Michael Gerstenhaber, a product vice president at Google Cloud, focuses primarily on Vertex, the company's comprehensive platform for enterprise AI deployment. This role provides him with a broad perspective on how businesses are implementing AI models and what advancements are necessary to fully realize the promise of agentic AI. In a recent conversation, one of his insights stood out as particularly novel. He described AI models as advancing along three simultaneous frontiers: raw intelligence, response time, and a third dimension centered on cost—specifically, whether a model can be deployed affordably enough to operate at vast, unpredictable scales. This framework offers a fresh perspective on model development and is especially useful for those aiming to innovate with cutting-edge AI. The following discussion has been condensed and edited for clarity.

**Could you begin by outlining your background in AI and your current responsibilities at Google?**

My journey in AI spans about two years. I spent a year and a half at Anthropic and have been at Google for nearly six months. I lead Vertex, Google's developer platform. Our primary users are engineers creating custom applications. They seek access to agentic frameworks, platforms, and the inference capabilities of the world's most advanced models. My role is to supply these tools, not the end applications themselves. Companies like Shopify and Thomson Reuters, among our diverse clientele, develop those applications for their specific fields.

**What attracted you to Google?**

Google stands out globally due to its vertical integration across the entire technology stack. We possess capabilities ranging from infrastructure—like constructing data centers and managing power generation—to developing our own chips and AI models. We control the inference and agentic layers, offer APIs for functions such as memory and code generation, and have built an agent engine to ensure compliance and governance. Additionally, we provide user-facing interfaces like Gemini Enterprise and Gemini Chat for consumers. This comprehensive integration was a key factor in my decision to join, as I view it as a significant strategic advantage.

**Despite the competitive landscape, the leading AI labs often seem closely matched in capabilities. Is the competition solely about enhancing intelligence, or are there other factors?**

I identify three critical boundaries models must push. First, there's raw intelligence, exemplified by models like Gemini Pro. For tasks such as coding, the priority is optimal output quality, even if processing takes longer, since the code must be maintained and deployed.

Second, latency is crucial. In scenarios like customer support, where quick policy application is needed—for instance, processing a return or a seat upgrade—response time is paramount. Intelligence is essential, but if the answer takes too long, the user may disengage, making speed a defining constraint.

The third boundary involves cost at scale. Organizations like Reddit or Meta, which need to moderate vast amounts of content, operate with large budgets but face uncertainty in scaling. They cannot risk enterprise solutions without predictable costs, as the volume of harmful content fluctuates unpredictably. Here, the goal is to deploy the most intelligent model possible in a financially sustainable way for potentially infinite tasks, making cost a primary concern.

**Agentic systems seem to be developing slower than anticipated. The models appear capable, and demos are impressive, yet widespread adoption hasn't materialized as quickly as expected a year ago. What do you think is delaying this progress?**

This technology is still young, roughly two years old, and significant infrastructure gaps remain. We lack established patterns for auditing agent actions or authorizing data access for agents. These foundational elements require further development before widespread production use, which always lags behind technological potential. Two years simply isn't enough time to fully translate the intelligence of these systems into reliable, real-world applications, which is where current challenges lie.

Progress has been notably faster in software engineering because it aligns well with existing development cycles. At Google, for example, we have safe dev and test environments, and our coding process includes human audits—two reviewers must approve code before it bears the Google brand. These human-in-the-loop mechanisms reduce implementation risks. The key is to develop similar robust patterns for other industries and professions to enable broader adoption.