Adam Ward, Cursor's head of talent, laid out the challenge he has with traditional recruiting on Lenny's Podcast this week. It starts with volume: contact 100 people, get 20 responses, screen down, hire from the remaining group. Ward calls it the "funnel of doom," and his objection is to potential quality—the 20 who responded were never the best 20, just the ones willing to talk. A company can run a perfect interview process on the wrong population.
His alternative is executive search applied to every role. Great exec teams define what exceptional performance requires, map the people who appear capable of it—Ward suggests assuming roughly 50 exist for any given job—and pursue them for weeks, months, sometimes years. He warns against using prestigious employers as a proxy: the question is not who the best product engineer someone knows is, but who they've worked with who is exceptional at collaborating with designers, or translating frameworks into products.
Cursor has organized itself around this. Employees drop impressive people into a #hiring-ideas Slack channel; the company works out who knows a promising name and may open a dedicated channel to coordinate the pursuit of one candidate. Brie Wolfson, who spent two months embedded at Cursor last year, reported spending about a quarter of her time recruiting. The company had grown from under 20 employees to nearly 250 in roughly a year, leadership still approved every hire, and the atomic unit of recruiting, in her phrase, was the person rather than the job spec.
Large-company folks will argue that this works because Cursor is small. New research suggests companies like Cursor may stay small. A working paper from INSEAD's Hyunjin Kim and Harvard's Rembrand Koning compared AI-native startups with similar venture-backed firms and found that among Y Combinator companies, the AI-native ones employed about 25% fewer people—more engineers, roughly 15% fewer managers and entry-level workers, flatter hierarchies—while posting comparable valuations and higher valuation per employee. Similar, smaller effects appear in a broader sample of US venture-backed startups.
If a company reaches scale with 500 people where a predecessor needed 2,000, it can spend four times the organizational attention per hire at the same total cost. This should still work at larger sizes than you might expect: a 1,000-person company replacing 15% of its workforce annually makes 150 hires, so mapping 50 candidates per vacancy means 7,500 targets a year—research AI can increasingly do, leaving humans to define excellence and build relationships. At 5,000 people the system would more likely differentiate than disappear, with executive-search treatment for roles where the gap between adequate and exceptional is large, rather than exclusively for levels above an arbitrary seniority level. Everything else gets a more standard throughput-oriented process.
Regardless of the scale at which it’s applied, the model's weak point is discovery. Selection research starts from nearly the opposite premise: humans are unreliable judges of future performance, so broaden access and improve assessment. Structured interviews rank among the strongest general predictors of job performance in recent meta-analytic work, with lower adverse racial impact than several alternatives—structure exists partly to keep confidence, reputation, and familiarity from doing the selecting. Ward does not reject assessment; Cursor runs hard interviews and work trials. But any system that identifies the 50 exceptional people before formally assessing them depends on who is visible to the people drawing the map.
Cursor illustrates both sides. Wolfson describes candidates surfaced through unusual signals—someone running Cursor workshops in Stockholm, a user who appeared as an extreme outlier in product telemetry. Wolfson also reported that nearly 40% of employees had attended one of eight highly selective universities, and that women in product and engineering were a known recruiting weakness.
There is a second-order problem if the model spreads. The same research finding smaller AI-native firms also finds fewer entry-level roles. Pair that with recruiting effort concentrated on already-proven performers and the result is a labor market that reallocates demonstrated talent efficiently while producing less of it. Broad funnels are wasteful, but they also discover people—large organizations have historically absorbed inexperienced workers and given people without elite networks a route to become legible through performance.
On this reading, funnel hiring was partly a technology for staffing very large organizations. If AI-native firms can grow without adding people at the old rate, the economics of hiring attention move with them. The unresolved question is who gets onto the map.
