The AI coding market is growing rapidly. [Photo: Shutterstock]
At Google, AI now generates 75 percent of newly written internal code, while human engineers review it, shifting the development system.
On April 22, Business Insider reported that Google disclosed the figure and said it is moving overall engineering work toward a more autonomous, AI-centred flow.
The figure shows Google is adopting AI coding faster than expected. Google said as recently as October 2024 that about 25 percent of internal code was generated by AI, and it previously explained that the share rose to 50 percent last fall. That means it has expanded to 75 percent of new code in about half a year.
Google has required developers to increase their use of AI not only for coding but also for other tasks. In a recent blog post, Google CEO Sundar Pichai (순다르 피차이) said the company’s engineering organisation is moving to a “true agentic workflow.” Under that structure, rather than engineers handling every task directly, more autonomous AI takes on some work while people focus on management and review.
Google also presented an example of improved productivity. Pichai said a complex code migration carried out recently by agents and engineers together was completed 6 times faster than when engineers alone handled it a year earlier. That means AI is being used beyond a simple support tool, including for large-scale software maintenance and migration work.
Google engineers are currently using the company’s Gemini model for code generation. Some employees were given separate goals for AI use, and that item is set to be reflected in performance evaluations this year. That shows AI adoption has become an organisational operating standard rather than an option.
Inside the company, tensions have also emerged over the use of other AI coding tools. Some Google DeepMind employees have been permitted in recent months to use Anthropic’s Claude code, and friction among employees was reported to have arisen in the process. The situation, in which Google uses external tools alongside its own model, shows that tool choices and performance standards within its AI development organisation are still being worked out.
In this flow, Google’s development organisation is shifting its focus away from writing code directly and toward validating AI-generated output and jointly designing complex tasks. As AI takes on most new code writing, the key challenge ahead is likely to be less about the share of code generation itself and more about who reviews and evaluates it, and by what standards.
Google’s case shows AI has moved beyond a development support tool and entered a stage of changing the overall way engineering operations work. As the share of new code writing shifts quickly toward AI, the focus is expected to centre on how to establish review responsibility, quality management and organisation-wide standards for use, rather than on generation itself.