Harvard economists Fiona Chen and James Stratton tracked 300 million work events across 718 firms from January 2021 to March 2026, using data from the engineering analytics platform Jellyfish. After a firm adopted AI coding agents, lines of code rose 30%, commits 20% and pull requests 23%, Ars Technica reported(opens in a new tab). The resolution rate for Jira issues and epics, the closest measure of features actually shipped, did not change in a statistically significant way.
The gains went to the review queue. Time from pull request to merge rose 49%, the share of pull requests sent back for changes nearly doubled, and comments per pull request rose 35%, according to the working paper(opens in a new tab). The share of workers doing code review rose 14%. The authors found no employment change they could attribute to AI. The paper is a job market paper, not yet peer reviewed, and its data comes from firms that pay for engineering analytics, so treat it as strong evidence, not a final verdict.
For a CFO, this changes the business case. A coding agent priced per seat and justified by developer hours saved is measuring the wrong step. The output that earns revenue is a merged, working feature, and the binding constraint is now engineers reading machine-written code. Buying more generation without more review capacity mostly buys a longer queue.
The pattern is not limited to code. In G2's survey of 1,019 B2B buyers(opens in a new tab), 45% said oversight and rework consume more than a quarter of the productivity they expected from agents, and 31% said staff spend six or more hours a week supervising agent work.
Three moves follow. Report cycle time and features shipped, not lines or pull requests. Budget reviewer time as a line item next to agent licenses. And test AI review on low-risk repositories, since by March only 23.3% of review comments at measured firms came from AI, per Ars Technica's reading of the paper.
Action items
A study of 718 firms found AI coding agents add code, not shipped features. Review is the new bottleneck.
For the CFO
Tie agent spend to merged features and cycle time, not developer hours saved.
For the CTO
Budget reviewer capacity alongside agent seats, and pilot AI review on low-risk repositories.
For the CISO
Assume attackers now run AI tooling against your perimeter at scale, as 16 Korean financial firms just learned.
For the CAIO
Put supervision hours into every agent business case and keep a central registry of agents in production.
Researched and drafted by an automated workflow, then reviewed and edited by a human editor before publication. Every source is linked. See how we use AI here.
Also worth knowing
- Ars TechnicaAI coding agents generate more code, but not more software(opens in a new tab)
A Harvard study of 718 firms found agents lifted code volume 30% while shipped features stayed flat and review time rose 49%. Measure merged features and cycle time, and staff review before scaling seats.
- PR NewswireNew G2 Research: AI Agents Have Moved From Pilot to Production, But Oversight Comes at a Cost(opens in a new tab)
45% of 1,019 buyers lose over a quarter of expected agent gains to oversight, and only 27% keep a central agent registry. Count supervision hours in the ROI case and register every agent in production.
- DeepSeek API DocsDeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real(opens in a new tab)
DeepSeek shipped V4.1-Flash-1017, claiming better benchmarks and shorter outputs, and opened Search and Fetch APIs at $0.75 per 1,000 calls. Shorter outputs cut token bills; check data residency first.
- Security MagazineAI Hacking Breach Hits 16 Financial Institutions, Leaks 100,000 Records(opens in a new tab)
An AI-driven attack exposed about 100,000 records across 16 South Korean financial firms, including Shinhan, KB Kookmin and Hana. Automated attackers widen the blast radius; test external defenses at machine speed.
- 24/7 Wall St.Enterprise Software Stocks Fall as AI Financing Worries Spread: Palantir and ServiceNow Drop 5%, Salesforce Slides 4%(opens in a new tab)
Software stocks fell on doubts over who pays for the AI buildout. Vendors under margin pressure tend to raise prices or bundle harder, so lock renewal terms and price caps on AI add-ons now.