The Autonomous Issue

Every vendor demo now ends the same way. The agent books the meeting, updates the CRM, and sends the follow-up, all on its own.

The question is not whether an agent can do a step. It is which steps it can do without a human, and where handing it the wheel quietly makes things worse.

This issue follows that line.

The feature is the closest thing to a real answer we have. A team ran a randomized experiment inside a live customer service operation. The AI cut the average handling time, but it lowered customer ratings on the chats it took, and the human rescue only worked for technical problems, not for a frustrated customer. Worse, the workers assigned to supervise the AI checked out. Autonomy in a real operation is a band, not a switch, and the human handoff is only as good as a human who is still paying attention.

The research corner explains why the band is narrow. In a study of long tasks, a model’s accuracy on each step fell as the chain got longer, and once its own earlier mistakes were in front of it, it made more. The problem was not that the model could not reason. A long chain compounds small errors into failure, and the model gets worse the moment it has to build on its own bad output.

The strategy piece prices it. A benchmark of real office tasks, each worth about two hours of human work, found that AI agents are already much cheaper and faster than people, and not yet as good. Cost is solved. Quality is the gap. So the question is never can it. It is: is it good enough here, and what does a mistake cost.

Then two operators show both ends of the line. Klarna automated two thirds of its customer service, then its CEO said the cost cutting had gone too far, quality dropped, and the company started rehiring people. Aaron Levie, one of the most bullish voices on agents in the enterprise, still designs for humans on top of them, not humans removed.

The instrument turns all of it into a map. Sort every step of your motion by how long the chain is and how much a mistake costs. Automate the short, cheap, reversible steps. Keep a person on the long or high-stakes ones. Then design the handoff so the human on the other side is still awake.

Autonomy is not a switch you flip. It is a line you draw, step by step, and keep moving as the models get better. See you Tuesday.