Margin of Safety #50: Goodbye to $150k baby investment bankers

AI commoditizes many apprenticeship roles like investment banking or software engineering

Last week, an LP of our funds asked us to provide an updated projection of one of our venture funds. As I (Jimmy) was updating the model I had, I had the idea to use Claude and build a completely new model from scratch. Before I was done updating an existing model, Claude had already finished. What would have taken me several hours, and at least 1-2 years to fully internalize back when I was a new banker, was completed in minutes. It did not feel like a marginal improvement. It felt like the early years of toil in my banking days were actually not necessary!

Financial modeling was once a proxy for competence. It required repetition, attention to detail, and a degree of structured thinking that separated the trained from the untrained. Today, much of that separation has collapsed. The floor of human capability has risen sharply. The ceiling, defined by judgment, taste, and decision-making, has not moved nearly as much. This asymmetry is where the real implications begin.

The historical contract of “baby jobs”

Across industries, most junior roles were never designed for immediate productivity.

The “baby investment banker”, or investment banking analyst, was overpaid for Excel and formatting. Compared to investment banking analysts, a resident doctor endured long hours and $60,000 salary in exchange for clinical training (Kathryn: we realize that the medical field has very strange labor economics due to the AMA’s position as what is effectively a modern day guild. But despite that, the asymmetry is hard to ignore). An apprentice electrician accepted lower wages to accumulate practical experience. At the extreme, a young musician might have been paid nothing at all, trading labor for exposure and future opportunity.

The implicit deal was simple. We subsidize you today so you become valuable tomorrow. Higher subsidies (we think) mostly worked because juniors still produced some economic value.

AI breaks the economics

Well, AI changes that equation in a fundamental way. Tools like Claude and Gemini remove much of the low-skill, high-volume work that justified many historically higher-paid junior roles. Tasks that once required hours of manual effort can now be completed instantly, often with fewer errors.

This compresses the early learning curve. Skills that previously required repetition and time can now be accessed on demand. The consequence is structural as junior workers are no longer subsidized by their output. Training becomes more clearly a cost center rather than a blended investment. Highly paid apprenticeship only works when juniors produce enough value to offset their cost. AI weakens that foundation.

The coming compression of junior salaries

If this dynamic holds, compensation structures will need to adjust. The $100,000 to $150,000 entry-level roles in banking, consulting, and software engineering may prove difficult to sustain. Firms will increasingly ask a straightforward question. Why pay a premium for work that can be completed by AI systems at a fraction of the cost? The AI system will also improve over time, not take time off, or require delicate HR policies. (NB: Kathryn is somewhat less pessimistic here: she believes that there is at least a possibility of task-mix shifting to favor areas where AI persistently lags, even at junior levels. Such shift would restore some of the economic value. However, she agrees that junior wages are likely to continue to face downward pressure and the question is simply how much.)

A more likely equilibrium begins to resemble other apprenticeship-heavy professions. Resident doctors, for example, earn modest salaries relative to their long-term earning potential. Apprentices in skilled trades follow a similar pattern. Compensation reflects the fact that training, not immediate output, is the primary purpose of the role.

In this emerging landscape, junior workers are no longer competing primarily with other humans. They are competing with tokens, compute, and model outputs.

The judgment gap

There is a second-order effect that may prove more consequential. If AI systems can handle the bulk of execution, junior workers may skip the repetition that historically built intuition. The “painful reps” that formed pattern recognition, taste, and the ability to make decisions under uncertainty become less frequent.

This creates a potential gap. Output can be accelerated, but judgment may not develop at the same pace. The risk is not that juniors become less productive. In many cases, they will be more productive. The risk is that they become less prepared for senior roles that require independent thinking, contextual understanding, and taste. In other words, AI may compress the learning curve on paper while elongating it in practice.

Early signals from engineering

Software engineering offers a useful leading indicator. In conversations with senior engineers, a consistent theme emerges. The marginal value of interns and early-career engineers declines as AI tools become more capable. Tasks that once served as training grounds can now be completed with minimal human involvement.

This does not eliminate mentorship. Many experienced engineers, and firms, still invest in junior talent, often for reasons beyond immediate economics. But the structure begins to shift. Early-career roles start to resemble residencies. Lower pay, higher emphasis on learning, and a clearer understanding that the role is primarily developmental. There is also a widening distribution of outcomes. The best juniors, those who can think beyond the tools, continue to outperform AI. The average junior, however, struggles to differentiate and the bar keeps getting higher as AI gets smarter.

The future of apprenticeship

If apprenticeship economics are changing, the structure of early careers will need to adapt.

One path is the formalization of training. Firms may introduce structured, lower-paid residency-style programs that explicitly frame the early years as educational. Compensation would reflect the reality that learning, not production, is the primary goal.

Another shift is selectivity. Firms may hire fewer juniors but invest more deeply in those they do hire. Mentorship becomes more deliberate and more resource-intensive.

There is also a design question. If traditional entry-level tasks are increasingly automated, firms need to identify new forms of work that are both valuable and difficult to automate. This likely includes tasks that require context, interpersonal judgment, or domain-specific nuance.

Culturally, mentorship may bifurcate. It becomes either passion-driven, sustained by individuals who value teaching, or highly selective, reserved for a small set of high-potential candidates.

The end of comfortable beginnings

For much of the past two decades, certain career paths offered a predictable entry point. High-paying junior roles in banking, consulting, and technology provided both financial stability and a structured path to advancement.

That model is narrowing as entry-level roles are becoming harder to access, lower paid, and more selective. The barrier to entry rises even as the tools to perform the work become more accessible. This has broader implications. Talent pipelines that rely on broad intake at the junior level may face structural pressure. If fewer people enter the system, fewer will progress through it. In this world, it becomes unclear whether firms will have sufficient access to the senior, skilled roles they still need to fill.

The more fundamental risk is not that AI replaces junior workers. It is that it disrupts the pathway through which junior workers become senior ones. We are still early in this transition, but the direction is clear. The future constraint may not be labor. It may be judgment.

If you are building or thinking in this area, particularly how to cultivate human talent in AI-disrupted industries, we would welcome the conversation.

Feel free to reach out to jpark@forgepointcap.com and kshih@forgepointcap.com.

This blog is also published on Margin of Safety, Jimmy and Kathryn’s Substack, as they research the practical sides of security + AI so you don’t have to.