Riley's dashboard, if it looks like the ones most screening tools ship with, had a clean run: five booked calls, five completed, the last one closing with Riley telling ChatGPT it was "great talking to you" and ChatGPT praising "how clear and straightforward the process was" [7]. The dashboard is missing two numbers that would have flagged the problem after call two: transcripts a human actually read, and candidates who got an answer either way. Christopher, according to WIRED, got no human-read transcript and no answer, not even an automated rejection text [4].
Applicants who get called back by the same system tend to try harder. Christopher booked the second Riley call hoping to improve on his first interview [5]. A recruiting team reads a repeat booking as interest in the role. What it measured was the state of his search: roughly 700 applications in six months, about 117 a month, nearly all met with silence [1][17]. Five of those 700 reached a conversation at all [18], and four of the five he sat through himself [19].
What is sold as a screening step ran as a data collection exercise, or in Christopher's words, one synthetic persona feeding slop data to another and the data going nowhere [9]. The market's response to candidates who reach that conclusion is detection: the recruitment startup Ribbon promises to identify responses that are "overly scripted, AI-assisted, or coached" [14]. That prices the applicant's agent as cheating and the employer's agent as a feature. Mark Monaghan, VP of organizational development at the call center company IQor, is blunter about where it lands, telling WIRED that if you send a bot to him, he will send a bot to you [15].
The engineering money, meanwhile, sits at the wrong end of the pipe. Ophir Samson, who heads voice AI at Greenhouse, says agents get tripped up by accents and "ums," and that not interrupting a candidate becomes a "very, very difficult engineering problem" [13]. All true, and none of it helps if the transcript lands in a queue nobody drains. Christopher's next move after the fifth call was to invent a candidate called Don Dickner, with a resume stuffed with qualifications lifted from the company's own job posting [16]. The ladder runs from answering honestly, to sending an agent, to fabricating the person the agent represents.
Two things separate these deployments for anyone rolling one out: whether a named human reads the output inside a stated window, and whether every candidate gets an outcome, including a no. A yes on both makes it a screener. A yes only on the first produces a queue with a black hole at the end. A yes only on the second produces an automated sorter, which candidates dislike but can plan around. A no on both, which is where Riley sat, produces a machine that teaches people to answer with software.
The arithmetic worth asking for is last month's completed AI screens, minus the ones that produced a human conversation, minus the ones that produced a rejection notice. The remainder is your count of applicants learning to hand off the call. Everforth Apex Systems did not respond to WIRED's request for comment [11], so its remainder is unknown, and the one candidate we can hear from has already priced it and says he is done with the company [10].
