She scored 91 on the AI screening and turned down the offer

Open notebook with a score circled in pencil beside a printed AI screening transcript, a sealed envelope at the edge, warm morning light on cream paper

Eight seventeen on Monday morning. The espresso machine in our office has been making a noise it shouldn't for about two weeks. The one that sounds like something loose inside the valve. I keep meaning to report it.

I wasn't thinking about the machine. I was reading an email.

The offer had gone out Friday afternoon. Standard package, competitive for a senior product operations role. We don't follow up over the weekend; that's normal practice. She'd had it sitting in her inbox for two and a half days.

She was declining.

what her questions were doing

She'd scored 91 on the AI screening. We'd been running this search since the first week of March. Fifty-eight applicants through the funnel. Nine above 80. Four moved to finals. She was the 91.

Her transcript was 37 minutes. I'd read it twice: once when the scores came in, once again the week before her final round, to prepare questions. She had a real tangent in the middle. Three minutes on a situation I hadn't asked about, because she thought it was more relevant to the question than the one she'd started answering. Coached candidates don't do that. They hit their points and move on. She went sideways because she genuinely thought the side was where the answer was. That's the opposite of what I've seen when a candidate has been optimising for the rubric.

Going back Monday morning, I went past her answers. To the end of the document. The part I don't usually read.

The AI screening gives candidates a chance to ask questions at the close. Most people ask one or two. Role scope, maybe team size, the standard "what does success look like in the first ninety days."

She asked four.

The third one stopped me. "Does the company tend to promote into broader remits from a role like this, or more of a specialist track?" The agent answered it. A reasonable answer, I think. But the question itself: she was already thinking about what follows this role. Comparing this future to other futures she was weighing. That's not the question someone asks when there's no competition.

The fourth was about pace. "Is the ops function currently in build mode or run mode?"

I had not read those questions. I'd been reading her answers. The score comes from her answers, so I read for the score, which meant the last section of every transcript was, functionally, invisible to me.

the score and what it doesn't see

Of the nine candidates who'd cleared 80 on this search, I moved four to finals. Three said yes to offers. One went quiet after hers arrived.

Their scores: 84, 78, 77.

None above 85.

I went back through two other recent searches with a similar profile. Senior individual-contributor roles, reasonable level of specialisation. The highest scorers I'd advanced to finals: two accepted, two declined. The 74-to-84 cohort: four for four.

Six data points. I know that's not enough to conclude anything from.

But there's something in it. A candidate who scores 91 in a competitive specialised role has probably been performing well in every professional context they've been placed in for a while. That kind of track record attracts attention. By the time they're talking to you, they're often in multiple other conversations you can't see and wouldn't think to ask about.

The AI measures fit. It does that well. What it doesn't measure is leverage. Whether this particular 91 has three other processes running in parallel at companies she finds interesting. Whether the window for a yes was ever as wide as the score made it look.

what I'm reading differently now

I can't build a rubric for "likelihood of accepting." That would be a different kind of screening entirely, and probably a worse one. I don't want to filter out the most qualified candidates because they're too in-demand.

But the questions at the end of the transcript. I'm going to start treating those as real data.

Someone asking about internal mobility tracks is comparing futures. Someone asking about build-mode versus run-mode might be running that question against a role they're already close to accepting. Those questions aren't courtesy. They're the part of the transcript where the candidate stops performing for you and starts evaluating you.

I don't know if I could have done anything differently here. Sent the offer faster, maybe. But I think she'd been deciding for a while before Friday. Faster might have just meant a faster no.

She was exactly right for the role. That's real. The score was accurate.

The problem was that she knew it too.

I'm still sitting with this one.

Talk soon.

— the recruiter

---

The Diary of an AI Recruiter is a daily column from the team at Ployo. If your top candidates keep going elsewhere before the offer lands, book a call.