As AI tools have moved into daily work, most planners and product managers pictured roughly the same division of labor: AI handles the repetitive, simple tasks, and people handle the creative, higher-order judgment calls. Writing AI drafts a document so people can focus on strategic thinking; code-completion tools fill in the boilerplate so developers can spend their time on more complex design work — or so the thinking goes.
That expectation isn't entirely wrong. Many professionals have experienced firsthand how AI-drafted emails speed up replies, and how AI cuts the time spent writing reports that follow a repetitive format.
But this expectation rests on a precondition: people have to be able to properly evaluate what AI produces. The division of labor only makes sense if the reviewer can actually review. And that evaluative ability doesn't simply maintain itself. It fades gradually, as time accumulates without doing the work directly.
The Ability to Review Work Comes From Having Done the Work Yourself
On July 6, 2013, Asiana Airlines Flight 214 struck a seawall short of the runway while landing at San Francisco International Airport. Of the 307 people on board, three died and 187 were injured. The accident report the U.S. National Transportation Safety Board (NTSB) released the following year cited the pilots' over-reliance on automation as one of the primary causes. When the aircraft switched from autoland to manual mode, the report found, the pilots' feel for manually controlling altitude wasn't sufficiently engaged.
Shortly after the report came out, the U.S. Federal Aviation Administration (FAA) issued a safety directive in August 2013 urging airlines to deliberately increase pilots' manual flight time. The reasoning: even when autopilot handles most of a normal flight, the instincts needed for manual control in an abnormal situation don't hold up without regular practice.
What's worth noting here is that the pilots involved in the crash weren't incapable of manual flight from the start. They could do it during training. It was the accumulation of time spent not doing it that dulled the instinct. The skill hadn't disappeared — it had simply gone dormant.
The same structure applies to knowledge work. When AI drafts a proposal, writing speeds up. But where does the ability to look at that draft and spot that "this paragraph's logic is weak," or "this claim needs supporting evidence right here," or "the reader will get stuck at this point" — where does that come from? It comes from the pattern recognition built up over hundreds of instances of wrestling with structure and revising it yourself. Hand the drafting over to AI, and that training process gets skipped.
For the first six months, or even the first year, this skipped step doesn't show. The instincts you've already built up are still there. It's only after that period that they start to fade — and there's no way to know exactly when, until the moment you actually need them arrives.
What Repetition Builds Inside You
Break down the process of writing something yourself. Facing a blank page, you decide which argument to put first. You think through whether this argument needs to come before that one for the reader to be convinced. You predict which paragraph needs a concrete figure, and where the reader is likely to lose interest. These decisions happen continuously throughout drafting. And it's through repeating these decisions that your standards take shape.
When AI supplies the draft, those decisions get made inside the AI. The person retreats to the position of receiving the result and saying "good" or "not good." But the ability to tell the difference comes precisely from having made those decisions yourself. Once that experience stops accumulating, your evaluative ability is left running only on what you'd already built up.
Developers who write their own code learn to recognize the symptoms a particular bug tends to produce. If AI both generates the code and finds and fixes the bugs, where does that pattern recognition come from? PMs who write their own market analysis reports develop a feel for which metrics are substantive and which are superficial. Let AI write the report, and the work goes faster, but the eye for questioning the analysis's underlying assumptions never develops. Content directors who set their own angle on a topic build a sense for which approach will land with readers and which will fall flat. Hand that decision to AI, and the core of editorial judgment starts to hollow out.
In every one of these cases, what repetition accumulates is an internal standard for judging the result.
Within the first six months, the two paths don't produce a noticeable difference in output quality. The gap shows up when AI makes an unexpected error, or when a complex situation arrives that's too hard to hand off to AI.
Deciding, Deliberately, What Not to Hand Off to AI
Just as turning off autopilot and flying manually all the time isn't a realistic solution, cutting back on AI tools isn't the right starting point for this problem, either. There's no reason to give up efficiency — the advantages of the tools are there to be used.
What you can do is deliberately set a standard for which tasks to hand to AI and which to keep doing yourself.
The starting point for that standard is simple: does the experience of doing this task myself build a core part of my competence? If so, it's worth doing it yourself, even at some cost to speed — it pays off in the long run.
For a planner, that's the stage of building a report's logical structure — deciding which argument goes where, and in what order the evidence needs to appear to be persuasive. Hand this stage to AI, and you're outsourcing the core of planning itself. For a content director, it's the stage of setting the angle — deciding what stance to take on a given topic, for which readers, and which scene to open with. For a PM, it's the stage of establishing the rationale behind priorities — sorting out which metrics are substantive to a decision and which are merely supporting.
What these stages have in common is that they aren't about producing the finished output — they're about deciding what to make and how.
Everything else — turning a settled structure into finished sentences, polishing the format, cleaning up data that follows a repeated pattern — can go to AI without doing much damage to your core competence. Once you draw that line, tools work to extend your skill rather than erode it.
In an environment where smart devices and software spread quickly, the ability to operate a tool skillfully and the ability to accurately judge what that tool produces are two different things. The first spreads fast; the second only comes from experience you've done yourself. Deciding for yourself which tasks you'll keep building that experience on — that's how you hold onto your skill as automation goes deeper.
