The Trajectory
Where does autonomous AI go when the organization has 100 people and society hasn't decided how much to trust it yet?
This page is a live hypothesis, not a finished article. The ideas below are directions we're tracking — come back as the picture sharpens.
Below about 30 people, agent adoption is mostly individual or team-level. One engineer wires up a Claude pipeline; another pair builds a Notion-to-ticket automation. These are tools, not infrastructure.
At 100+, something shifts. Agents stop being the private property of whoever set them up and start becoming shared institutional capability — infrastructure that can outlast the person who built it, that other teams depend on, that shows up in org charts (or conspicuously doesn't). That transition is the difficult one.
Large organizations that move early will likely develop two distinct agent classes: embedded agents that live inside a team's workflow and stay under that team's governance, and cross-functional agents that operate across org boundaries and require central oversight. The second class is where things get complicated.
Cross-functional agents create accountability gaps. If an agent coordinates between Legal, Finance, and Engineering and produces an output that causes a compliance issue, the chain of responsibility is genuinely unclear. Traditional organizational structures — designed around human actors with clear job descriptions — aren't built for this.
Even if every technical capability problem were solved tomorrow, agents wouldn't operate freely. Society sets a ceiling on what autonomous systems are allowed to do — not through explicit rules in most cases, but through an accumulating pattern of trust and precedent.
That ceiling currently sits at a surprisingly low level. An agent that books a meeting is unremarkable. An agent that books a flight is slightly more interesting. An agent that signs a contract on your behalf is crossing into territory most people aren't comfortable with yet — even if they believe the agent could do it correctly.
The distinction isn't about technical accuracy. It's about reversibility, stakes, and legibility. When something goes wrong with a human decision, we have centuries of norms for assigning responsibility and making it right. When it goes wrong with an agent, we don't — yet.
There's a structural shift coming that most organizations aren't thinking about yet. Today, humans coordinate through agents: you use tools to help you get work done and communicate with others. In a few years, the ratio starts to flip in certain domains — agents begin coordinating through humans.
This isn't science fiction. It's already happening in narrow contexts: incident response systems that page the right human at 3am; trading systems that escalate to a human only when a threshold is breached. The human is still in the loop, but the loop's default state is autonomous, with human participation as the exception rather than the rule.
What that looks like at the organizational level — where it creates value, where it creates liability, what new roles emerge — is the question we're trying to think through. This page is where we'll update as the picture sharpens.