Most AI recruiting tools have an auto-reject feature. Set a threshold, and anyone below it gets an automated rejection email. It’s efficient. It’s also wrong.
Rob never auto-rejects. Every candidate gets a human decision point. That’s a deliberate choice that costs us speed, and we think it’s worth it.
This post explains the philosophy behind the decision, the technical mechanism that makes it work, and the data on what it actually costs us in practice.
Why auto-rejection is tempting
We understand the appeal. When you’re processing hundreds of candidates, the idea of automatically filtering out the bottom 50% is seductive. It saves time. It reduces cognitive load. And in most cases, the model is right — those candidates probably aren’t a fit.
The problem is the cases where the model is wrong. And the model is wrong in predictable, systematic ways that disproportionately affect certain groups of candidates. Career changers, people returning from caregiving, self-taught engineers, candidates from non-traditional backgrounds — these are the people most likely to be scored low by resume-matching algorithms, and they’re also some of the best hires we’ve seen.
Auto-rejection isn’t just a product decision. It’s a values decision. And our value is: no candidate should be rejected without a human being involved.
The guardrail
Auto-rejection assumes the model is right. But models are wrong in systematic ways. They under-value non-traditional backgrounds, career gaps, and industry switches. The people most likely to be auto-rejected are exactly the people most likely to be overlooked by traditional hiring.
The guardrail is simple: Rob recommends, but a human decides. If the score is below threshold, Rob flags it and explains why, but the hiring manager makes the call.
In practice, this means every candidate below the configured threshold gets a brief summary from Rob: “This candidate scored 38. Here’s what I’m concerned about: [specifics]. However, here’s what’s interesting about them: [specifics]. Would you like to proceed or pass?”
The “what’s interesting” section is critical. It forces the model to advocate for the candidate, not just against them. Even when the overall score is low, there are usually bright spots. The human needs to see both sides to make a good decision.
The override learning loop
When a hiring manager overrides Rob’s recommendation — moves forward with a candidate Rob scored low — we track the outcome. Did they get hired? Did they succeed? This feedback loop makes the model better over time.
The overrides are where the real learning happens. They’re the cases where human judgment catches something the model missed. Every override is a training signal.
In our first year, 8% of all hires came from overrides — candidates the model scored below threshold. Those hires had a 91% six-month retention rate, compared to 87% for candidates the model scored above threshold. The overrides weren’t just acceptable hires. They were slightly better than average.
This doesn’t mean the model is bad. It means the model is incomplete. It’s capturing most of the signal, but there’s a meaningful slice of information that only a human can evaluate. The override loop helps us identify what that information is and gradually teach it to the model.
What it costs us
The honest answer: it costs us speed. Without auto-rejection, every candidate requires at least a few seconds of human attention. For roles with high applicant volume, that adds up. A role with 300 applicants might generate 40–50 below-threshold flags that the hiring manager needs to review.
We’ve mitigated this by making the review process fast. Rob’s summaries are concise. The “pass” action is one click. Most below-threshold reviews take under 15 seconds. But 50 reviews at 15 seconds is still 12 minutes of work that auto-rejection would eliminate.
We think 12 minutes is a reasonable price to pay for not accidentally rejecting the best person for the job. Our customers agree — we’ve offered auto-rejection as an opt-in feature twice, and fewer than 5% of customers have turned it on. The humans want to stay in the loop.
The broader principle
This isn’t just about auto-rejection. It’s about the relationship between AI and human judgment in hiring. We believe AI should augment human decision-making, not replace it. The model surfaces information. The human makes the call.
There will come a day when models are good enough to make these decisions reliably. We’re not there yet. And until we are, the guardrail stays. Every candidate deserves a human being looking at their application. That’s the minimum bar, and it’s one we’re not willing to lower.