The cost of moving fast

For most researchers, the work ends in identifying and understanding a problem. Matt Beane wanted a solution. So he started SkillBench.

What he set out to solve was an answer to this: As companies race to prove returns on billions of dollars in AI spending, many are losing a pipeline of talent as they trim junior ranks in favor of automation. 

The result, argues Beane, a professor of technology management at the University of California Santa Barbara, is an emerging skills crisis: Organizations are optimizing for today's productivity at the expense of tomorrow's expertise.

It's a paradox that sits at the center of Beane's research, which appears in his book “Skill Code: How to Save Human Ability in an Age of Intelligent Machines,” and every week, companies are coming to him with the same lament.

“They recognize that there’s a threat, and they worry not just about de-skilling junior people, but also just keeping the flow of junior talent going at all,” says Beane. 

SkillBench is his own answer to the problem. Using telemetry and AI-powered analytics, the company aims to help organizations measure whether their AI adoption is building workforce capability or quietly eroding it, and what to do about it.

“Right now, companies have no way to know if you're getting AI right,” he says. “That's what we're trying to fix."

We caught up with Beane to hear what he’s finding in the field and what it would actually take for organizations to stop making the wrong bet.

This interview has been edited for length and clarity.

You spoke several months ago about the risk in eliminating junior roles—both for organizations as well as for individuals at all stages of their careers. That was even before Claude Code and a raft of new layoffs in the name of AI. What has changed since then?

In really important ways, nothing has changed, and that's a huge problem. I talk with four to six organizations a week at the leadership level about AI transformation, and the best players are still using 20th-century tools to make sense of a 23rd-century opportunity. A really good leader is going around talking with people, breaking rank, creating labs, running experiments, rewarding failure, even trying aggressively to hire junior people. They recognize the threat, not just de-skilling their own junior people, but keeping the flow of junior talent going at all. But they're doing all of it manually. I haven't run into a firm that's trying to automate the learning process itself. That's what you need to do to keep up.

What's holding them back, in your view?

A really wonderful mentor of mine is very fond of saying, “I've never seen a better way to waste money than trying to improve a process you don't understand.” The fact is, we don't really understand the work we're doing. We never really have. We've always relied on job descriptions, standard operating procedures, and training materials that become stale almost as soon as they're created. AI is accelerating change so quickly that those tools are being ripped to shreds day by day. Everybody bought a gas pedal, but nobody has a speedometer or a steering wheel. 

How do you square that with the fact that many companies are under pressure to prove AI return on investment?

So, 2025 was the year of: Okay, let's get serious and come up with use cases. 2026 is the year the board started asking the CTO and CHRO: Where's the ROI? The fastest answer is to look at AI exposure studies, decide some functions should be 50 percent more productive, and conclude you need half as many people. I'm not saying that's fundamentally wrong or deeply misguided. It just should be the last move a leader makes, not the first.

And things are so hectic right now. There's so much change going on, so much chaos in the economy that the default maneuver, reflected in the data, is to hold on to your senior talent. They’re good at dealing with rough waters. We've got rough waters ahead. Keep them. Junior people are slower and make more mistakes and need more management. But the organizations that take that approach are going to lose. 

New research

The mid-market's $16 billion drain

Get the report

What makes you so sure that strategy backfires?

Because even I used to talk about skill development as distinct from or complementary to productivity, but in fact, all that skill development is is the change in your productivity over time. So if you're getting better at being good, that learning curve for an organization or for an individual is kind of the same thing, and meanwhile how much better are you getting at being productive over time. 

All of these layoffs in some way might enhance long-term productivity. I'm not going to stand here and say that they won't. But there’s not nearly enough research. I’ve studied the knock-on effect of robotic automation on surgery on the profession's capacity and capability, and that is not a pretty picture. They took this deal and sacrificed learning on the altar of productivity. Robotic surgery boomed, but as a result, you have fewer surgeons out there fit for duty, and among those that are fit, fewer surgeons are out there operating. And then, of those surgeons, there are fewer who are confident and competent with the new method, and the public is paying the price for that. Surgeons are paying the price for that. Hospitals are paying the cost for that. 

So, if I told you, would you take a bet where your productivity goes up now by 1.4x as a firm, but your new ceiling for productivity is either set or will go down slowly over time? You'd probably say, “No, I'm not going to do that,” but that's effectively the deal many organizations are making.

Are workers themselves worried about this tradeoff?

I’ve talked to lots of software engineers—good ones—who are very concerned about trading away their skills and their future ability growth, their learning curve, in exchange for quick results from AI. They're trying to get work done faster while making sure that tomorrow they're better at the important things—taste, discretion, problem selection, architecture—the kinds of capabilities that AI can execute against but not originate.

And a good team leader is thinking, how do we work together as a team, so that we're dramatically more productive while also making sure that everybody's getting better at the important stuff, not at just shipping more content?

So what does it take for humans to get better at that important stuff? Is there a role for AI in that?

Humans need a place they can go where, just by using AI, I get better, and that includes higher quality relationships with other human beings. That's how humans get good at stuff. You can't just get better at home with a screen and no human interaction. That's not how humans function. At the very least, we don't get motivated enough. Most of us need contact with other people—to impress them, or to not want to embarrass oneself in front of them—that stuff drives us to get better. So you need a way of using AI that drives you to getting better through human relationships, not in spite of them.

But trends like remote work and AI assistants both reduce day-to-day interaction. What are we losing?

Right, going virtual doesn't help there. You’ve got to do more. I run a virtual company and you have to do a lot of extra work to make up for that, in particular to cultivate the kind of learning by collateral contact that happens when you’re all in an office. For example, you walk down the hall and ask someone to get you a file about a case, and by the way, to find that file, you had to physically go through four other folks, and in the process maybe you discover something interesting or you learn some stuff. That is how you used to learn. I'm not saying let's bring back the card catalog, but we do have to find some ways to re-inject this human contact back into this digitally mediated work we live in now in a way that enriches both sides of the divide.

Doesn't AI make it easier for senior employees to simply do everything themselves?

The “I could always do it myself” temptation is a recipe for hair-on-fire management and micromanagement. That was true in 2015, but now, because that manager has Claude Code, they can actually go and crush a lot of this stuff dead in software. I'm seeing this up close. A senior dev can do two weeks of work in two hours and has no need to reach out to a junior person. Anthropic's own internal report on the effect of AI use was that people aren't going to each other anymore. Because it's not just about results, it's about what makes work worth it. You learn something from newbies, you learn how they see the organization. You can call BS on stuff that you take for granted. Which executive learned TikTok on their own? None—they all learned it from an admin. You have to have this flow both ways, otherwise the organization starts to get sclerotic.

How do you actually get organizations to prioritize skills development when they're under pressure to move fast?

You should have to take a hit on your bonus if you don't manufacture that kind of interaction that supports the expert, supports the junior person, and gets the work done, even if in some cases it slows things down. That's what a resident in the operating room is in a hospital. You could hire another surgeon, it wouldn't cost that much more. But a teaching hospital is deliberately going to slow things down only a little to get somebody up to speed. We've got to start making that deal in other organizations, too.

What’s an example like that when it comes to skilling workers around AI? 

People need guidance on whether they're using AI well—and whether they're using it in ways that help them grow over the long term. We need something like fitness rings for AI: a way to understand not just output but whether we're building better judgment and better habits. It also needs to be social, it needs to be so that I can learn from you and you can learn from me. The goal isn't simply to become faster. It's to become better. Right now it's just a lot of motion over progress.