July 16, 2026 • 5 Min Read

Every senior person currently working got there the same way. They did the entry-level work, they made mistakes, and someone corrected them. Over years, that accumulation of small failures and course corrections became the pattern recognition that makes a person senior. It’s not a glamorous process and it doesn’t show up on a resume, but it’s the only mechanism anyone has ever found for producing people who actually know what they’re doing.

The evidence is already on Wall Street

On Wall Street, the conversation has become unusually candid. Goldman Sachs President John Waldron has described traditional bank operations as a “human assembly line” ready for automation. Citigroup’s Jane Fraser has told staff: “some roles will change, new ones will emerge and others will no longer be required.”

What makes that candor worth paying attention to is the paradox sitting beneath it. According to QuantumBlack, McKinsey’s AI consulting arm, banks are shrinking junior analyst classes by as much as two-thirds. Those same junior cohorts supply roughly 62% of the banks’ own AI talent.

The rationale for cutting is hard to argue with in the short term. If a senior analyst plus an AI tool can produce the output of three juniors, the economic case for hiring trainees weakens considerably. The concern is that in seven years there may be no directors, because nobody learned how to become one.

Automating entry-level work doesn’t eliminate the need for senior judgment. It eliminates the path to it.

The cohort entering the workforce now may be the last to be trained in anything resembling the traditional way. They’ll arrive expecting the apprenticeship that built every senior before them and instead find themselves supervising AI systems in roles designed for productivity, not development.

That’s a problem with a long fuse. The organization looks fine. Output is up. Efficiency metrics are strong. But the internal capacity to develop the next generation of senior people, the people who can catch what agents miss, who can read a situation the model wasn’t trained on, who can decide which problems are actually worth solving, is quietly eroding. And it likely won’t announce itself until it’s expensive to fix.

Apprenticeship has to be engineered deliberately

The response most organizations reach for is a training program, but judgment doesn’t form in a classroom. It forms through experience.

An alternative looks something like this:

• Resist the urge to measure early career professionals by output. Output is what agents are for. Junior professionals are for judgment formation, and judgment forms slowly. If the metric is volume of work cleared, you’re measuring the wrong thing and incentivizing the wrong behavior.
• Make critiquing agent output part of the job description. Not approving it. Critiquing it. Grade on the quality of the critique: did this person find what was wrong, name why it was wrong, and redirect correctly? A junior who gets good at catching what agents miss is developing exactly the pattern recognition the organization needs.
• Don’t let junior employees specialize too early. Judgment develops by seeing many kinds of decisions, not repeating the same one. Rotate them through different AI-assisted workflows so they learn to recognize different patterns of failure.

Redesigning apprenticeship also means redesigning hiring. If AI increasingly performs the execution, hiring for execution no longer makes much sense.

The job description has to change before the hiring does

In today’s world, existing recruitment systems are built around skills. Resumes list skills. Job descriptions require skills. Interviews test skills. However, skills are increasingly what agents do well. This begs and obvious question: If we’re not hiring for execution anymore, what are we hiring for?

The answer? We’re hiring for the kind of capabilities AI doesn’t develop naturally:

1. Judgment is the ability to decide what’s worth doing, not just execute what’s been assigned. An AI wealth management agent can identify every portfolio that should be rebalanced. What it can’t recognize is a client who just lost a spouse shouldn’t receive an automated investment recommendation this week, regardless of what the model suggests. The recommendation isn’t technically wrong, but it’s wrong for the moment.
2. Taste is the ability to recognize when something’s wrong before you can explain why. In an interview with a recruiter, AI-generated summary will accurately capture every answer the candidate gave but miss the quality that made that candidate memorable. Nothing in the summary is technically wrong. It’s simply missing what mattered most.
3. Direction is the ability to instruct, supervise, and redirect non-human collaborators. An agent given a vague brief will produce a confident answer to the wrong question. The person who can write a precise brief, review the output against intent rather than against instructions, and redirect without starting over is doing something genuinely difficult.

None of these capabilities appear on a traditional job description. But they’re increasingly the qualities that determine whether someone can become an expert in an AI-first organization.

Human-Centered Design

The organizations that will thrive tomorrow won’t be the ones that automate the most work. They’ll be the ones that deliberately redesign how expertise is built.

Don’t stop hiring juniors. Just hire them for different reasons.

If your junior roles no longer produce the judgment your future leaders will need, that’s a workforce design challenge, not a hiring one.

See how RGP applies human-centered design to workforce transformation.

Research & Insights

You can’t hire for judgment until you understand what good judgment looks like inside your organization. That requires more than a competency model. It requires observing how experienced people actually make difficult decisions.

That’s where research begins.

Explore RGP’s Research & Insights approach.