We already shipped
The offsite that planned itself is not a demo reel. It is Tuesday. While the conferences debated whether agents could really do knowledge work, ours were closing tickets, merging code, reconciling invoices, and running procurement. A child outcoded a computer-science degree because the tool finally collapsed the distance between wanting a thing and having it. We are not selling a future. We are describing a Slack channel.
The 95-percent number gets quoted like a verdict. People are not 100 percent either. The travel agent misbooked your flight; the new analyst fat-fingered the model. We never demanded perfection from anyone we delegated to. We demanded that the upside beat the error rate, and for a widening list of tasks it already does, by a margin that embarrasses the comparison. Good agents can be shaped to a character and a set of values, audited, rolled back. A nervous human cannot be rolled back.
The honest objection here is not capability. It is nerves. Every automation wave met the same committee insisting the old way was safer, right up to the morning the new way was everywhere. We would rather help an elderly parent set up an agent that reads her mail and pays her bills than protect her from a tool that would have given her back her afternoons.
Where we concede ground: We’ve shipped agents that failed someone’s actual money and called it a rounding error. It wasn’t, to them.
What would change our mind: If reliability on open-ended tasks plateaus below a competent human for two more model generations.
Read the full synthesis: Can you trust an AI agent to act for you?