1. News
  2. AI models & agents
  3. Jellyfish · AI coding
Jellyfish · AI coding2 min read

AI now helps write over a quarter of code changes at a typical company

A study of 1,300 companies finds AI agents touched 27.7% of code changes in September, up from about 2% in January, while people still approve almost everything.

By the Dashline deskSource: R&D World
AI now helps write over a quarter of code changes at a typical company
Photo: Joonspoon, CC BY-SA 4.0

AI now helps write more than a quarter of the code changes at a typical company. Jellyfish, a Boston firm that tracks engineering work, found that AI agents opened or added code to 27.7% of pull requests at the median company in September. A pull request is a proposed change to a software project. In January, the figure was about 2%.

What changed

Jellyfish drew on more than 1,300 companies, 276,000 developers and 100 million pull requests for its October report. The median share of code written by AI tools was 29%. Yet people still sign off on nearly everything. At the median company, only 0.5% of merged changes were approved on AI review alone.

Jellyfish's Nicholas Arcolano said a turning point came around the end of 2025, when coding agents became reliable enough to handle more substantial work.

What a software factory is

The report describes software factories as pipelines where agents pick, write, review and ship the work with little human involvement. They remain rare. At the 90th-percentile company, their share of pull requests rounded to 0.0%. At the 99th-percentile company, they opened 9.4% in September, up from about 1.7% in March.

In most pipelines, people still choose what work gets done. Jellyfish's Nikolas Albarran said handing that choice to an agent signals trust in the surrounding setup.

Who checks the work

Developers are more positive about AI than about how their employers handle it. In a survey of 1,417 developers at 30 companies, 88% agreed AI coding tools helped their work. Only 62% agreed their organization effectively manages the risks of AI-generated code.

Setup matters, too. Engineers who edited an agent's instruction files six or more times merged 2.3 more pull requests a week than those who never did. Those files tell an agent about a project and how to work within it.

What it costs

Usage is not free. By late September, spending on Claude Code, an AI coding tool, reached a median of $63 per active developer per week. Using the report's four-week month, that is $252. At the 90th percentile, the monthly equivalent hit $1,666.

What to do

If your business builds or buys software, expect faster output. Two questions are worth asking your developers or vendors:

  • Who reviews the code AI writes before it ships?
  • How does the company manage the risks of AI-generated code?
The takeaway

AI is writing a growing share of code, so ask who checks it and how risks are managed.

More articles

You should really read these.

Next step

Let’s talk