Insights

5 Takeaways from Our Conversation With Cassie Kozyrkov, the Architect of Google’s AI-First Transformation

AI is reshaping how technical organizations think, build, and operate. In our recent Blueprint Plus+ session, Cassie Kozyrkov, CEO of Kozyr and former Chief Decision Scientist at Google, shared what she learned leading Google’s AI-first transformation. Her insights point toward a future where AI is less about efficiency hacks and more about reimagining how work gets done.

You can watch the whole episode here.

In the meantime, here are our top five takeaways.

1. AI = Complexity. Don’t just automate, look for totally new solutions.

“When executives say AI adoption, I want them to hear complexity adoption… The reason we are using very complex tools is that there is no other way to solve what we need to solve.”

Cassie points out that AI’s value isn’t simply in replicating what you already do, but faster. Its best upsides are tackling problems so complex that they were previously impossible even to imagine. Traditional workflows assume humans can define every step in advance. AI breaks that constraint and allows teams to rearchitect processes that were limited by human memory and the capacity of a human mind. 

This idea is also something that McKinsey noted back in 2024: the largest value comes not from automation, but from reinventing workflows end-to-end. It shouldn’t be about getting rid of jobs, but about trying totally new things. 

2. When it comes to trusting AI outputs, testing isn’t enough. You need safety nets.

“Don’t trust. Don’t even trust the test… Wrap everything in safety nets so that even if things go wrong, you can mitigate problems.”

Don’t get her wrong, testing is important. However, Cassie’s point here is that testing can sometimes be too incomplete or reveal issues too late, especially if you’re using agentic AI that can take action proactively. She thinks that trust must be built into the structure of your engineering workflows. That means implementing layered safeguards, including monitoring, fallback logic, human-in-the-loop review, and clear boundaries around what the AI is allowed to do.

For engineering teams, this means shifting their focus from “Did this code pass the tests?” to “What happens when it doesn’t?”

3. For AI adoption to be successful, you need leaders in the room.

“The big cultural signal is leadership participation… If the people who know what’s worth doing aren’t there from the beginning, your AI project won’t change anything.”

AI transformation is not just the domain of the IT department. It needs buy-in from everyone, especially leaders. Cassie warns that AI efforts fail when leaders assume engineers alone can bridge the gap between what’s technically possible and what’s strategically valuable.

Her experience at Google demonstrates that when leadership, UX, reliability engineering, and technical teams work together from day one, organizations produce solutions that couldn’t have existed before. Without that strategic alignment, AI becomes a tool-chasing exercise instead of a transformation.

Cassie is clear that leaders don’t need to understand the math behind billions of parameters, but they do need to understand the possibilities and the potential impact for things like developer experience and productivity. 

4. AI is about more than pinching yesterday’s pennies. It’s about having superhuman memory.

“What data gets you is superhuman memory and superhuman attention… It allows us to automate what we cannot come up with by thinking through the steps.”

AI is fueled by data, and when it’s leveraged properly, it can perform sophisticated tasks with datasets that might exhaust a human. AI-first teams let the technology handle the cognitive load while humans focus on judgment and creativity. The competitive advantage comes from expanding what the organization can perceive and retain, and reducing errors in repetitive projects.

This might help explain why our State of the Developer Ecosystem report found that the top things that developers want to delegate include converting code to other languages (which is important for modernizations and migrations), documentation, and writing boilerplate code. 

5. 2026 might be the year the open office plan has to change.

“Imagine everyone talking into their phones or laptops in an open office… It’ll be very weird. We may need to redesign our spaces for natural interaction modes.”

Voice assistants are booming. According to a 2025 Market.us forecast, the global market for them is projected to grow to roughly USD 79 billion by 2034.

Cassie notes that many people think at the speed of speech, not typing, and as tools evolve, speaking to AI agents will become a normal part of day-to-day knowledge work. But as she also points out, you can’t have 60 developers all dictating to their agents in one open-plan room. From 2026 onward, she thinks that office spaces will be better optimized for these kinds of interactions. Blueprint++ is a new series where we interview leading thinkers in business and technology about topics like responsible AI adoption, the latest trends in agentic AI, measuring ROI, and transforming the developer experience. If you’re curious, check out future episodes on the JetBrains YouTube channel.