AI Is Taking Over The First Three Years Of The Job. Year Four Has To Be Rebuilt On Purpose.

The case for pointing AI at junior work is easy to make and usually goes unchallenged: the first years of a technical career are full of repetitive, low-judgment tasks, and automating them frees people for higher-value work. What the argument skips is where senior judgment comes from in the first place. The tedious early tasks are more than just output a company needs. They're the reps through which a novice becomes an expert. Remove them and the immediate math looks like efficiency, but the long-term consequence is a pipeline that stops producing the very people whose judgment makes the automation safe to trust.

Contemplating this problem is Ted Rusch, Vice President of Preconstruction at Batson-Cook Construction, an Atlanta-based general contractor that builds across the Southeast. He's led the company’s preconstruction group since 2018, following roles at JE Dunn Construction Group and Brasfield & Gorrie, and he's candid that the apprenticeship question is one he has no clean answer to.

"AI is going to replace the first or second or third year person doing menial tasks. And that’s a great thing," Rusch says. "But how do I get the first-year person to a fourth-year person if they’re not doing those three or four years of tasks?"

The reps that make an expert

Rusch’s worry is not that AI handles the drudgery, but that the drudgery was doing something nobody priced in. The monotonous early work of an estimator, like counting items on a set of plans or quantifying linear feet of wall and pipe, is exactly the work he'd most like to offload, and also exactly the work that teaches a young estimator how a building goes together. "By the time they’re at the end of it, they’re boring and monotonous and nobody likes doing them," he says of those early tasks. "But if somebody’s not learning those basics, how do we make up for that?"

He's clear that instead of refusing the technology, the definition of early-career development has to be rebuilt rather than deleted. The industry hasn't done that work yet. "What that’s really going to do is change the definition of what year one, two and three are. That’s going to take some time to figure out," Rusch says.

Philip Samuelraj, Founder and CEO of AI delivery firm Techjays, thinks the gap is real but counters that the thing worth preserving was never the manual reps. In his view, the skill that made a senior valuable was always the judgment layered on top of the task, and that's what early careers now have to be redesigned to teach directly. "The manual counting was a slow way to build judgment about how the work fits together. If that task goes away, you have to teach the judgment on purpose instead of waiting for it to accumulate as a byproduct," Samuelraj says. "The new version of year one is learning to direct the system and catch where it's wrong, not doing by hand what the system now does."

Samuelraj sees it as a deliberate design job: the apprenticeship isn't disappearing so much as moving up a layer, from performing the task to supervising and correcting the agent that performs it.

Why the judgment is still required

The reason the lost reps matter is that the systems replacing them can't yet be trusted unsupervised, which means human judgment isn't optional at the point of review. Rusch is direct that his team still checks the machine’s output, because the cost of a missed detail in a budget is real and can be significant. "I don’t trust the data without review. It’s very much a trust-but-verify scenario right now."

What that verification actually requires is the kind of pattern recognition that only comes from having done the work. He points to a real example from a set of plans, an eight-inch drain pipe that looked correct at a glance and in reality was anything but. "Look at that, it’s doing everything it should. But then, oh wait, it's coming out the other side of the building and it’s eight feet in the air, not four feet underground," Rusch recalls. Catching that isn't a task you can hand to someone who's never learned to read the plans, which is precisely why the apprenticeship can't simply disappear.

Samuelraj’s firm builds systems for exactly this problem. What moves a system past the trust-but-verify stage, he says, is instrumenting the work so the checking is built in rather than left to a human to perform every time. He acknowledges that manually re-checking outputs is where everyone starts, "but it doesn't scale, and it burns the expert’s time on the wrong thing. You instrument the system so every action it takes is observable, you run an eval layer that scores it against what the experts would've done, and when it gets something wrong, that becomes feedback that trains it not to repeat the mistake. That's how it earns your trust."

Rusch, from the field, keeps a human in the loop because the data is not yet dependable. Samuelraj, from the build side, treats the human checking as the raw material for making the system dependable enough that the checking can eventually move up to the boundaries rather than every output.

The point of the time saved is human work

Rusch frames the upside of automation in a way that depends on people still having the expertise to use it. Rather than reducing headcounts, the goal is moving skilled people off counting and onto the coordination that only humans do well. "We look at it as a tool that can free people up to do people things," he says, describing estimators moved out of heads-down takeoff and into rooms with architects and trade contractors to work through how a building actually gets built.

He's deliberately wary of letting that reclaimed time be read as slack to cut, and says he's intentionally careful with leadership about how it gets measured and redeployed rather than pocketed as headcount savings. The value shows up only if the freed-up people have the judgment to spend that time well, which loops back to the apprenticeship problem: the coordination work he wants people doing is senior work.

The expertise hierarchy is inverting

One unexpected effect is that AI has scrambled who learns from whom. For decades, seniority ran one direction in construction, with thirty-year veterans holding knowledge that juniors had to absorb over time. Fluency with AI tools has flipped part of that relationship, and Rusch sees his most experienced people reaching for help from the least experienced. "Believe it or not, it’s the older people who are recognizing that they need the younger more so than the opposite direction."

That inversion is healthier than it sounds, because it replaces a one-way hierarchy with an exchange. The veterans hold the domain judgment the models can't replicate, while the younger staff hold the comfort with the tools. Rusch describes his own posture as refusing to be the holdout. "I’m 50. I’m trying not to be the guy who fights it off till it’s too late and then I’m useless in five years. I’ve seen that happen with people, and I don’t want that to be me."

Permission to be bad at first

What makes the exchange possible, in Rusch’s telling, is a culture that treats early incompetence with a new tool as expected rather than punishable. His own agent-building attempts have mostly failed, and he says so plainly, because the alternative is a workforce too afraid to learn. "One of my favorite sayings is, don’t be afraid to suck at something new," he says. "I think we’re in the sucking-at-something-new phase."

For Rusch, the apprenticeship gap and the permission to experiment are two ends of the same problem. AI can remove the tasks that used to build expertise, but it can't yet supply the expertise itself. Rebuilding how people learn rather than assuming the pipeline takes care of itself is the work that separates a durable AI strategy from a short-sighted one.