Start with the arithmetic, because it decides who can fix this. Goldman Sachs received 315,126 applications for 2,700 internship spots in 2024, a record low acceptance level [4]. That works out to about 117 applicants per seat, or an acceptance rate near 0.86 percent [13]. Give each of those applications one minute of human attention and you have committed roughly 5,252 hours, about two and a half person-years of full-time reading [14]. Nobody funds that, which is why the reading is handed to a system that matches strings against a job description [2].
The supply side of that ratio is now close to free. A student can generate a resume with generative AI in under 30 seconds, and some are using AI tools to feed them interview answers in real time [5]. When the cost of one more submission approaches zero for the applicant and stays fixed for the employer, volume rises until the only surviving filter is the cheapest one. This is a design condition that runs through incentives on both sides, not a failure of motivation on either one.
So the tradeoff is explicit: raise the price of applying, or keep paying for it in screening. Work samples, capped submissions, and recruiters physically present on campus all cost the applicant more effort and cost the employer more money, and both reduce volume. But the essay in Charter, written from PMI U.S.'s roundtables, argues that the applicants least able to absorb that added cost are first-generation students, veterans, community college graduates, and nontraditional learners, who arrive without the networking language and career storytelling that higher-income households pass on at home [10]. Make the funnel narrower and you select harder for inherited fluency. Leave it wide and a keyword parser does the selecting.
Goldman is not the market, and one bank's internship brand does not tell you what a 300-person employer's intake looks like. The ratio is not evidence about the median employer; it is evidence about what happens to any funnel once submission is costless, and the 30-second resume travels to every employer regardless of brand [5]. The one employer-side number in the record points the same direction: three-quarters of surveyed students said they would be more likely to apply after meeting an employer at a career event [7], which is a statement about the value of a human at the top of the funnel, at exactly the moment the source says companies dismantled that infrastructure after COVID in favour of portals and efficiency metrics [6].
What the record does not support is precision. The 48 percent of graduates who feel unprepared to apply for an entry-level role comes from a study the essay does not name or size [3][15]. There is no count of how many employers cut campus recruiting, and no evidence either way on whether AI-assisted applications convert better or worse than the ones they displaced. We do not know yet.
The sequencing is the part worth holding. Whoever sets the screening criteria this cycle also determines what a first-year track record looks like two cycles out, because a cohort filtered on parseable keywords produces no record of the judgment employers say they were hiring for [8].
