WHAT 21% CONFIDENCE ACTUALLY MEANS

Axios reported this week that only 21% of recruiters are very confident their AI screening tools aren't cutting qualified candidates. Buried in that number is a bigger story than a stat about AI accuracy. It's a story about what most hiring teams actually built when they automated the top of the funnel, and it wasn't a decision. It was a filter.

The numbers

The context matters. Recruiters are now processing an average of 291 applications per hire, up from roughly 100 in early 2021, according to data cited from the recruiting firm Ashby. Indeed reports it now takes 25% longer to fill a role than it did before the pandemic, despite that surge in applicant volume. On paper, more applicants and more automation should mean faster, better hiring. The data says otherwise.

Greenhouse's survey data, also cited in the Axios report, shows why. An AI system has now interviewed sixty-three percent of U.S. applicants. Thirty-eight percent have walked away from a job entirely because an AI interview was required. 46% say their trust in the hiring process has declined. And ghost jobs, postings with no real intent to hire, still make up 18 to 22% of what's live in any given quarter.

Here's the tension worth sitting with: 49% of hiring managers say AI has improved the quality of candidates. Fewer than half of recruiters trust the tool doing the screening. The people closest to the output like what they're seeing. The people running the system don't trust it. That gap doesn't happen by accident. It happens when a tool is built to solve a volume problem and is also asked, without much scrutiny, to solve a judgment problem.

The diagnosis

A filter and a decision look similar from the outside. Both result in a shorter list. But they are solving for different things, and most hiring teams never separated the two before building around AI screening.

A filter asks: how do we get 291 applications down to a workable number? That's a real, practical problem, and AI is genuinely useful for it.

A decision asks: why this person, and not that one, for this specific role, at this specific company, right now. That's not a volume problem. It's a judgment call, and it requires knowing what you're actually optimizing for before you automate anything.

Most teams reached for AI to solve the first problem and quietly let it answer the second one too. Nobody sat down and defined what "qualified" means for the role before the tool started making that call at scale. The result is a system that is fast, defensible on paper, and running on a definition of fit that was never actually written down.

What it costs

The candidate side of this is not abstract. Gorick Ng, the Harvard career advisor cited in the Axios piece, put it plainly: candidates have moved past ordinary job search burnout into systemic cynicism. They assume every posting might be fake and every rejection might be automated. That assumption is not paranoia. It is a rational response to a hiring process that, by recruiters' own admission, has a roughly one-in-five confidence rate in its own accuracy.

That erosion shows up on your side of the table too. A 46% drop in candidate trust does not stay limited to those who get rejected. It shapes how your best prospects, the ones with options, talk about the process before they even apply. It shows up in offer acceptance rates, in referral quality, and in whether a qualified candidate finishes the process or drops out at the eleventh hour because the experience gave them insight into how the company actually operates.

The question that actually matters

Before adding another tool, or defending the one you have, ask what your screen is actually deciding, and whether anyone could explain that decision out loud, in definite terms, to a candidate who asked. If the honest answer is "it gets the list shorter," you have a filter. If the answer names the criteria, ties them to the role, and holds up under a direct question, you have a decision.

Most organizations we work with have never had to answer that question, because nobody asked it before the system went live. That's not a technology problem. It's a diagnosis problem, and it's the one worth solving before the next automated round of resumes goes out.

For most talent leaders, that diagnosis work doesn't happen because nobody gave them the mandate, the budget, or the time to do it properly. Fixing what's broken with the screen means defending a specific decision to leadership before you can even open the tool. That's the harder problem, and it's the one worth solving first.