When I started my career, I had very little. Trial and error was the curriculum, and the tools I work with today didn't exist. I failed out of school before eventually returning. There was no shortcut, just a lot of figuring it out.
I think about that often when I see the conversation about AI and entry-level work, because the narrative tends to flatten into either panic or dismissal. What I actually believe is more complicated: the generation entering the workforce right now has access to something I didn't. The ability to experiment cheaply, build things quickly, connect with people across industries, and learn from a volume of information that would have been unimaginable a decade ago. Curiosity, which has always mattered, is now more rewarded than ever, because there is simply more to be curious about.
That doesn't mean it's easy. The readings this week make clear that the structural challenges are real, and the knowledge transfer mechanisms that used to build careers quietly in the background are under genuine pressure. But I don't think the right response is nostalgia for an apprenticeship model that had its own serious limitations.
The question I keep coming back to, and I don't have a single answer for it, is how practitioners and professionals at every stage of their careers navigate a moment where the path forward is genuinely less prescribed. For those early in their careers, the runway for experimentation has never been wider. For those further along, the work is figuring out how to transfer what you know in a world where the old scaffolding is shifting. There's no one-size-fits-all here. But curiosity, resilience, and a willingness to build in the open still look like the most durable assets I know.
THIS WEEK'S READING:
A new working paper from IESE Business School builds an economic model around a deceptively simple observation: tacit knowledge cannot be fully codified. As the paper puts it, "no set of instructions can ever be complete... judgement, dexterity, experience, and other forms of tacit knowledge inevitably come into play when technique is executed." When AI automates the tasks where junior workers would have acquired that judgment, the pipeline breaks, quietly and gradually.
The paper also surfaces a subtler point: labor isn't just about output. It produces knowledge. Novices working alongside experts diffuse best practices across the organization, and that diffusion is what sustains long-run productivity growth. AI systems optimized for verifiable task completion can't replicate it, because tacit knowledge is, almost by definition, not easily verified. The result is a productivity curve that spikes early and erodes later, with the paper estimating a potential drag of 0.05 to 0.35 percentage points off annual U.S. per-capita output growth.
Why It Matters: The damage doesn't require entry-level employment to fall. It only requires that novices get matched with worse mentors, or skipped over for the tasks where deep learning actually happens. That's a structural shift that won't show up in hiring dashboards.
The research describes the mechanism. The next piece shows the scale.
Goldman Sachs economists tracked 16,000 U.S. job losses per month attributable to AI, with workers aged 22 to 25 absorbing a disproportionate share. The same economists were careful to note the true aggregate impact is likely smaller than their estimates suggest, which matters, because the headline number tends to flatten out what's actually a more complicated picture.
Here's the complication: Gen Z is the cohort most natively fluent in AI tools. The same group absorbing the most displacement is also the group most likely to be running AI agents, building side projects with LLMs, and entering the workforce with AI literacy that many of their managers are still catching up to. Displacement and capability are landing on the same generation at the same time.
Why It Matters: This isn't a story about a generation being left behind. It's a story about a transition that's moving faster than the institutional structures around it, and a generation that may be better equipped to navigate that transition than the headlines suggest.
The displacement is real. So is the question of who's responsible for addressing it.
MIT Technology Review names the pipeline problem directly: when AI absorbs the entry-level work, organizations don't just lose headcount capacity, they lose the mechanism by which expertise gets built and transferred. The long-term case for investing in that pipeline is clear. The short-term incentives to do so are not.
This is where job tenure becomes a complicating factor. When average organizational tenure keeps declining, the ROI calculation on developing junior talent gets harder to justify. Why build a two-year development arc for someone who may leave in eighteen months? The article doesn't fully resolve this tension, and that's honest, because the tension is real. It points toward systemic redesign without prescribing exactly how organizations make that pencil out.
Why It Matters: The knowledge transfer problem and the retention problem are compounding each other. Solving one without addressing the other is only a partial fix.
What does it actually look like to redesign the apprenticeship? One CEO has been thinking about it.
Written in late 2025, this piece from Thomson Reuters CEO Steve Hasker covers much of the same ground as the other readings in this issue, and that's worth noting. The concerns about apprenticeship erosion, tacit knowledge loss, and the need for deliberate redesign were already being raised at the leadership level before the displacement data fully materialized. The conversation isn't new. The urgency is.
Hasker's framing is useful for practitioners and professionals thinking about their own organizations: the question isn't whether AI changes how junior talent develops, it's whether leadership is being intentional about what replaces the old model. Most aren't, yet.
Why It Matters: When a CEO of a major knowledge-industry firm is making this argument publicly, it signals that the organizations best positioned to adapt are the ones already asking the structural questions.
The last piece offers a counterweight, and a useful one.
This piece doesn't dismiss the displacement data. It contextualizes it. Wages in AI-exposed fields are rising. Students are adapting, not retreating. And the distinction the article draws is one worth sitting with: when AI is used to augment a worker, output and employment tend to go up. When AI is used as a pure automation substitute, headcounts drop. The technology is the same. The organizational choice is different.
That nuance is easy to lose in a conversation that tends toward either panic or dismissal. The honest picture is more granular: this is a transition with real costs and real opportunities, distributed unevenly across roles, organizations, and cohorts, depending largely on how AI gets deployed.
Why It Matters: The variable isn't the technology. It's the intent behind how it's used, and whether the people closest to the work have any say in that.
Across these readings, a consistent pattern emerges: the short-run math on AI automation often looks compelling, and the long-run costs are diffuse enough to be easy to ignore. Tacit knowledge erodes slowly. Apprenticeship pipelines don't collapse overnight. The productivity drag accumulates over years, not quarters, which means it rarely triggers the kind of organizational alarm that drives real redesign.
Earlier this year, Bryan Dosono, Jennifer Deal, and I co-authored a panel at ACM SIGMIS-CPR '26 that approached this same set of questions from a practitioner framework. The argument we kept returning to was that the transition from creator to curator only holds if someone, somewhere, preserved the original craft. If the pipeline that builds creators gets hollowed out, curation becomes a skill built on borrowed time. The full paper is here if you want to go deeper on the sociotechnical side.
The question worth sitting with isn't whether to use AI. It's whether your organization is being intentional about what it's also choosing not to automate, and why.
Have a take on any of this? Reply directly, I read every response.
Until next Tuesday,
Charles Costa, MLIS