The Mentorship Gap: Why Reps Beat AI for Junior Estimators

Walk into almost any precon department right now and you’ll see the same tableau. On one side of the room: a 60-year-old chief estimator with three binders, two monitors, and a coffee cup older than most of the plans on his desk. He’s muttering about steel prices in 2008 because the current quote “feels off by about nine percent.” On the other side: a 25-year-old analyst, two years out of school, who just asked an AI tool to level fifteen subcontractor bids and is staring at a clean output that looks, well, clean.

The senior is going to retire. The junior is going to inherit the estimate. And somewhere in between those two chairs is a knowledge transfer we are not doing. I’ve been watching this gap widen for a while now, and I’ve landed on a single rule that I think the industry needs to tattoo somewhere visible. Learn the work. Then automate it. Not the other way around.

The mentorship gap in construction isn’t a staffing problem. It’s a knowledge transfer problem. We are handing juniors tools that generate answers faster than they can learn to question them, and we’re doing it right as the people who could teach them to question walk out the door.

---

The Cliff We’re Walking Toward

The demographics are not subtle. NCCER estimates that roughly 41% of the current construction workforce will retire by 2031. That’s not a slow leak. That’s a dam. And the replacements aren’t showing up ready: in the AGC’s 2025 Workforce Survey, 57% of firms reported that available candidates lack essential skills or required licenses.

Put those two numbers in the same sentence and you get the real problem. Retirements are taking the instinct with them, and we’re replacing it with autocomplete.

I want to be careful here. I’m not doing the “kids these days” bit. The junior estimators I meet are sharper with software at 25 than I am now, and there are four-year-olds with iPads who could probably out-click all of us before lunch. That’s a gift. The issue isn’t the talent. The issue is the sequence. When a senior estimator flinches at a number, that flinch is the output of a thousand Tuesdays of being wrong about concrete, a decade of negotiating with the same handful of subs, and a few jobs that taught them what “acceptable risk” actually feels like when the weather turns. You cannot download the flinch. You have to build it. And the senior estimator who has built it is carrying a database nobody else in the building can open.

---

You Can’t Be the Human in the Loop if You Don’t Understand the Loop

Everyone loves saying “keep a human in the loop” when they talk about AI in estimating. It’s the phrase that makes the room nod. But nobody asks the obvious follow-up: what if the human doesn’t know what the loop is?

A Microsoft Research and Carnegie Mellon study presented at CHI 2025, surveying 319 knowledge workers across 936 real tasks, found that higher confidence in generative AI was associated with less critical thinking, and that workers “refrain from critical thinking” when they lack the skills to inspect the AI’s output. Read that sentence twice. The tool doesn’t just skip the work. It quietly skips the thinking about the work, too.

Now translate it to our world. A junior who has never cold-called thirty subs to chase coverage on a mechanical package cannot feel when an AI-leveled bid sheet is missing a line. A junior who has done that exercise once, painfully, catches the gap in ten seconds because something on the page doesn’t smell right. Same output. Two different humans. Only one of them is actually in the loop.

This is the car analogy I keep coming back to. Cruise control is a beautiful piece of engineering, but you still want the person behind the wheel to have driven in the rain before. Otherwise you haven’t automated driving. You’ve automated confidence.

---

The Reps That Build the Instinct

So what do we do about it? We stop treating the boring parts of the job as waste. At least the first time through. Here is my short list of reps I think every junior estimator should still do by hand, once, before they’re allowed to hand that task to a model.

• Manual takeoff on a small, self-contained scope. Doors and frames, or site concrete. Small enough to finish, real enough to hurt when you miss a count.
• Calling subs to chase coverage. Not emailing. Calling. You learn more about a trade partner’s bench and mood in one phone call than in a week of portal notifications.
• Leveling a bid package line by line. Actually comparing scope inclusions and exclusions with a highlighter and a pot of coffee. This is the single most formative exercise in a young estimator’s first two years.
• Missing a scope item and feeling the cost on a Tuesday. I wish I could assign this one. You can’t. But when it happens, senior estimators should frame it as tuition, not failure.

None of this means juniors should spend three years doing by hand what a tool can do in three minutes. That’s the other ditch, and plenty of firms are still in it. I’ve written before about what AI can and can’t credibly do for estimators today, and the honest answer is: it can do a lot, but only for someone who can smell when it’s off. The reps are how you build the nose.

---

Scale the Senior, Don’t Replace the Junior

Here’s the part the industry keeps getting backwards. Most firms look at AI and ask, “how do we use this to get by with fewer junior estimators?” That’s the wrong question, and the BLS projects cost estimator employment to decline 4% from 2024 to 2034 partly because that’s the question everyone’s answering. The better question: how do we use AI so our seniors stop being chief spreadsheet updaters?

Think about what a senior estimator actually carries. Thirty years of knowing which sub is good for the number they put on the page and which one is going to come back in week six looking to renegotiate. The memory of exactly how that escalation clause bit a hospital job in 2011. A read on which drawings are optimistic and which ones are going to generate an RFI the size of a small novel. Which materials are holding, which prices are about to move, whose handshake still counts, and how long to wait before picking up the phone when a number looks too good. None of that lives in a database. Most of it isn’t even written down. It shows up as a hunch, a pause, a quiet “hold on, run that one again.” That pause is not vibes. That is three decades of scar tissue doing its job in real time, and it is the most valuable thing a firm owns.

Because that’s what’s actually happening in most precon departments I visit. That instinct, the one we just described, is spending Tuesday afternoon reformatting a subcontractor’s PDF into a column that matches the template. That is not what we hire senior estimators to do, and pretending otherwise is how a firm’s bench quietly ages out.

If AI takes the senior’s time back from the spreadsheet, two things happen. First, their judgment gets to touch more deals, which is how a firm’s win rate actually moves. Second, they have time to sit next to the junior and explain why the number feels off by nine percent. That second thing is the entire ballgame. The org chart we should be drawing isn’t “senior plus tool replaces three juniors.” It’s “tool frees senior to train two juniors into the next seniors.” One of those org charts has a five-year horizon. The other has a cliff.

The junior estimator isn’t the thing to automate. They’re the next senior you’re kneecapping if you do.

---

The Short Version

We are not short on software. We are short on the patience to build the humans who use it well. If you run a precon team, the move is the same whether you have two estimators or twenty. Put the reps first. Put the tools in their hands second. And protect the senior’s time like it’s the rarest input on the job, because it is.

Learn the work. Then automate it. Not the other way around. If you want to see how we’re building Buildr’s platform around that idea rather than against it, come take a look. Bring your chief estimator. And your 25-year-old.