Sundar Pichai has been quietly updating one number on earnings calls for two years, and I think it's the most important trend line in software. In October 2024 he said AI was writing a bit over a quarter of Google's new code, with engineers reviewing everything. By spring 2025 it passed 30 percent. Last fall it hit half. At Cloud Next this April he said 75 percent. You can argue about what counts as "AI generated," and people do. But nobody looks at that curve and walks away thinking nothing changed.
Cheap code didn't make engineers faster the way everyone assumed. METR ran a randomized trial last summer with experienced open source maintainers working in codebases they knew well, and the result was kind of hilarious. With AI tools the devs took 19 percent longer, while believing they'd been 20 percent faster. The tools work fine. Typing just was never the slow part. What actually eats the time is understanding the system and deciding whether the thing deserves to exist, and AI didn't touch either.
Which explains what's happening to job titles. The engineers who get the most out of agents are the ones who understand customers best, because that's what it takes to hand an agent a good spec. PostHog defined product engineers well a while back. Same goals as a product manager, but they ship code. On the seed to Series B teams we talk to, that's just the job now. The person running agents all day is the same person reading the customer's Slack thread. Nobody changed their title, the work just moved out from under it.
You can see it in hiring too. A few years ago a seed stage team meant eight engineers, a designer, and a PM to keep the machine fed. The teams we meet now run five or six people total. Everyone talks to customers, and the founder still reviews half the PRs. They can afford more people. Past a certain point every new hire adds more coordination than output.
It goes both directions. PMs are getting more technical too. Kevin Weil at OpenAI says writing evals will become a core PM skill, and he's right. None of this started with AI either. Airbnb dropped the classic PM role back in 2023 and merged it into product marketing, well before agents wrote anyone's code. AI sped up a shift that was already happening.
Marty Cagan said the sharpest thing I've read about AI and product. Teams using AI for roadmaps "still have their crappy roadmaps. They're just doing them faster." Productboard's recent survey of enterprise product teams backs him up. 72 percent used AI to generate content but only 18 percent trusted it to prioritize. Teams happily let the model produce things but won't let it decide what to build, because deciding well means knowing your customers cold, and the model doesn't come with that.
So the bottleneck moved. For fifteen years engineering time was the scarce resource, and whole org charts existed to protect it. Now the hard part is knowing what to build. Whoever holds that, engineer or PM or founder, ends up setting direction, whatever their business card says.