Editorial illustration for GM Engineers Now Spend Just 15% of Time Writing Code After AI Overhaul
GM Engineers Spend 85% Less Time Coding With AI
GM Engineers Now Spend Just 15% of Time Writing Code After AI Overhaul
Rashed Haq oversees autonomous vehicle engineering at General Motors, and by his account, the people writing code for GM's self-driving software spend most of their week doing everything but that. Speaking at VB Transform 2026, held at the Hotel Nia in Menlo Park, California, Haq put the number at 15%. The rest goes to sifting through vehicle data, tracking down bugs, running experiments, and checking whether fixes actually hold up.
That shift didn't happen by bolting a chatbot onto GM's existing setup. Haq said the division rebuilt its engineering workflows around AI agents that handle chunks of that non-coding work directly, and the payoff shows up in the numbers: about three times as many merged pull requests across the AV engineering group, quicker releases, and fewer bugs slipping through to later stages of development.
The 15% figure lines up with older data on how developers actually spend their time, well before agentic AI entered the picture. GM's case, though, points to what happens when a company stops treating AI as a coding sidekick and starts using it to redesign the workflow itself.
The result, Haq said, is roughly three times as many merged pull requests across GM’s autonomous vehicle engineering organization, faster releases and fewer defects escaping into later stages of development.
Why this matters
Haq's 15% figure is the most concrete data point we've seen yet on what "AI-native" engineering actually looks like inside a company shipping physical products, not just SaaS. Tripling merged pull requests sounds like a win, but that metric alone doesn't tell us whether GM's autonomous vehicle code is getting better, worse, or just bigger. Pull request volume has always been a shaky proxy for engineering output, and it gets shakier when agents are triaging, testing, and running experiments alongside humans.
For founders and engineering leads watching this, the real question is what GM's review and validation process looks like now that machines are doing most of the pre-code work. Self-driving software carries safety stakes that a typical web app doesn't, so a tripled merge rate needs scrutiny, not applause. If GM is willing to talk numbers this specific on stage, we'd want to see the failure rate, rollback frequency, and incident data alongside it.
Until then, treat this as a workflow story, not proof that AI agents are ready to own critical decisions in safety-critical systems.
Related Reading
AI Tools & Apps Claude gains shared context in Excel, PowerPoint; Microsoft adds Copilot Cowork
AI Tools & Apps Windows Copilot AI unable to pinpoint image source in user test
AI Tools & Apps LG's recent webOS update adds Microsoft Copilot app, now removable
From Archives Runway's AI video bug becomes a feature, guided by LLM context.
From Archives Survey: 700+ CS Educators in 49 Countries Rethink AI-Era Testing
Common Questions Answered
How much time do GM's autonomous vehicle engineers now spend actually writing code after the AI overhaul?
According to Rashed Haq, GM engineers now spend only 15% of their time writing code, down from significantly higher percentages previously. The remaining 85% of their time is spent on activities like sifting through vehicle data, tracking down bugs, running experiments, and verifying whether fixes are effective.
What specific improvements did GM achieve by redesigning its engineering workflows around AI agents?
GM achieved roughly three times as many merged pull requests across its autonomous vehicle engineering organization, faster software releases, and fewer defects escaping into later stages of development. These improvements demonstrate the tangible impact of integrating AI agents into their engineering processes rather than simply adding a chatbot to existing workflows.
Why is GM's 15% figure significant compared to other companies' AI implementations?
Haq's 15% figure represents the most concrete data point available on what "AI-native" engineering actually looks like inside a company shipping physical products, not just software-as-a-service solutions. This provides rare transparency into how AI is transforming real-world engineering workflows at scale in the autonomous vehicle industry.
What limitations exist in using merged pull requests as a metric for GM's engineering productivity gains?
While tripling merged pull requests sounds like a significant win, this metric alone doesn't indicate whether GM's autonomous vehicle code is actually getting better, worse, or just larger in volume. Pull request volume has historically been a shaky proxy for engineering output, and becomes even less reliable when AI agents are handling tasks like triaging, testing, and running experiments.
Further Reading
- GM's AI Design Slashes Car Development Time In Half - IEEE Spectrum
- GM's AI tools could cut the car development timeline in half - Fast Company
- How GM uses AI to improve vehicle software – and safety - GM News
- Study finds AI tools made open source software developers 19 percent slower - Ars Technica
- How much does AI impact development speed? An empirical study of 21% faster software development with AI - arXiv
