AI coding productivity. 44 percent more code, and not much more finished work
A Harvard study of 655 firms shows that AI-written code waits for review. What it means for developers and for owners who order software.
Part of the guideClaude Code, the complete guide. From installation to your own skills and plugins
In this article
In short
A Harvard study using data from 655 firms shows that after AI agents arrive, developers write 44 percent more code, while finished tasks grow by only 13.7 percent. Code waits for review, and the time from submission to approval grows by 56.1 percent. Code review is the bottleneck, so give AI small tasks, read the plan first and let the machine run the tests.
A developer with an AI agent finishes a change before lunch. Then the pull request waits two days, because the only person who knows that module has four more like it in the queue. In a Harvard study this is exactly where AI coding productivity gets stuck.
The study is by Fiona Chen and James Stratton of Harvard University, titled "Artificial Intelligence in the Firm. Bottlenecks in Software Production". We use the 8 November 2026 version. The paper is not peer reviewed yet, and older versions and press stories carry different numbers, so it is worth reading the latest one.
What they measured
The authors used data from Jellyfish, a platform firms use to track the work of their engineering teams. That is 445 million events from GitHub, Jira and Google Calendar, from 720,542 people at 655 firms, from January 2021 to August 2026. They compared firms that adopted AI tools earlier with firms that adopted them later.
A year after adopting AI agents, firms write 44 percent more lines of code and open 35.7 percent more pull requests. Resolved Jira issues grow by only 13.7 percent, and that result is borderline significant. Completed larger projects grow by 18 percent. The authors found no detectable effect on employment.
There is a lot more code. There is a bit more finished work.
Where AI-written code gets stuck
The authors point to code review. After agents arrive, the time from submitting a change to its approval grows by 56.1 percent. Code waits longer before anyone opens it, and then the review itself takes longer. Each change gets 38.6 percent more comments.
AI review tools have not solved this. 79.1 percent of firms in the study already use them, and most reviews and comments are still written by people. People also tend to review only code they know. A senior engineer's change is more often reviewed by another senior engineer.
According to the authors, engineers spend 30 to 40 percent of their time actually writing code. The rest goes to planning, review and testing. AI speeds up the first part, and the rest keeps the same pace or slows down.
How we work with AI-written code
We use Claude Code every day, and we know this bottleneck from our own work. A few habits help us read less and with more confidence.
Plan first, code second. For a bigger change Claude Code works in Plan mode. It reads the project and writes a plan, and we accept or correct it before a single line is written. Reading half a page of plan takes less time than reading a change across ten files. The permission modes are in our article What Claude Code does beyond chat.
Tests before review. The CLAUDE.md file says how to run the tests and that they must pass before a commit. A person only sees a change that has already passed everything a machine can check. How to write that file is in our CLAUDE.md article.
Small changes. One task, one branch, one thing to check. An agent happily fixes something next door along the way. We move those fixes to a separate change, because every extra line is an extra minute of someone's review.
A helper for the dull part. A subagent reads long logs and test output in its own context and hands back a few lines of summary. We described this in Subagents in Claude Code. Whether a change goes in is still a person's decision.
What it means for a business owner
If you employ developers or order software from someone, do not measure progress in lines of code or commits. That is exactly what AI raises most easily. Ask about finished, working features, and about who checks the code and how before it reaches your clients.
It also pays to make sure one person is not the only reviewer for the whole system. In the study, that was the situation that held code up the longest.
If you want to see where AI could speed up work in your business, also outside programming, book a free AI audit. It is a conversation and a written list of places, with no obligation.
What would you like to change in your business?
Leave your number. We will talk about your process and see whether we can help. The first call is free.
The service this text is about
Automation and AI / OCRAutomation of document re-typing, OCR and checking rules, and your data stays with you.