The real risk with automation isn't that it makes you slower or more error-prone in the moment. It's that it quietly lets your judgment atrophy, and by the time you notice, the instinct for what "looks wrong" is already gone. We've seen the posts, and we've felt the tension ourselves: the data cleaning tool that saves you an afternoon, the dashboard that builds itself, the agentic workflow that runs on its own. Each one is a small win for speed. But each one also moves you further from the messy, hands-on work that used to teach you how the numbers were supposed to behave.
The MIT SMR piece you referenced gets at something important: augmentation beats replacement, but only when the human stays engaged enough to intervene. That's not a feel-good slogan. It's a practical requirement. If you're not the one spotting the weird distribution or questioning the dodgy groupby, you're not actually supervising the machine. You're just hoping it's right. And the longer those pipelines get, the more autonomous they become, the harder it is to audit them after the fact. By the time something quietly goes sideways, you're not fixing a data problem. You're untangling a process you never fully understood in the first place.
What bothers us isn't the tool. It's the quiet shift in who does the thinking. Junior work disappears, and that's framed as progress. But someone still has to catch the failures, and increasingly, that someone is a person who has spent less time in the weeds themselves. The ethical question isn't just about jobs. It's about whether the people left in the loop are actually equipped to catch what the model misses, or whether we're all just pretending they are. That's not a philosophical aside. That's a daily operational risk.
So here's the concrete point: if you're leaning on automation, build deliberate checkpoints where you do the work yourself. Not every day, not every task, but often enough that the patterns stay fresh. Run a manual sanity check on a sample. Rebuild a dashboard from scratch every few weeks. Question one assumption per pipeline, out loud, in writing. That's not Luddite nostalgia. That's maintenance for your own judgment. Because the moment you stop being able to tell the difference between a good result and a plausible one, the automation isn't augmenting you anymore. It's replacing the part of you that actually knew what you were doing.