AI and the Career Ladder Problem

Brynjolfsson, Chandar, and Chen published a paper last year that the finance-planning conversation still has not priced. Using ADP payroll data, they tracked employment by age and occupation exposure to generative AI through 2024. Workers aged 22 to 25 in high-exposure occupations saw a 13 percent relative employment decline. Older workers in the same occupations saw almost no change. Stanford HAI's 2026 AI Index cross-validates the finding with a roughly 20 percent figure for entry-level software engineers in the same age band. The effect is sharp, it is young, and it is showing up in real payroll data.

For most of 2023, the dominant story about AI and work was that it would help novices most. That story came from the Brynjolfsson, Li, and Raymond customer-support study, which found 34 percent productivity gains for novices and almost nothing for senior reps. Dell'Acqua's jagged-frontier paper pointed in a similar direction. The lab evidence suggested AI as a great leveler. Then field studies caught up. The ADP result and a parallel Upwork analysis showed experienced freelancers losing share to the tool. The novice-advantage read was a lab artifact that did not survive contact with organizational hiring practice.

Firms did two things at once. They rolled out AI tooling that let senior employees do more of the work that used to require a junior to assist. And they slowed or froze entry-level hiring because the seat the junior would have filled stopped needing a person. The 22-year-old is being routed around. The work a junior would have done is being absorbed upward into the senior workflow. That reframing matters because it changes the policy and the corporate response. Retraining programs aimed at helping young workers use AI better are solving for the wrong constraint. The constraint is that the job is gone, and the candidate's preparation is secondary.

For a CFO running a five-year workforce model, this is a structural line item. The mid-career hires of 2030 are the entry-level hires of 2026. If entry-level hiring collapses in 2026, the internal pipeline to staff VP-level roles in 2035 also collapses. External hiring fills the gap, but external hiring is expensive and unreliable, and the pool is shrinking for exactly the same reason across every firm in the industry. The firms that assumed AI would mostly substitute for lower-wage work underestimated how much of the career ladder was built on the same work.

Anthropic's March 2026 Economic Index adds texture. Usage of Claude for automation rather than augmentation has been climbing. Task diffusion is pushing into lower-wage occupations faster than into higher-wage ones, and inside high-wage occupations the specific tier most exposed is the entry-level one. The pattern shows up across vendors. The OpenAI economic index and Microsoft's Work Trend data point the same direction. The entry-level seat is the seat most exposed because that is where the work is most defined, most repeatable, and most susceptible to being done by a tool that someone else directs.

Three moves flag as useful for a finance org treating this seriously. First, audit the hiring funnel by level and by AI exposure. If entry-level hiring is flat or falling in functions tagged as high-exposure, note whether the change was deliberate or emergent. Emergent cuts are where the succession risk hides. Second, model the 2030 internal talent pipeline under the assumption that 2024 to 2026 entry-level hiring was structurally lower than trend. Identify the skill categories where the internal bench is thinnest. Third, decide whether your firm wants to preserve an entry-level training rung even without a short-term business case. The answer depends on your cost of external senior hires and the strategic value of institutional knowledge. Neither input is trivially zero.

Some firms should let the ladder shrink. A consulting firm that expects AI to compress engagement timelines may rationally run with fewer juniors. A bank with a heavy regulatory workflow may need the juniors specifically to satisfy compliance review. The point is that the decision should be explicit. Firms are letting hiring drift without naming the succession tradeoff. The ones that name it make the right call for their specific business. The ones that do not are accepting a 2030 pipeline they never modeled.

Humlum and Vestergaard's Denmark panel, which followed the same worker cohort for two years, found null earnings effects for AI users overall. The ADP and Stanford HAI numbers show the entry-level tier taking the hit instead. The distribution of AI's labor effect is lopsided, and the asymmetry sits at the career-stage boundary. A finance team that reads its own payroll data by age and exposure today will find either reassurance or a problem it can still fix. Either outcome is useful. The pipeline takes years to rebuild. The data is available now.