlighted risks. One observed that "every line of code is a potential liability", while another noted the danger of AI hallucinations: agents "can do the most messed up stuff without any hint of malice". A third suggested that mean time to recovery may become an increasingly important metric as engineers understand less of the code they ship.

Analysis

Why This Matters

The findings challenge the narrative that AI coding tools automatically boost software delivery throughput.
Downstream bottlenecks in code review and testing may require new tools or process changes, not just faster code generation.
Organisations investing heavily in AI coding agents may not see the expected return on investment without addressing the full development lifecycle.

Background

The study, conducted by Harvard researchers, used the Jellyfish analytics platform to measure engineering activity across a large dataset of firms. Jellyfish aggregates data from tools such as GitHub, Jira and Google Calendar to provide productivity metrics. The researchers supplemented the platform data with a survey of 100 engineers to validate time-use estimates, finding that code writing occupies only a minority of an engineer's time.

Key Perspectives

The study authors: AI coding tools increase code volume but create bottlenecks in downstream processes, particularly code review. The gains in completed work are much smaller than the increase in lines of code, suggesting the development process itself needs rethinking. Developers and engineering leaders: The findings confirm practical experience, many teams have observed that AI-generated code requires careful review and testing. Some may see this as validating concerns about code quality and maintainability. Critics and skeptics: The study measures only certain tools and time periods. It does not account for long-term productivity benefits such as reduced cognitive load or faster prototyping. The metrics used (Jira issues) may not capture all forms of valuable output.

What to Watch

Whether AI-assisted code review tools emerge to address the bottleneck the study identifies.
Organisations reporting on their own productivity metrics after adopting AI agents, do they see the 12% improvement in completed work or less?
Further research into mean time to recovery and other quality metrics as AI-generated code becomes more common.

Sources

Study Finds AI-generated Code Fails to Boost Productivity Due to Downstream Bottlenecks · InfoQ

Checked

12 claims in this story checked against 1 source: 9 supported.

Show each claim

Partly supportedAI coding tools boost code volume but do not improve productivity.InfoQ: “Coding assistant were found not to increase the number of resolved issues, while agents brought a 12% increase. However, this is significantly smaller than the estimated 30% increase in lines of code.”
Partly supportedThe summary says code review bottlenecks offset AI gains.InfoQ: “When AI raises productivity in code writing, complementary steps may serve as bottlenecks.”
SupportedA Reddit user said, “every line of code is a potential liability”.InfoQ: “Reddit user Prime_1 observed that "every line of code is a potential liability"”
SupportedA Reddit user said agents “can do the most messed up stuff without any hint of malice”.InfoQ: “They can do the most messed up stuff without any hint of malice”
SupportedThe study analysed about 300 million events from 700,000 workers at more than 700 firms between January 2021 and March 2026.InfoQ: “The study was based on approximately 300 million work events of 700,000 workers across 700+ firms from January 2021 to March 2026.”
SupportedCoding assistants increased lines of code by 12%, commits by 9% and pull requests by 5%; only the commits effect was statistically significant.InfoQ: “AI assistants lead to small increases in code output — 12% for lines of code, 9% for commits, and 5% for pull requests — with only the effect on commits being statistically significant.”
SupportedAI agents increased lines of code by 30%, commits by 20% and pull requests by 23%.InfoQ: “AI agents lead to large and significant increases in all three measures: 30% for lines of code, 20% for commits, and 23% for pull requests.”
SupportedAssistants produced no increase in resolved issues, while agents brought a 12% increase in resolved work.InfoQ: “Coding assistant were found not to increase the number of resolved issues, while agents brought a 12% increase.”
SupportedCode review time rose 49%, comments per pull request rose 53%, and the share of pull requests requiring changes doubled.InfoQ: “the average PR time to review increased by 49%, the number of comments per PR rose by 53%, and the share of PRs requiring changes doubled.”
SupportedThe share of workers performing code reviews grew 14%, with no improvement from AI-assisted reviews.InfoQ: “the share of workers performing code reviews increased by 14%, with no improvement coming from the adoption of AI to assist with the reviews.”
SupportedThe authors estimated engineers spend 30–40% of their time writing code.InfoQ: “We impute time usage from the Jellyfish data and validate this imputation using a survey of 100 Prolific engineers, finding that engineers spend 30-40% of their time writing code”
Partly supportedThe story says complementary development steps become choke points when AI raises code-writing productivity.InfoQ: “When AI raises productivity in code writing, complementary steps may serve as bottlenecks.”