The central argument: why increasingly capable models and machines still need an institutional operating layer.
I work on the part of AI that begins after the model is good enough.
For nearly a decade, through Drive TLV and Drive Europe, I have worked across hundreds of collaborations between technology startups and global industrial organizations, in automotive, mobility, energy, manufacturing, logistics, and semiconductors.
The same family of failures kept repeating. Strong technology succeeded in a demonstration or pilot, then stalled at the threshold of production. Some pilots were shelved. Some were renewed year after year without ever crossing. Some shrank until nothing depended on them. Different endings, same place.
I saw the pattern in the hundreds. I verified it in fourteen: the Physical AI deployments I could observe closely enough to code. Eleven technically successful pilots never reached sustained production. The model was the primary cause in only two.
The obstacle was rarely the technology. It was the absence of an operating environment that could answer basic institutional questions:
Who owns the outcome? What is the system authorized to do? When must it defer to a person? What evidence survives an incident? How does the organization learn from experience? What remains when the model or vendor changes?
These questions are usually treated as deployment details, to be resolved after the technology works. They are part of the system itself.
That realization became a question:
What must exist around an intelligent machine before an institution can responsibly allow it to act?
Everything on this site is an answer to that question.
Physical AI is usually described through models and machines: the intelligence that decides and the body that acts.
Neither supplies what an institution needs before it can let that system participate in consequential work, where a bad output does not remain an output, but becomes an event.
A system can be extraordinarily capable and still not be deployable. Capability answers what a machine can do. It does not answer what it may do, how its authority is earned, what evidence must survive its actions, or who answers when those actions produce consequences.
I call the persistent operating layer that answers those questions the Habitat.
A Habitat is more specific than context or everything surrounding the model. It:
A model can make action possible. The Habitat makes it permissible, repeatable, and accountable.
I am the author of The Case for Habitat and the originator of Agent-Habitat Dynamics, the research and operating framework behind the thesis: how intelligent machines progress from impressive capability to trusted participation in the physical world.
This is not an argument for slowing AI. It is an argument for building the conditions under which institutions can responsibly allow it to do more.
For more than 25 years, I have built technology companies, backed frontier systems, taught computer science, and helped global industrial organizations move emerging technology into real operations. The work has placed me on both sides of deployment: inside the startup that must prove a new technology works, and alongside the institution deciding whether to trust that technology with machinery, safety, capital, and reputation.
Earlier in my career I co-founded and led ClickFox, an early customer-experience analytics company that grew to a valuation in the hundreds of millions of dollars. That experience taught me how quickly a technical insight becomes an organizational problem once it enters a real institution.
For more than a decade I served as an adjunct associate professor at the Georgia Institute of Technology’s College of Computing. I am an inventor on multiple patents and have published across artificial intelligence, analytics, mobility, and the institutional adoption of emerging technology.
Today I am a founding partner of Drive TLV and Drive Europe and founder and General Partner of NextGear Ventures, where I work with entrepreneurs building at the boundary between software and the physical world.
Builder, investor, teacher, deployment partner. Different roles, one problem: how does a promising technology become something society can safely, economically, and institutionally depend upon?
Three connected parts.
Fieldwork. Working with technology companies and industrial operators on the practical conditions for deploying Physical AI in real environments.
Research. Developing Agent-Habitat Dynamics into a testable framework for authority, evidence, adaptation, and institutional learning.
Public argument. Writing about the economic, legal, strategic, and societal consequences of intelligent machines entering the physical world.
One aim across all three. Intelligent machines are entering work that carries real consequences, and people will be working alongside them. Whether that makes human life better depends less on how capable the machines become than on whether trust, authority, and accountability are built into the environment around them. That machinery does not yet exist. I am working on it.
If you are new to the work, these are the best entry points:
Tal Cohen is an entrepreneur, investor, author, and the originator of the Habitat thesis for Physical AI. He is the founder and General Partner of NextGear Ventures and a founding partner of Drive TLV and Drive Europe. Over more than 25 years, he has built technology companies, invested in frontier systems, taught computer science, and worked across hundreds of collaborations between startups and global industrial organizations. He previously co-founded and led ClickFox and served for more than a decade as an adjunct associate professor at Georgia Tech’s College of Computing. He is the author of The Case for Habitat and the creator of Agent-Habitat Dynamics.
Tal works between Tel Aviv and Atlanta.
For research, speaking, collaboration, or media inquiries: tal@talcohen.ai