The Pipeline You Don’t Know You’re Cutting

There’s a version of this story that looks like good management. Junior roles doing work that AI can now do faster. You make the call. Costs down, velocity up, senior people freed from reviewing mediocre first drafts. On paper, it’s a clean trade.

Most leadership teams took it.

I’m not saying that was irrational.
I’m saying it wasn’t complete.

What aviation learned after the fact

In the 1980s and 90s, the airline industry automated the cockpit. The logic was sound: autopilot reduced human error, improved fuel efficiency, let pilots focus on what actually required judgment. Crashes went down. The data was unambiguous.

What wasn’t visible was what was eroding underneath it. Pilots were spending less time flying and more time monitoring. Monitoring is not flying. The pattern recognition, the physical instinct, the situational awareness you build from thousands of hours of manual control — none of that develops while watching a screen.

By 2009, researchers had a name for it: automation dependency. In surveys conducted across the commercial aviation industry, 77 percent of pilots reported their skills had deteriorated. Seven percent felt they had improved.

The FAA launched a major investigation after a string of accidents, including Air France Flight 447, where pilots faced an unexpected stall at 38,000 feet and couldn’t recover. Not because the plane failed. Because they hadn’t flown manually enough to know what to do when the automation handed it back.

The industry hadn’t removed pilots.
It had removed the conditions that keep pilots capable.

What makes this pattern dangerous is that the aviation industry didn’t discover the problem when they made the call to automate. They discovered it years later, when the system failed and nobody in the cockpit knew what to do.

The gap between when the pipeline breaks and when the deficit shows up is long enough that the two events don’t look connected. By the time the problem is visible, the decision that caused it feels like ancient history.

I don’t think most organizations see themselves in this story.

They should.

What was actually in that role

Junior positions in knowledge work aren’t primarily about output. The output is a byproduct. What those roles are actually for is absorption.

A junior designer, researcher, or strategist spending two years inside an organization is learning things that can’t be transferred through a document or an onboarding deck. How decisions actually get made versus how they’re supposed to get made. Which constraints are real and which are negotiable.

What “good” looks like in this specific context, with these specific people, under these specific pressures. And they’re being wrong at low stakes, repeatedly, which is the only way anyone builds judgment about what good actually costs.

When you replace that work with AI output, the output looks fine.
That’s the problem.

In June 2025, MIT Media Lab researchers published a preprint study tracking brain activity in students writing with and without AI tools. Students using ChatGPT produced clean, well-structured essays. They also showed lower brain activity, weaker memory retention, and less ownership of what they’d written. The output was indistinguishable. The formation didn’t happen.

You can’t compress the reps. The work of becoming someone who can make good decisions inside a specific organization is time-dependent in ways that don’t yield to efficiency tools. Compressing it doesn’t speed up the process. It skips it entirely.

That practice isn’t sustainable long-term.

What the research says about organizations actually winning with AI

Here’s where the argument gets uncomfortable for the people who made the call.

A 2025 MIT Sloan Management Review study examined how high-performing organizations were actually deploying AI. The finding doesn’t match the story most leadership teams are operating from.

The organizations getting the most from AI weren’t using it to get answers faster. They were using it to design better decision environments — what the researchers call Intelligent Choice Architectures. The system surfaces options leadership hadn’t considered, identifies tradeoffs that would have stayed invisible, and structures the choice space so human judgment gets applied to something actually worth judging.

The deliverable from AI isn’t the answer. It’s the set of choices. And the value of that set depends entirely on whether the people working with it know which questions are worth asking.

That isn’t a junior skill. But it isn’t purely a senior skill either — it’s organizational fluency. The kind that takes years to develop by being inside a problem, watching decisions land, understanding why a particular constraint exists even when nobody can articulate it anymore. You build it by doing the work at every level. You don’t build it by watching AI do it.

The study put it plainly: framing the problem is more consequential than solving it. If your organization has been systematically removing the people who were in the process of learning to frame problems, you now have a fast machine pointed in directions that nobody can fully evaluate. That is a different kind of risk than the one on your radar.

The trade you actually made

You didn’t just cut costs when you compressed junior roles.
You cut the absorption period.

The senior people in your org have the judgment they have because they spent years doing the work that junior people now don’t do. That judgment isn’t renewable on your current timeline. When those people leave, or move up, or burn out, the people behind them haven’t done the work. They’ve watched AI do it.

This is what makes the competence lag so hard to catch. Your organization looks fine right now. Output maintained, velocity up. The problems show up later, when you need someone to step into a senior role and the person next in line has sharp execution skills and almost no organizational judgment. Or when a major decision gets made confidently and nobody in the room had the depth to see what was being missed.

By then, the spreadsheet that made the original call will be three budgets ago.

Not an argument against AI

To be clear: this isn’t nostalgia. Not an argument for protecting junior roles as a matter of fairness, not a case against AI, not a suggestion that you go back to doing things by hand.

The aviation industry is dramatically safer than it was before autopilot. Nobody serious is arguing otherwise. But the FAA’s response to automation dependency wasn’t to remove autopilot. It was to mandate regular manual flying. Scheduled disconnection from the automation. Deliberate, recurring practice of the skills that the system was letting atrophy.

The organizations that get this right won’t be the ones that slowed down on AI. They’ll be the ones that treated judgment development as infrastructure — and budgeted for it the same way they budget for the tools.

If you’ve been cutting the pipeline, you’re not buying speed. You’re borrowing against the judgment of people who did the work you’ve now decided isn’t necessary.

That debt comes due.