Yesterday, the AI Caught Something I Missed

Open notebook beside a laptop showing a candidate scorecard — second look

A note before the entry

This is the first entry of The Diary of an AI Recruiter. The premise is simple: I'll write down what actually happened in a week of hiring — the moments I want to remember, the calls I got wrong, the patterns I keep seeing. No frameworks. No listicles. Just the small things that change how I work.

Names have been changed. Everything else is real.

Yesterday

I almost rejected Marwa.

Her CV was on the pile I'd already moved into the "no" column — three short stints in three years, no exposure to the specific platform the role calls out, a one-line "career break" that I didn't ask about. I'd given her about nine seconds. By the standards of how recruiters actually scan a CV, that was generous.

The AI gave her an 87.

I'd configured the screening so anyone above 80 lands in a Tuesday morning review queue. I clicked into her profile expecting to find the scoring rubric had over-weighted something — verbal fluency, maybe, or the soft-skills bucket. The thing recruiters quietly accuse AI tools of doing: rewarding people who interview well, not people who can actually do the job.

That wasn't it.

The screening call lasted nineteen minutes. The transcript was longer than I expected for the time. She was efficient. I read the section the model had flagged as "high signal" — a question about a system she'd built at her second job, the one I'd dismissed as too short to count.

She had decided not to use a queue.

I read it twice.

She had decided not to use a queue, in a context where any junior engineer would have reached for one, because the throughput pattern she was actually seeing made the queue's failure mode worse than the problem it was solving. She knew the constraint. She knew the next-worst option. She knew why she'd chosen the thing that looked, on a CV bullet, like a missing piece of architecture.

The CV said: Built event-processing pipeline.

The conversation said: I understood the tradeoff that everyone in your shortlist has just been parroting back at you from a textbook.

Why I'd written her off

I want to be honest about this part. I didn't filter her out because the rubric told me to. I filtered her out because, in the last two years, the four hires who came through the door with her profile — career switcher, three short stints, big gap in the obvious-keyword column — did not work out. I built a model. The model isn't documented anywhere. It just runs.

The problem with the model is the sample size. Four people is not a dataset. It's a feeling.

The AI doesn't have my feeling. It can't. It has the conversation she had with our screening agent, and the comparison between what she said and what the role requires, and a rubric I can look at and argue with line by line. When it disagrees with me, the disagreement is legible. I can read why. I can read it twice. I can decide it's wrong.

I couldn't do that with my own gut. My gut just said "no" and moved on.

What I'm taking from this

A few things.

One. The value of the AI in my workflow is not that it's smarter than me. It isn't, on the questions where I'm right. The value is that it forces me to look at the candidates I would not have looked at — and to look at them with evidence in hand, not vibes. The cost of that second look is small. The cost of a false negative on a senior hire is not.

Two. I'm going to start logging the candidates I move from "no" to "yes" after the AI surfaces them. I want to know, in six months, whether the people I would have rejected are performing better, worse, or the same as the people I would have advanced on instinct. If my gut is right, the data will show it. If it isn't, I want to know that too.

Three. I'm sending Marwa to the next round.

What I'm watching for this week

• A junior engineering role with sixty CVs and not enough hours.
• Whether the candidate I rated highest from last week's longlist actually responds to the offer email today.
• A reference check on someone the AI scored low and I scored high. I want to see which of us is wrong before I write next week's entry.

If you're reading this and you build hiring tools: the most useful thing your product can do for me is not "find the perfect candidate." It's "tell me, with reasons, which of my no's you disagree with." That's the work.

Talk soon.

— the recruiter

---

The Diary of an AI Recruiter is a weekly first-person column from the team building Ployo. If you want the AI to take the same kind of second look at your own funnel, book a call with the founder.