Six months to working AI, six years of controls: the gap in the JPMorgan cohort data

A firm founded in 2019 ran its own books and answered its own email for more than six years before a model got near the workflow ^[1]. Whatever review habits it has were paid for in those months, mostly by accident rather than design. The 2025 cohort reaches the same point in about six months ^[2], which leaves at least 66 fewer months of unassisted operating history to reason from ^[11].

Rhett Power, writing in Forbes, argues that the tools amplify whatever discipline already exists, and that automation brought in without structure scales blind spots ^[4]. For a ten-year-old company that is a warning about retrofitting. For a six-month-old company it is a different problem: there is very little accumulated practice to amplify, so what gets scaled is whichever assumption the founder held on the day of the rollout. The speed that removes the setup cost also removes the stretch of time in which somebody would have noticed the assumption was wrong.

Marko Kling of Serrala, who the column says has spent 17 years on enterprise automation strategy, makes the structural version of the point: firms treat oversight as a culture problem when it is a design problem ^[5]. His formulation is that "trust erodes quickly when accountability cannot keep pace with automation" ^[6]. The operational translation is a naming exercise, done before the switch is flipped rather than after something breaks: who owns the outcome of this automated process ^[7]. That costs nothing at rollout and is expensive to reconstruct in front of a customer, a regulator or an investor.

The Zendesk figure in the same piece shows where the design decision actually gets made. The company says its newest AI agent is built to resolve up to 80% of support issues on its own ^[9], and the column notes that the remaining fifth is the messy, emotional or high-stakes end of the queue ^[10]. Look at who draws that line. Whatever the agent declines to resolve is, by definition, what a human sees. Absent a deliberate policy, a vendor's confidence threshold is setting the company's escalation rule, and it will not be written down anywhere a board can read it.

Which makes the measurement question the practical one. Power says he has watched founders treat the amount they have automated as the measure of success ^[13], and proposes asking at review time what people chose not to automate ^[12]. That question is cheap and hard to game, because the system under review cannot generate the answer. The 2019 cohort had six years of operating to work out which judgments were worth keeping in human hands. The 2025 cohort has to settle that inside the same six months it takes to wire the tools in ^[2], and that is the part of the compression that does not show up in an adoption curve.