When a VP starts pulling a number through an AI model instead of asking the analyst who used to answer that question, the org chart doesn’t change, but the work does. The analyst is still there. The team is still there. But the question that used to travel from the VP to the analyst and back, carrying context and sometimes returning with a small correction the VP didn’t know they needed, now loops inside the VP’s head with a model standing in for the analyst. The VP gets an answer faster. The analyst doesn’t know the question was asked.
The question was never just the question. It carried context the VP hadn’t thought to include, because the analyst already knew which quarter they were talking about and which team had last revised the number. It sometimes came back with a correction: ‘Yes, that number’s right, but we’re not counting pending contracts the same way we were last quarter.’ The model doesn’t have that context, and it doesn’t know it doesn’t have it. The answer comes back clean, which is the problem.
The analyst would have signalled uncertainty in some way. A hedge in the email, or a follow up an hour later after a quick check with finance. The model does none of that. It gives you the same tone for a number it’s confident in and a number it’s reconstructed from adjacent context. Its tone comes from the format, and the format doesn’t change when the underlying uncertainty changes.
This isn’t just VPs and analysts. Anywhere senior people got information by asking someone else, the flow has started to reroute through models. A department head who used to read a project summary from their team now asks a model to condense the Slack channel. A CEO who used to ask a VP to explain a market shift now reads a model’s take on three analyst reports. Each substitution saves time, and each one closes a loop that used to exist between the senior person and the team.
Some part of middle management existed to be the traffic medium for those loops, but not the whole of it. Reports had to be summarized, people had to be managed, resources had to be divided, and political risk had to be absorbed. But part of the middle layer was there so senior people stayed in touch with the business through a running exchange of small questions and small corrections. When that exchange moves into a model, the middle layer looks the same on the org chart. The reports still get written. The role of being the loop between levels quietly stops being anyone’s.
At the senior level, what this does is easy to miss because nothing obvious breaks. Decisions keep getting made and the numbers are close enough. But the answers those decisions rest on are cleaner than the answers used to be, and cleaner isn’t the same thing as accurate. The answer an analyst would have given you over coffee comes with small signs of uncertainty, like a pause or a ‘technically, but’. The model’s answer has none of that, and the senior person acts on it as if it’s been through the same loop as the analyst’s, when it hasn’t.
The pattern is hard to see from inside the senior role because nothing looks worse in the moment. Decisions made on the model’s answers will be made faster and with less friction, and most of them will be fine. The ones that drift off will drift off slowly, in ways that don’t map back to a single bad answer. They’ll map back to a stretch of months during which the person making the decision stopped being corrected by anyone whose job included noticing. The analyst who would have corrected them was never asked. The department head who would have pushed back was briefed on an output the senior person had already acted on.
Rebuild the loops on purpose. Name them, and put them on someone’s calendar the way any other operating thing gets built. A weekly walkthrough with the analyst whose work you’ve been skipping is one shape. A monthly hour with the department head whose judgment you’ve been replacing with a model summary is another. Pick the loops to keep and accept the ones you’re dropping. Do it on purpose, because by accident is how it’s happening now.
The thing to watch for in your own work is the first time you take a model’s number and act on it without any second thought about where it came from. The number might even be right. The signal is that you stopped asking the question that used to travel, and that the loop closed without you noticing.