You're just changing so much, and the turnover of the code is fast enough that you're rewriting it several times before you ship. It's hard for many people to collaborate on the codebase compared to before, when the code velocity was slower.
As you go outside of engineering, the big challenge I see is access to data. For example, you might want your agent to answer, "How many times a day do companies around the world ask, 'Hey, what's the status of this deal?'" That kind of information is something the company knows and should be easy for an agent to answer. We actually have some cool tools that can answer that pretty well, but habits and data access are real issues. In a bigger company, managing who has permission to access data all needs to be worked out, and then rule definitions come into play. You mentioned how engineering, product management, and design are a bit distant from each other, and sometimes you might want to merge those rules as things improve. That’s my view of where we’ll be in 2026—the models will be capable, but we’ll still only be using them so much.
When it comes to adopting this intelligence, a lot of us are working on exactly what the Gemini Enterprise teams and anti-gravity teams are tackling. They're not working on these problems precisely, but this is the roadmap. We’re using AI internally, running into barriers, and working faster. That’s the product we’re shipping. For example, the SRE team at Google found parts of their job that can have automated workflows, and that’s happening in certain areas. The challenge is doing it systematically—developing skills, centralizing them, and making them available for everyone to use.
Identity and access controls are real challenges, and we're working on them. Security is a major concern, and that adds another layer on top of everything. The cost of mistakes when running these services is high, so we need to handle it carefully. But once we solve these issues, we’ll bring the technology out in a more robust way, and people will be able to do much more.
We also see that other companies are improving these processes, and as things get better, adoption will spread. For example, Google reforecasts its business a few times each year. At Stripe, we do a budget for the year and then generate three formal forecasts. A forecast is a snapshot in time, combining data from people's knowledge and what’s written down—questions like, "How is this product doing? Will this deal close?" AI could potentially do a fully automated, no-human-in-the-loop forecast.
I expect that 2027 will be a big inflection point for these technologies. Even as people check the work in the conventional way for a while, you could switch over fully to AI-generated forecasts. Some of these shifts will happen profoundly in 2027. This connects to the question about whether larger companies or startups will get there first. One advantage for startups is that they tend to have more AI-native teams, and might be able to integrate these tools more quickly. For bigger companies, it requires deeper training and transformation, which can be harder to do.
As for small initiatives inside Google that excite me, probably something that would surprise people: when we decided to do data centers in space, we started as a very small team. That kind of project is what keeps things interesting.
