Overview

This paper reports the first randomized field test of a large‑language‑model recruiter in an active junior‑developer search. A total of ~37,000 applicants were randomly assigned to either a conventional résumé‑screen followed by a human interview or to an AI‑led structured video interview led by micro1’s AI interviewer, Zara, with the same human interview at the end. Recruiters and final‑round interviewers never saw a candidate’s treatment assignment. 

Experiment and pipeline

Treatment candidates spent up to 40 minutes in an AI conversation that generated a skill report covering React, JavaScript, and CSS plus a soft‑skills and proctoring score. Recruiters used only that report to decide who advanced. Control recruiters relied on traditional résumé scores alone. Thirty‑five candidates from each pipeline reached the blind final interview. The figure below plots subsequent labour‑market outcomes for every treatment subgroup.

Primary results

Candidates who cleared the AI interview passed the final human interview 54 percent of the time, compared with 34 percent for controls, a 20‑percentage‑point lift. Interviewers therefore needed to conduct 44 percent fewer human interviews to find each hirable applicant, cutting recruiter time and workload, ultimately saving recruitment costs. 

The advantage extended beyond the vacancy. Five months later almost 40 percent of AI‑selected finalists reported new jobs on LinkedIn, 17 points higher than their control peers. Widening the lens to everyone who passed Zara versus everyone with the top résumé score still shows a 5.9‑point edge. These trajectories are visible in the figure above, where the treatment line outpaces all résumé‑based groups month after month. 

Quality and mechanism evidence

Zara’s interviews were markedly sharper than human first‑rounds in a separate corpus of 1,150 transcripts. Independent scoring gave the AI a mean conversational‑quality score of 7.80 versus 5.41 for humans, an improvement of more than thirty percent, and variability on both conversational and technical dimensions was far lower, indicating more consistent dialogue. 

Our research shows why these gains arise. Self‑selection at the AI stage did not siphon off high performers: drop‑outs were slightly older and more experienced, yet their subsequent job‑finding rate lagged completers. The AI interview uncovered résumé inflation as well. In the treatment sample 21 percent of candidates claimed at least one of the three required skills that the conversation proved they lacked, and a historical audit replicated that share across other roles. 

Interpretation

Placing Zara ahead of human screens improves downstream hiring accuracy by nearly a fifth, almost halves recruiter time per hire, and produces interviews that are both richer and steadier than human first‑rounds. Because those who pass the AI step land jobs more often, the experiment suggests that conversational LLM assessment surfaces genuine talent rather than simply reordering the existing pool. For high‑volume technical searches AI assessment is not just a faster filter; it is a smarter one.