Interviewing
Ten Minutes of Pitch, Ten of Evidence, and the Sentence Between
The screening loses. In a twenty-minute first call that also has to sell the company, the evaluation is the half that degrades without anyone noticing, because a founder in persuasion mode asks questions that invite agreement and gets it. Split the call and say out loud where the line falls: ten minutes where you talk and they ask, ten where they talk and you write. Ask one AI question and spend the follow-ups on it.
The takeThe blended call survives because it feels good, and the good feeling is the failure. A candidate who leaves excited and a founder who leaves impressed have produced nothing either of them can hand to a third person, or to themselves in a week. What actually persuades a strong candidate is rarely enthusiasm about the company. More often it is one question about their own work that nobody else in the process thought to ask, followed by the patience to let them finish answering.
Where Olive fits
Open a role and see what the work shows
Twenty minutes can capture a candidate describing how they check a confident claim, and not much more than that. Olive puts the checking itself in front of them as work: an occupational assignment, an AI assistant that will overreach if nobody stops it, and a human reviewer who writes or edits the findings on what happened with the timestamp behind each finding.
Assess your shortlistSplit the twenty minutes and say where the line falls
Put a seam in the twenty minutes and say out loud where it falls, rather than letting the candidate guess which half of the call they are in. The screening is the half that loses when nobody marks the line, and it loses quietly. One sentence at minute ten does most of the work.
Something close to: "That was the pitch. The next ten minutes are questions about how you work, and I will be typing while you answer."
The candidate stops matching your enthusiasm and starts answering. You stop selling, which is the half nobody does by accident.
Order matters less than the boundary, though there is a sensible default.
Sell first when the candidate was not looking for a job, because ten minutes of evaluation aimed at someone who has not decided to be interested reads as an audition. Screen first when forty people applied and the call is a filter, because the pitch is wasted on most of them.
Decide which call it is before you dial. Two products are sharing one twenty-minute invitation, and only one of them can go first.
A two-person company is not obliged to build a loop around a first call, and the smallest hiring process a two-person company can defend is a real answer to a real constraint. The smallest defensible version still keeps the pitch and the evidence in separate halves, because the cost of merging them is paid entirely in the half nobody can see.
Naming the line costs a beat of awkwardness and buys the candidate permission to be concrete.
Most people describing their own work hedge until somebody tells them hedging is not what is wanted.
Why the evaluation degrades while the call still feels good
A call that has to sell is a call where you are the persuasive one, and persuasion runs on agreement: you ask questions that invite a yes, you fill the pauses, you finish the sentence when the candidate stalls. The pitch survives all of that intact. The evaluation comes out as a warm impression with nothing underneath it.
The warm part of a call is not sealed off from the scored part.
In a 2010 study, 189 accounting students went through structured mock interviews. The impression an interviewer formed during the small talk, before a single structured question, correlated .42 with that same interviewer's later structured score 1.
When different interviewers supplied the score, the figures fell to .25 and .24.
The authors do not read the result as pure contamination. The effect ran through rated competence rather than liking, and it held when the person doing the small talk supplied no structured score at all.
What the study does establish is that an impression formed in the opening minutes is still present in the score that follows them.
Interviewers also assemble a coherent account out of almost anything. In a 2013 experiment, 76 undergraduates predicted classmates' semester grade averages, and predictions made after an unstructured interview correlated .31 with the real result while prior grades alone correlated .65 2.
Later in the same paper, 96 of 169 participants chose to run an interview whose answers were generated at random over running no interview at all.
Undergraduates predicting a classmate's grades are not founders predicting job performance, and the numbers do not carry over. The mechanism does.
Low-value information dilutes good information, and a person listening to a stranger will build a story out of whatever arrives, including noise.
Twenty unmarked minutes are ideal conditions for building that story. If you are also the only person who will ever speak to this candidate, the problem compounds, and stopping a first impression from deciding it becomes a matter of procedure.
Use the evaluation half on one AI question
One question with four follow-ups beats four questions with none. A first answer about AI is often a rehearsed one, because a tools list and a story about saving time are the easiest things to prepare. The follow-ups reach the part nobody rehearses: what the candidate did when the model was confident and wrong.
Pick a question that produces an event rather than a philosophy. Ask for the last time a model handed them something that looked right and was not, and the answer arrives with a date, a document and a decision attached to it.
Then spend every follow-up on that same event. What the claim was. What it got checked against. What happened when the check disagreed with the model. Someone who did the work can answer in specifics, given time to find them, and someone who only read about the work moves to generalities and stays there.
A 2022 re-analysis of the personnel selection literature put structured interviews at .42 for predicting job performance and unstructured interviews at .19 3.
Before any statistical correction, the observed figures were .32 and .13.
Structure in that re-analysis means a research coding of format: the same questions of everyone, rated on a common scale. No particular script is implied, none of the underlying studies touched AI in any form, and the unstructured estimate is unstable enough that some of the studies behind it find nothing at all.
What survives is narrow and worth having. Across those studies, interviews that asked the same thing of everyone and rated it the same way correlated with rated job performance roughly twice as strongly as interviews that did not.
Ten minutes holds one question and its follow-ups, or two if the second is short and factual. Two AI questions worth their minutes in a fifteen-minute screen runs the same arithmetic at a slightly longer length.
One question has a failure mode of its own. A question that circulates can become a password, and when every candidate now clears the AI question, either it stopped working or the market moved.
Write the note while the candidate is still talking
Write during the call, not that evening. A note reconstructed six hours later is a summary of your impression, and the impression is the thing the selling half already colored. Typing while the candidate answers costs a little warmth and preserves what you came for: what was said, in the words it was said in.
Detailed notes are among the rarest parts of a structured interview, including in the research describing one. A content analysis of 104 interviews reported in studies published between 1997 and 2010 found detailed notes present in 19 percent of them, against 71 percent for rating each answer and 66 percent for anchored rating scales 4.
The count measures what researchers described in published studies, and the authors say it holds only as far as each study described its structure in full. Even so, practice is unlikely to be more careful than the literature describing practice.
What goes into the note is a separate question from whether one exists.
Researchers text-mined post-interview notes on 7,650 candidates hired at a large Chinese internet technology company. The more job-relevant capabilities a note named, the better that hire's later performance, promotions and retention: roughly a 2 percent rise in performance per standard deviation of the match 5.
Nobody who was rejected appears in that data, so the finding says nothing about them, and the result is correlational, at one firm, across two years of hires.
What the study measured is note content, which happens to be the part a founder controls inside twenty minutes.
So write behavior and leave the verdict for later. "Said the citation checked out until the page said something else, then went to the docket" is a note. "Strong on verification" is a rating nobody can reconstruct, including you.
What the selling half is actually for
The selling half exists for the candidate's decision, not yours, and it lands better with a real question attached to it. A person weighing three conversations remembers the one where somebody asked about their actual work and then listened to the answer. Ten minutes of focused pitch plus ten of attention beats twenty minutes of enthusiasm.
Founders resist the split because the pitch feels like the part only they can do. Ten minutes covers exactly that part: why the company exists, what the first ninety days contain, what this person would own outright.
The rest of the pitch is a document, and a document can be sent afterwards.
A senior candidate who was not looking, on a twenty-minute call with a company they have not yet decided to want, may decline to be evaluated at all. Some of them will hang up warm and never reply. The split costs those people, and that is the market argument for keeping the call blended.
Book the second conversation before the first one ends, while the answer is still on the screen in front of you.
The twenty minutes stay twenty minutes. Ten of them now produce a sentence somebody else can read and disagree with. The call that felt best produced none.
Common questions
Should the pitch come first or the screening questions?
Sell first when the candidate was not looking for a job, because ten minutes of questions aimed at someone who has not decided to be interested reads as an audition. Screen first when the role drew forty applications and the call is a filter, since the pitch is wasted on most people who will not advance. Either order works. What does not work is running both at once, because the persuading voice and the evaluating voice ask different kinds of questions, and only one of them produces an answer comparable to the next candidate's.
How many AI questions fit in ten minutes?
One, about a specific time a model's output looked right and was not, with three or four follow-ups on that same event. A second question fits only if it is short and factual. Four questions with no follow-ups tend to produce four rehearsed answers, which is the outcome that feels productive and compares to nothing later. The follow-ups hold the evidence: what the claim was, what it was checked against, and what happened when the check disagreed with the model.
What if the candidate spends the screening half asking about the company?
Answer briefly and put the time back. A candidate with real questions is a good sign, and the questions belong in the selling half, so say that out loud: offer to take them at the end and hold the remaining minutes for the work. If the questions keep coming, the call has told you which stage this person is at, which is worth knowing and is not an evaluation. Book a second conversation rather than treating the first one as evidence.
Is one AI question enough to decide on?
No. One question in a first call is a single observation from twenty minutes, and it decides whether to spend a longer round rather than whether to hire. Treat it as a filter with a low bar: an answer carrying a real event advances, an answer that stays general does not. The heavier evidence comes later, from work the candidate does in front of you.
Does typing during the call make it feel like an interrogation?
It can when the typing goes unexplained, so announce it. Saying that the pauses are notes turns silence into procedure instead of judgment, and a candidate can read it as being taken seriously. Writing the note afterwards is what actually costs: six hours later the record is a summary of your impression, and the impression formed during the half of the call where you were the one doing the persuading.
References
- 1. Initial Evaluations in the Interview: Relationships with Subsequent Interviewer Evaluations and Employment Offers ✓ homepages.se.edu Supports the claim that the impression formed in the rapport-building small talk correlates .42 with the same interviewer's structured score and .25 and .24 across interviewers, so an unmarked warm opening is not sealed off from the evaluation.
- 2. Belief in the unstructured interview: The persistence of an illusion ✓ sjdm.org Supports the dilution mechanism: predictions after an unstructured interview correlated .31 with actual grades against .65 for prior grades alone, and 96 of 169 participants preferred an interview with randomly generated answers to no interview.
- 3. Revisiting Meta-Analytic Estimates of Validity in Personnel Selection: Addressing Systematic Overcorrection for Restriction of Range ✓ gwern.net Supports the claim that interviews coded as structured correlated with rated job performance roughly twice as strongly as unstructured ones: .42 against .19, and .32 against .13 before correction.
- 4. The Structured Employment Interview: Narrative and Quantitative Review of the Research Literature ✓ doi.org Supports the claim that detailed notes are rare even in published structured-interview research, present in 19 percent of 104 interviews reported between 1997 and 2010, against 71 percent for rating each answer and 66 percent for anchored rating scales.
- 5. Predictive Validity of Interviewer Post-interview Notes on Candidates' Job Outcomes: Evidence Using Text Data From a Leading Chinese IT Company ✓ frontiersin.org Supports the claim that what a note names matters: across 7,650 hired candidates, notes naming more job-relevant capabilities went with roughly a 2 percent rise in later performance per standard deviation of match.
5 sources, numbered by first appearance. Every one was opened and checked against the claim it carries. How Olive sources claims
General guidance for hiring teams. What works at one company and one volume may not transfer to yours.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.