ChatGPT shipped in November 2022 and the second AI stage opened.
The change from tool to assistant was a vocabulary shift before it was an architectural one. You could now talk to the thing in plain English. You could ask it open questions. You could iterate. The interface looked like a chat window because the metaphor was a chat with a knowledgeable junior. That metaphor sold. Two years later every white-collar workflow had at least one tab open with a chat in it.
Underneath, the loop had not changed. A human asked. The assistant answered. The human decided what to do with the answer. The assistant did not act between turns. It did not have a calendar or a queue. It did not check on anything while the human was asleep.
The assistant stage compressed white-collar productivity faster than any tool stage shift in history. Drafting time fell. Research time fell. Code review time fell. Customer-support time fell. Most of the headline AI productivity studies you read are measuring the assistant stage, not the agent stage, because the agent stage barely had data yet at the time those studies were designed.
The assistant stage also kept the org chart untouched, for the same reason the tool stage did. The human still owned every action. The assistant could write a draft, the human still had to send it. The assistant could propose a plan, the human still had to commit it. There was no seat for an assistant on the chart, because a seat implies accountability for outcomes, and an assistant only owned its turn of the conversation.
What changed at the assistant stage was the speed at which an existing org could absorb information. A 20-person company with assistants could do the analytical work of a 200-person company without assistants. That gap was real and mattered. But both companies still ran on the same human-driven coordination layer. The 20-person company was faster because each human was faster, not because the coordination layer was different.
This is the stage most large companies are still in, and most of them are still proud of it. They have AI strategies, AI committees, AI risk assessments, AI vendor approvals. They are buying assistants for thousands of employees and measuring the productivity lift per seat. The numbers go up. The chart does not change.
Big eats little is failing at the assistant stage, but slowly enough that most operators are still not feeling it. A 2,000-person company with assistants is still beating a 200-person company without assistants. The compounding has not yet flipped because the coordination layer in both is the same, and the coordination layer is the part that scales unevenly.
The flip happens at the next stage.
The first time an AI does not wait for a human to ask, the org chart bends.
This is post 2 of 4 in the From Tool to Robot series.
Previous: AI was a tool first
Next: The agent is the first AI that ever joined the org chart