Managers Don’t Get Years to Learn the Job Anymore

I spent twenty years growing in one software organization, and what I know about judgment I learned from great mentors and by getting things wrong in ways I could recover from.

An integration I built shipped with a bug that took most of a day to track down. In that time, it wrote bad data that had to be corrected. Another time, I misunderstood a customer’s requirements and sent a quote for work that turned out to be wrong, and I sat through the uncomfortable call that followed. One of the worst mistakes early on was giving a customer a firm delivery timeline that I quickly found out was never going to work.

None of those were good days. But they were survivable, and each one taught me something I could not have absorbed from being told. The mistake was the lesson, and the mistake was small enough to learn from without taking down anything that mattered for long.

It is tempting to look back on that and call it a slower, gentler era. It wasn’t. The work was fast and high pressure. While AI has increased the speed and pressure, one of the biggest differences now is how widely a mistake can spread.

Early in my career, a lapse in judgment stayed roughly contained. One bad call produced one bad outcome, and there was usually time to catch it before it spread. Greater responsibility brought higher-risk decisions, but that was after I had learned how to make better decisions. Automation has changed the shape of that. When a person directs an agent and misses something, the error does not sit politely in one feature waiting to be found. It propagates, potentially fast, before anyone looks at it closely. The tooling we have built is good at catching the kind of error that is mechanically wrong, the test that fails or the type that doesn’t match. It is far worse at catching the error that is plausible, confident, and simply wrong. That kind passes review for the same reason a person accepts it. It looks right.

So the thing that used to build judgment, the small recoverable mistake, has gotten more expensive to allow at the exact moment we need people to develop judgment faster. This squeeze lands hardest in two places.

The first is on the individual contributor. We are now asking people early in their careers to act as managers of agents. They set the goal, they review what comes back, they correct it, and they decide whether it is good enough to use. That is supervisory work, and it leans on discernment, the ability to look at a confident answer and know whether to trust it. Discernment is built through experience, and we are asking for it before the experience has had time to accumulate.

The second is on the frontline manager, who now carries a harder version of an already hard job. They are responsible for building that discernment in their people faster, and for keeping output high regardless. They are doing that while the executives above them, under their own pressure to show results from these tools, expect more and expect it sooner. The manager is the point where the pressure from below and the pressure from above meet.

Discernment is the visible edge of a larger thing, which is judgment, and there is early research suggesting that edge is more fragile than we assume. In a recent Wharton study, researchers gave people confident answers from an AI that were deliberately and undetectably wrong. Participants went along with the wrong answer around eighty percent of the time. Their accuracy tracked the AI’s rather than their own, and when the AI was wrong they performed worse than they would have with no AI at all. The researchers call the pattern cognitive surrender. It is recent work and the experimental conditions were specific, so I hold the number loosely, but it points at something I recognize from practice. The hard part is not getting people to use the tools. It is keeping their judgment engaged while they do.

Management and AI are both evolving, so there is a long way to go before this is solved. Moreover, I am suspicious of anyone claiming they know where it will land this early. What I am fairly sure of is that the old model, where a manager had years to grow into the role one survivable mistake at a time, is not coming back in its old form. The runway was already shrinking before AI, for reasons that had nothing to do with it, and AI has shortened it further while raising the cost of the mistakes that used to do the teaching.

In the next post I want to take this from the manager’s chair specifically and look at what changes when the people you are developing are learning judgment under these conditions. For now I will leave it as the problem I keep seeing, stated as plainly as I can. The role is asking for judgment earlier than the work can safely build it.