fter working on the launch of Claude 3 through Claude 4, Michael Gerstenhaber left Anthropic to join Google five months ago, where he's now focused on bringing AI to more people and organizations.
During his time at Anthropic, Gerstenhaber became invested in the value of AGI and the importance of sharing it with the world. Given its vertical stack, Google was better positioned in the space to achieve that goal.
“So I left because I accidentally got AGI pilled along way, Dario [Amodei, Anthropic’s CEO] has a very specific effect on people, and I believe that the technology is one of the biggest of our time, probably the biggest,” said Gerstenhaber. “Distributing the technology has become, if not a moral endeavor, a very exciting endeavor for me because of its importance.”
Like OpenAI, Anthropic is racing toward AGI, but the two companies frame their missions differently. Amodei has spoken out about the risks of AGI, including the displacement of entry-level white-collar jobs. At the same time, OpenAI explicitly centers AGI as its goal. We reached out to Anthropic for comment on Gerstenhaber's assessment, but the company did not have a response.
At Google, Gerstenhaber serves as Vice President of Product for Vertex AI and Agents, the company's platform for building and deploying AI in the enterprise. The role puts him at the center of Google's AI cloud infrastructure, everything from inference APIs to agentic capabilities, where he works directly with customers to find the right solutions.
“At Google, we do have that ability to distribute. We're the only Cloud that's vertically integrated among the power plants with the data centers, with the TPUs in the data centers, with access to the smartest models in the world, whether it's ours or my former colleagues, and the platform itself with customers on the cloud,” said Gerstenhaber.
He has already seen AI drive meaningful workflow transformations across companies, including through agentic solutions. For instance, he cited a large pharmaceutical company that delegated statistical analysis and coding of clinical data to agents. Another example was Thomson Reuters’s development of agentic products, such as CoCounsel and Westlaw, for legal research.
He acknowledged AI agents haven't reached their full expected value, not because the technology isn't ready, but because of trust issues. Organizations lack clear ways to define scopes, struggle with accountability when AI fails, and can't easily evaluate whether workflows are performing correctly. His advice for implementation? Take bite-sized steps.
“People should find the scope over which they don't need a human at all, and that might be a very narrow scope, not a very ambitious scope," said Gerstenhaber and, "then you'll widen the aperture from there.”