A manager looks at a team’s output over a quarter and sees a gap that wasn’t there a year ago. Two people are shipping more than they used to, at a pace that wouldn’t have been possible a year ago. Two others are shipping about the same as always, which now looks slow by comparison. The team is the same team as last year on paper, but half of them are using the AI tools and half aren’t, and that’s where the gap came from.
A team where AI use is optional splits into two shapes. On one side are the people who have adopted the tools and folded them into how they work. They spend a lot less time on the kinds of tasks that took them most of the week last year, and they ship more than they used to. On the other side are the people who haven’t, either because the tools make them uncomfortable or because they don’t think the tradeoff is worth it. They work the way they always worked, which produced respectable output last year and looks slower now.
The heavy user gets a big output boost and a smaller increase in skill than the output would suggest. Because the model does a chunk of the work, the heavy user spends less time in the part of the task where the learning happens, which is the struggle of producing the first pass yourself. The output stays high, and the person running the tools isn’t building the understanding they used to build when they did the work by hand. Six months in they look strong on output and a bit thinner on the kinds of questions a senior asks.
The refuser keeps the old learning curve. They spend the same hours on the same kinds of task, and they come out of the year with the judgment that came from doing the work the slow way. What they lose is output. The team around them is producing more, and at the same hours of work the refuser’s output now looks like less of a contribution than it did last year. If the company reads the team on output alone, the refuser looks like a performance issue, while they’re doing exactly what they were rewarded for doing a year earlier.
The old management frame reads the team on output and on visible effort. Output is easy to count, and effort is what the manager sees in the meetings. Both signals are now broken. Output is broken because the heavy user is producing more without doing more of the work, and the refuser is producing the same work for less output. Effort is broken because the heavy user looks productive by shipping a lot and the refuser looks slow by shipping less, but neither signal maps to the quality of the judgment either of them is building.
A management frame that handles the new shape has to read the team on two things at once. One is the volume of work the person is producing, which is what the old frame already read. The other is the shape of the judgment the person is building, which the old frame didn’t need to read because everyone did the work themselves.
The second read takes longer than the first. The manager has to watch how the person works, rather than only what they produce. For the heavy user it means watching whether they review the model’s output with care, or whether they trust it too much and ship things they shouldn’t have. For the refuser it means watching whether the judgment they’re building from the old way of working is the judgment the company still needs, and whether their slower output is a cost worth carrying for that judgment.
Optional AI use on a team is a thing that has to be managed rather than left to sort itself out. The team that manages it gets both kinds of value: a slice that ships at high volume, and a slice that ships at the old rate but carries the judgment the team will need for the harder work. The team that doesn’t manage it ends up with a thinning bench and a skewed read on who’s performing. A manager who notices the split and reads both halves honestly gets a team that runs on two tempos at once and knows why, while the one who reads on output alone ends up with a team that all looks like the heavy user, and a year of onboarding loss they won’t notice until the next hard problem lands.
Pillar V
**Edtech / Learning Systems**
> *Two memos on what makes a learning system work, and the operator > lesson for corporate training.*